arXiv:2501.09223v1 [cs.CL] 16 Jan 2025

Foundations of
Large Language Models

Tong Xiao and Jingbo Zhu

January 17, 2025

NLP Lab, Northeastern University & NiuTrans Research

Copyright © 2021-2025 Tong Xiao and Jingbo Zhu
NATURAL L ANGUAGE P ROCESSING L AB , N ORTHEASTERN U NIVERSITY
&
N IU T RANS R ESEARCH
Licensed under the Creative Commons Attribution-NonCommercial 4.0 Unported License (the
“License”). You may not use this file except in compliance with the License. You may obtain a copy of the License at http://creativecommons.org/licenses/by-nc/4.0. Unless
required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS ” BASIS , WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either
express or implied. See the License for the specific language governing permissions and limitations under the License.
January 17, 2025

Preface
Large language models originated from natural language processing, but they have undoubtedly
become one of the most revolutionary technological advancements in the field of artificial intelligence in recent years. An important insight brought by large language models is that knowledge
of the world and languages can be acquired through large-scale language modeling tasks, and
in this way, we can create a universal model that handles diverse problems. This discovery has
profoundly impacted the research methodologies in natural language processing and many related
disciplines. We have shifted from training specialized systems from scratch using a large amount
of labeled data to a new paradigm of using large-scale pre-training to obtain foundation models,
which are then fine-tuned, aligned, and prompted.
This book aims to outline the basic concepts of large language models and introduce the
related techniques. As the title suggests, the book focuses more on the foundational aspects of
large language models rather than providing comprehensive coverage of all cutting-edge methods.
The book consists of four chapters:
• Chapter 1 introduces the basics of pre-training. This is the foundation of large language
models, and common pre-training methods and model architectures will be discussed here.
• Chapter 2 introduces generative models, which are the large language models we commonly
refer to today. After presenting the basic process of building these models, we will also
explore how to scale up model training and handle long texts.
• Chapter 3 introduces prompting methods for large language models. We will discuss various prompting strategies, along with more advanced methods such as chain-of-thought
reasoning and automatic prompt design.
• Chapter 4 introduces alignment methods for large language models. This chapter focuses
on instruction fine-tuning and alignment based on human feedback.
If readers have some background in machine learning and natural language processing, along
with a certain understanding of neural networks like Transformers, reading this book will be quite
easy. However, even without this prior knowledge, it is still perfectly fine, as we have made the
content of each chapter as self-contained as possible, ensuring that readers will not be burdened
with too much reading difficulty.
In writing this book, we have gradually realized that it is more like a compilation of "notes" we
have taken while learning about large language models. Through this note-taking writing style, we
hope to offer readers a flexible learning path. Whether they wish to dive deep into a specific area
or gain a comprehensive understanding of large language models, they will find the knowledge
and insights they need within these "notes".
We would like to thank the students in our laboratory and all our friends who have shared
with us their views on large language models and helped with corrections of errors in writing. In
particular, we wish to thank Weiqiao Shan, Yongyu Mu, Chenglong Wang, Kaiyan Chang, Yuchun
Fan, Hang Zhou, Xinyu Liu, Huiwen Bao, Tong Zheng, Junhao Ruan, and Qing Yang.

ii

Notation
a

variable

a

row vector or matrix

f (a)
max f (a)

function of a
maximum value of f (a)

arg maxa f (a)

value of a that maximizes f (a)

x

input token sequence to a model

xj

input token at position j

y

output token sequence produced by a model

yi

output token at position i

θ

model parameters

Pr(a)

probability of a

Pr(a|b)

conditional probability of a given b

Pr(·|b)

probability distribution of a variable given b

Prθ (a)

probability of a as parameterized by θ

ht

hidden state at time step t in sequential models

H

matrix of all hidden states over time in a sequence

Q, K, V
Softmax(A)

query, key, and value matrices in attention mechanisms
Softmax function that normalizes the input vector or matrix A

L

loss function

D

dataset used for training or fine-tuning a model

∂L
∂θ

gradient of the loss function L with respect to the parameters θ

KL(p || q)

KL divergence between distributions p and q

iii

Contents
1 Pre-training
1.1

1

Pre-training NLP Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.1.1

Unsupervised, Supervised and Self-supervised Pre-training . . . . . . . .

2

1.1.2

Adapting Pre-trained Models . . . . . . . . . . . . . . . . . . . . . . . .

3

Self-supervised Pre-training Tasks . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.2.1

Decoder-only Pre-training . . . . . . . . . . . . . . . . . . . . . . . . .

7

1.2.2

Encoder-only Pre-training . . . . . . . . . . . . . . . . . . . . . . . . .

8

1.2.3

Encoder-Decoder Pre-training . . . . . . . . . . . . . . . . . . . . . . .

15

1.2.4

Comparison of Pre-training Tasks . . . . . . . . . . . . . . . . . . . . .

20

Example: BERT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

21

1.3.1

The Standard Model . . . . . . . . . . . . . . . . . . . . . . . . . . . .

21

1.3.2

More Training and Larger Models . . . . . . . . . . . . . . . . . . . . .

27

1.3.3

More Efficient Models . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

1.3.4

Multi-lingual Models . . . . . . . . . . . . . . . . . . . . . . . . . . . .

28

1.4

Applying BERT Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

30

1.5

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

35

1.2

1.3

2 Generative Models
2.1

2.2

2.3

36

A Brief Introduction to LLMs . . . . . . . . . . . . . . . . . . . . . . . . . . .

37

2.1.1

Decoder-only Transformers . . . . . . . . . . . . . . . . . . . . . . . .

38

2.1.2

Training LLMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

40

2.1.3

Fine-tuning LLMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

42

2.1.4

Aligning LLMs with the World . . . . . . . . . . . . . . . . . . . . . .

46

2.1.5

Prompting LLMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

51

Training at Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

56

2.2.1

Data Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

56

2.2.2

Model Modifications . . . . . . . . . . . . . . . . . . . . . . . . . . . .

58

2.2.3

Distributed Training . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

2.2.4

Scaling Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

Long Sequence Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

66

2.3.1

Optimization from HPC Perspectives . . . . . . . . . . . . . . . . . . .

67

2.3.2

Efficient Architectures . . . . . . . . . . . . . . . . . . . . . . . . . . .

68

2.3.3

Cache and Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

71

2.3.4

Sharing across Heads and Layers . . . . . . . . . . . . . . . . . . . . .

80

iv

v

2.4

2.3.5

Position Extrapolation and Interpolation . . . . . . . . . . . . . . . . . .

82

2.3.6

Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

92

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

94

3 Prompting
3.1

3.2

3.3

3.4

96

General Prompt Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

3.1.1

Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

97

3.1.2

In-context Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

99

3.1.3

Prompt Engineering Strategies . . . . . . . . . . . . . . . . . . . . . . . 101

3.1.4

More Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

Advanced Prompting Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.2.1

Chain of Thought . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

3.2.2

Problem Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . 117

3.2.3

Self-refinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

3.2.4

Ensembling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

3.2.5

RAG and Tool Use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

Learning to Prompt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
3.3.1

Prompt Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

3.3.2

Soft Prompts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

3.3.3

Prompt Length Reduction . . . . . . . . . . . . . . . . . . . . . . . . . 152

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

4 Alignment

155

4.1

An Overview of LLM Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . 155

4.2

Instruction Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

4.3

4.4

4.2.1

Supervised Fine-tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

4.2.2

Fine-tuning Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . 161

4.2.3

Fine-tuning with Less Data . . . . . . . . . . . . . . . . . . . . . . . . . 166

4.2.4

Instruction Generalization . . . . . . . . . . . . . . . . . . . . . . . . . 167

4.2.5

Using Weak Models to Improve Strong Models . . . . . . . . . . . . . . 169

Human Preference Alignment: RLHF . . . . . . . . . . . . . . . . . . . . . . . 172
4.3.1

Basics of Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . 173

4.3.2

Training Reward Models . . . . . . . . . . . . . . . . . . . . . . . . . . 179

4.3.3

Training LLMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

Improved Human Preference Alignment . . . . . . . . . . . . . . . . . . . . . . 187
4.4.1

Better Reward Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 187

Notation

vi

4.5

4.4.2

Direct Preference Optimization . . . . . . . . . . . . . . . . . . . . . . 193

4.4.3

Automatic Preference Data Generation . . . . . . . . . . . . . . . . . . 196

4.4.4

Step-by-step Alignment . . . . . . . . . . . . . . . . . . . . . . . . . . 198

4.4.5

Inference-time Alignment . . . . . . . . . . . . . . . . . . . . . . . . . 200

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

Bibliography

203

C HAPTER 1

Pre-training
The development of neural sequence models, such as Transformers [Vaswani et al., 2017], along
with the improvements in large-scale self-supervised learning, has opened the door to universal
language understanding and generation. This achievement is largely motivated by pre-training:
we separate common components from many neural network-based systems, and then train them
on huge amounts of unlabeled data using self-supervision. These pre-trained models serve as
foundation models that can be easily adapted to different tasks via fine-tuning or prompting. As a
result, the paradigm of NLP has been enormously changed. In many cases, large-scale supervised
learning for specific tasks is no longer required, and instead, we only need to adapt pre-trained
foundation models.
While pre-training has gained popularity in recent NLP research, this concept dates back
decades to the early days of deep learning. For example, early attempts to pre-train deep learning
systems include unsupervised learning for RNNs, deep feedforward networks, autoencoders, and
others [Schmidhuber, 2015]. In the modern era of deep learning, we experienced a resurgence of
pre-training, caused in part by the large-scale unsupervised learning of various word embedding
models [Mikolov et al., 2013b; Pennington et al., 2014]. During the same period, pre-training also
attracted significant interest in computer vision, where the backbone models were trained on relatively large labeled datasets such as ImageNet, and then applied to different downstream tasks
[He et al., 2019; Zoph et al., 2020]. Large-scale research on pre-training in NLP began with the
development of language models using self-supervised learning. This family of models covers
several well-known examples like BERT [Devlin et al., 2019] and GPT [Brown et al., 2020], all
with a similar idea that general language understanding and generation can be achieved by training the models to predict masked words in a huge amount of text. Despite the simple nature of
this approach, the resulting models show remarkable capability in modeling linguistic structure,
though they are not explicitly trained to achieve this. The generality of the pre-training tasks
leads to systems that exhibit strong performance in a large variety of NLP problems, even outperforming previously well-developed supervised systems. More recently, pre-trained large language
models have achieved a greater success, showing the exciting prospects for more general artificial
intelligence [Bubeck et al., 2023].
This chapter discusses the concept of pre-training in the context of NLP. It begins with a general introduction to pre-training methods and their applications. BERT is then used as an example
to illustrate how a sequence model is trained via a self-supervised task, called masked language
modeling. This is followed by a discussion of methods for adapting pre-trained sequence models for various NLP tasks. Note that in this chapter, we will focus primarily on the pre-training
paradigm in NLP, and therefore, we do not intend to cover details about generative large language
models. A detailed discussion of these models will be left to subsequent chapters.

1.1 Pre-training NLP Models
The discussion of pre-training issues in NLP typically involves two types of problems: sequence
modeling (or sequence encoding) and sequence generation. While these problems have different
1

Pre-training

2
forms, for simplicity, we describe them using a single model defined as follows:
o = g(x0 , x1 , ..., xm ; θ)
= gθ (x0 , x1 , ..., xm )

(1.1)

where {x0 , x1 , ..., xm } denotes a sequence of input tokens1 , x0 denotes a special symbol (hsi or
[CLS]) attached to the beginning of a sequence, g(·; θ) (also written as gθ (·)) denotes a neural
network with parameters θ, and o denotes the output of the neural network. Different problems
can vary based on the form of the output o. For example, in token prediction problems (as in
language modeling), o is a distribution over a vocabulary; in sequence encoding problems, o is a
representation of the input sequence, often expressed as a real-valued vector sequence.
There are two fundamental issues here.
• Optimizing θ on a pre-training task. Unlike standard learning problems in NLP, pre-training
does not assume specific downstream tasks to which the model will be applied. Instead, the
goal is to train a model that can generalize across various tasks.
• Applying the pre-trained model gθ̂ (·) to downstream tasks. To adapt the model to these
tasks, we need to adjust the parameters θ̂ slightly using labeled data or prompt the model
with task descriptions.
In this section, we discuss the basic ideas in addressing these issues.

1.1.1 Unsupervised, Supervised and Self-supervised Pre-training
In deep learning, pre-training refers to the process of optimizing a neural network before it is
further trained/tuned and applied to the tasks of interest. This approach is based on an assumption
that a model pre-trained on one task can be adapted to perform another task. As a result, we do
not need to train a deep, complex neural network from scratch on tasks with limited labeled data.
Instead, we can make use of tasks where supervision signals are easier to obtain. This reduces the
reliance on task-specific labeled data, enabling the development of more general models that are
not confined to particular problems.
During the resurgence of neural networks through deep learning, many early attempts to
achieve pre-training were focused on unsupervised learning. In these methods, the parameters of a neural network are optimized using a criterion that is not directly related to specific tasks.
For example, we can minimize the reconstruction cross-entropy of the input vector for each layer
[Bengio et al., 2006]. Unsupervised pre-training is commonly employed as a preliminary step
before supervised learning, offering several advantages, such as aiding in the discovery of better
local minima and adding a regularization effect to the training process [Erhan et al., 2010]. These
benefits make the subsequent supervised learning phase easier and more stable.
A second approach to pre-training is to pre-train a neural network on supervised learning
tasks. For example, consider a sequence model designed to encode input sequences into some
1
Here we assume that tokens are basic units of text that are separated through tokenization. Sometimes, we will use
the terms token and word interchangeably, though they have closely related but slightly different meanings in NLP.

1.1 Pre-training NLP Models

3

representations. In pre-training, this model is combined with a classification layer to form a classification system. This system is then trained on a pre-training task, such as classifying sentences
based on sentiment (e.g., determining if a sentence conveys a positive or negative sentiment).
Then, we adapt the sequence model to a downstream task. We build a new classification system
based on this pre-trained sequence model and a new classification layer (e.g., determining if a
sequence is subjective or objective). Typically, we need to fine-tune the parameters of the new
model using task-specific labeled data, ensuring the model is optimally adjusted to perform well
on this new type of data. The fine-tuned model is then employed to classify new sequences for
this task. An advantage of supervised pre-training is that the training process, either in the pretraining or fine-tuning phase, is straightforward, as it follows the well-studied general paradigm
of supervised learning in machine learning. However, as the complexity of the neural network
increases, the demand for more labeled data also grows. This, in turn, makes the pre-training task
more difficult, especially when large-scale labeled data is not available.
A third approach to pre-training is self-supervised learning. In this approach, a neural network is trained using the supervision signals generated by itself, rather than those provided by
humans. This is generally done by constructing its own training tasks directly from unlabeled
data, such as having the system create pseudo labels. While self-supervised learning has recently
emerged as a very popular method in NLP, it is not a new concept. In machine learning, a related
concept is self-training where a model is iteratively improved by learning from the pseudo labels
assigned to a dataset. To do this, we need some seed data to build an initial model. This model
then generates pseudo labels for unlabeled data, and these pseudo labels are subsequently used to
iteratively refine and bootstrap the model itself. Such a method has been successfully used in several NLP areas, such as word sense disambiguation [Yarowsky, 1995] and document classification
[Blum and Mitchell, 1998]. Unlike the standard self-training method, self-supervised pre-training
in NLP does not rely on an initial model for annotating the data. Instead, all the supervision signals are created from the text, and the entire model is trained from scratch. A well-known example
of this is training sequence models by successively predicting a masked word given its preceding
or surrounding words in a text. This enables large-scale self-supervised learning for deep neural
networks, leading to the success of pre-training in many understanding, writing, and reasoning
tasks.
Figure 1.1 shows a comparison of the above three pre-training approaches. Self-supervised
pre-training is so successful that most current state-of-the-art NLP models are based on this
paradigm. Therefore, in this chapter and throughout this book, we will focus on self-supervised
pre-training. We will show how sequence models are pre-trained via self-supervision and how the
pre-trained models are applied.

1.1.2 Adapting Pre-trained Models
As mentioned above, two major types of models are widely used in NLP pre-training.

• Sequence Encoding Models. Given a sequence of words or tokens, a sequence encoding
model represents this sequence as either a real-valued vector or a sequence of vectors, and
obtains a representation of the sequence. This representation is typically used as input to
another model, such as a sentence classification system.

Pre-training

4

Prompting
Zero/Few
Shot Learning

Pre-training

Training

Pre-training

Tuning

Pre-training

Tuning

Unsupervised

Supervised

Supervised

Supervised

SelfSupervised

Supervised

Unlabeled
Data

Labeled
Data

Labeled
Data

Labeled
Data

Unlabeled
Data

Labeled
Data

Task 2
Task 1

(a) Unsupervised Pre-training

(b) Supervised Pre-training

(c) Self-supervised Pre-training

Fig. 1.1: Illustration of unsupervised, supervised, and self-supervised pre-training. In unsupervised pre-training, the
pre-training is performed on large-scale unlabeled data. It can be viewed as a preliminary step to have a good starting
point for the subsequent optimization process, though considerable effort is still required to further train the model
with labeled data after pre-training. In supervised pre-training, the underlying assumption is that different (supervised)
learning tasks are related. So we can first train the model on one task, and transfer the resulting model to another task
with some training or tuning effort. In self-supervised pre-training, a model is pre-trained on large-scale unlabeled data
via self-supervision. The model can be well trained in this way, and we can efficiently adapt it to new tasks through
fine-tuning or prompting.

• Sequence Generation Models. In NLP, sequence generation generally refers to the problem of generating a sequence of tokens based on a given context. The term context has
different meanings across applications. For example, it refers to the preceding tokens in
language modeling, and refers to the source-language sequence in machine translation2 .

We need different techniques for applying these models to downstream tasks after pre-training.
Here we are interested in the following two methods.

1.1.2.1 Fine-tuning of Pre-trained Models
For sequence encoding pre-training, a common method of adapting pre-trained models is finetuning. Let Encodeθ (·) denote an encoder with parameters θ, for example, Encodeθ (·) can be a
standard Transformer encoder. Provided we have pre-trained this model in some way and obtained
the optimal parameters θ̂, we can employ it to model any sequence and generate the corresponding
representation, like this
H = Encodeθ̂ (x)

(1.2)

where x is the input sequence {x0 , x1 , ..., xm }, and H is the output representation which is a
sequence of real-valued vectors {h0 , h1 , ..., hm }. Because the encoder does not work as a standalone NLP system, it is often integrated as a component into a bigger system. Consider, for
example, a text classification problem in which we identify the polarity (i.e., positive, negative,
2
More precisely, in auto-regressive decoding of machine translation, each target-language token is generated based
on both its preceding tokens and source-language sequence.

1.1 Pre-training NLP Models

5

and neutral) of a given text. We can build a text classification system by stacking a classifier
on top of the encoder. Let Classifyω (·) be a neural network with parameters ω. Then, the text
classification model can be expressed in the form
Prω,θ̂ (·|x) = Classifyω (H)
(1.3)

= Classifyω (Encodeθ̂ (x))

Here Prω,θ̂ (·|x) is a probability distribution over the label set {positive, negative, neutral}, and
the label with the highest probability in this distribution is selected as output. To keep the notation
uncluttered, we will use Fω,θ̂ (·) to denote Classifyω (Encodeθ̂ (·)).
Because the model parameters ω and θ̂ are not optimized for the classification task, we cannot
directly use this model. Instead, we must use a modified version of the model that is adapted to
the task. A typical way is to fine-tune the model by giving explicit labeling in downstream tasks.
We can train Fω,θ̂ (·) on a labeled dataset, treating it as a common supervised learning task. The
outcome of the fine-tuning is the parameters ω̃ and θ̃ that are further optimized. Alternatively,
we can freeze the encoder parameters θ̂ to maintain their pre-trained state, and focus solely on
optimizing ω. This allows the classifier to be efficiently adapted to work in tandem with the
pre-trained encoder.
Once we have obtained a fine-tuned model, we can use it to classify a new text. For example,
suppose we have a comment posted on a travel website:
I love the food here. It’s amazing!
We first tokenize this text into tokens3 , and then feed the token sequence xnew into the fine-tuned
model Fω̃,θ̃ (·). The model generates a distribution over classes by
Fω̃,θ̃ (xnew ) =

h

Pr(positive|xnew ) Pr(negative|xnew ) Pr(neutral|xnew )

i

(1.4)

And we select the label of the entry with the maximum value as output. In this example it is
positive.
In general, the amount of labeled data used in fine-tuning is small compared to that of the
pre-training data, and so fine-tuning is less computationally expensive. This makes the adaption
of pre-trained models very efficient in practice: given a pre-trained model and a downstream task,
we just need to collect some labeled data, and slightly adjust the model parameters on this data. A
more detailed discussion of fine-tuning can be found in Section 1.4.

1.1.2.2 Prompting of Pre-trained Models
Unlike sequence encoding models, sequence generation models are often employed independently
to address language generation problems, such as question answering and machine translation,
without the need for additional modules. It is therefore straightforward to fine-tune these models
3
The text can be tokenized in many different ways. One of the simplest is to segment the text into English words
and punctuations {I, love, the, food, here, ., It, ’s, amazing, !}

Pre-training

6

as complete systems on downstream tasks. For example, we can fine-tune a pre-trained encoderdecoder multilingual model on some bilingual data to improve its performance on a specific translation task.
Among various sequence generation models, a notable example is the large language models
trained on very large amounts of data. These language models are trained to simply predict the next
token given its preceding tokens. Although token prediction is such a simple task that it has long
been restricted to “language modeling” only, it has been found to enable the learning of the general
knowledge of languages by repeating the task a large number of times. The result is that the
pre-trained large language models exhibit remarkably good abilities in token prediction, making
it possible to transform numerous NLP problems into simple text generation problems through
prompting the large language models. For example, we can frame the above text classification
problem as a text generation task

I love the food here. It’s amazing! I’m

Here
indicates the word or phrase we want to predict (call it the completion). If the predicted
word is happy, or glad, or satisfied or a related positive word, we can classify the text as positive.
This example shows a simple prompting method in which we concatenate the input text with I’m
to form a prompt. Then, the completion helps decide which label is assigned to the original text.
Given the strong performance of language understanding and generation of large language
models, a prompt can instruct the models to perform more complex tasks. Here is a prompt where
we prompt the LLM to perform polarity classification with an instruction.

Assume that the polarity of a text is a label chosen from {positive, negative,
neutral}. Identify the polarity of the input.
Input: I love the food here. It’s amazing!
Polarity:

The first two sentences are a description of the task. Input and Polarity are indicators of the input
and output, respectively. We expect the model to complete the text and at the same time give the
correct polarity label. By using instruction-based prompts, we can adapt large language models to
solve NLP problems without the need for additional training.
This example also demonstrates the zero-shot learning capability of large language models,
which can perform tasks that were not observed during the training phase. Another method for
enabling new capabilities in a neural network is few-shot learning. This is typically achieved
through in-context learning (ICT). More specifically, we add some samples that demonstrate how
an input corresponds to an output. These samples, known as demonstrations, are used to teach
large language models how to perform the task. Below is an example involving demonstrations

1.2 Self-supervised Pre-training Tasks

7

Assume that the polarity of a text is a label chosen from {positive, negative,
neutral}. Identify the polarity of the input.
Input: The traffic is terrible during rush hours, making it difficult to reach the
airport on time.
Polarity: Negative
Input: The weather here is wonderful.
Polarity: Positive
Input: I love the food here. It’s amazing!
Polarity:
Prompting and in-context learning play important roles in the recent rise of large language
models. We will discuss these issues more deeply in Chapter 3. However, it is worth noting
that while prompting is a powerful way to adapt large language models, some tuning efforts are
still needed to ensure the models can follow instructions accurately. Additionally, the fine-tuning
process is crucial for aligning the values of these models with human values. More detailed
discussions of fine-tuning can be found in Chapter 4.

1.2 Self-supervised Pre-training Tasks
In this section, we consider self-supervised pre-training approaches for different neural architectures, including decoder-only, encoder-only, and encoder-decoder architectures. We restrict our
discussion to Transformers since they form the basis of most pre-trained models in NLP. However, pre-training is a broad concept, and so we just give a brief introduction to basic approaches
in order to make this section concise.

1.2.1 Decoder-only Pre-training
The decoder-only architecture has been widely used in developing language models [Radford et al.,
2018]. For example, we can use a Transformer decoder as a language model by simply removing
cross-attention sub-layers from it. Such a model predicts the distribution of tokens at a position
given its preceding tokens, and the output is the token with the maximum probability. The standard way to train this model, as in the language modeling problem, is to minimize a loss function
over a collection of token sequences. Let Decoderθ (·) denote a decoder with parameters θ. At
each position i, the decoder generates a distribution of the next tokens based on its preceding
tokens {x0 , ..., xi }, denoted by Prθ (·|x0 , ..., xi ) (or pθi+1 for short). Suppose we have the goldstandard distribution at the same position, denoted by pgold
i+1 . For language modeling, we can think
gold
of pi+1 as a one-hot representation of the correct predicted word. We then define a loss function
L(pθi+1 , pgold
i+1 ) to measure the difference between the model prediction and the true prediction. In
NLP, the log-scale cross-entropy loss is typically used.
Given a sequence of m tokens {x0 , ..., xm }, the loss on this sequence is the sum of the loss

Pre-training

8
over the positions {0, ..., m − 1}, given by
Lossθ (x0 , ..., xm ) =
=

m−1
X

i=0
m−1
X

L(pθi+1 , pgold
i+1 )
LogCrossEntropy(pθi+1 , pgold
i+1 )

(1.5)

i=0

where LogCrossEntropy(·) is the log-scale cross-entropy, and pgold
i+1 is the one-hot representation
of xi+1 .
This loss function can be extended to a set of sequences D. In this case, the objective of
pre-training is to find the best parameters that minimize the loss on D
θ̂ = arg min
θ

X

Lossθ (x)

(1.6)

x∈D

Note that this objective is mathematically equivalent to maximum likelihood estimation, and can
be re-expressed as
θ̂ = arg max
θ

= arg max
θ

X

log Prθ (x)

x∈D
i−1
XX

x∈D i=0

log Prθ (xi+1 |x0 , ..., xi )

(1.7)

With these optimized parameters θ̂, we can use the pre-trained language model Decoderθ̂ (·)
to compute the probability Prθ̂ (xi+1 |x0 , ..., xi ) at each position of a given sequence.

1.2.2 Encoder-only Pre-training
As defined in Section 1.1.2.1, an encoder Encoderθ (·) is a function that reads a sequence of
tokens x = x0 ...xm and produces a sequence of vectors H = h0 ...hm 4 . Training this model is
not straightforward, as we do not have gold-standard data for measuring how good the output of
the real-valued function is. A typical approach to encoder pre-training is to combine the encoder
with some output layers to receive supervision signals that are easier to obtain. Figure 1.2 shows
a common architecture for pre-training Transformer encoders, where we add a Softmax layer on
top of the Transformer encoder. Clearly, this architecture is the same as that of the decoder-based
language model, and the output is a sequence of probability distributions




pW,θ
 1. 
 .  = SoftmaxW (Encoderθ (x))
 . 

(1.9)

pW,θ
m

4

If we view hi as a row vector, H can be written as

H

=





h0
 .. 
 . 
hm

(1.8)

1.2 Self-supervised Pre-training Tasks

9

Self-supervision
E.g., evaluate how well the
model reconstructs the masked token

Output for Downstream Tasks

Softmax

Prediction Network

Encoder

Pre-trained Encoder

e0

e1

e2

x0

x1

x2

e3

e4

e0

e1

e2

e3

e4

x3

x4

x0

x1

x2

x3

x4

(masked)

(a) Pre-training

(b) Applying the Pre-trained Encoder

Fig. 1.2: Pre-training a Transformer encoder (left) and then applying the pre-trained encoder (right). In the pre-training
phase, the encoder, together with a Softmax layer, is trained via self-supervision. In the application phase, the Softmax
layer is removed, and the pre-trained encoder is combined with a prediction network to address specific problems. In
general, for better adaptation to these tasks, the system is fine-tuned using labeled data.

Here pW,θ
is the output distribution Pr(·|x) at position i. We use SoftmaxW (·) to denote that
i
the Softmax layer is parameterized by W, that is, SoftmaxW (H) = Softmax(H · W). For
notation simplicity, we will sometimes drop the superscripts W and θ affixed to each probability
distribution.
The difference between this model and standard language models is that the output pi has
different meanings in encoder pre-training and language modeling. In language modeling, pi is
the probability distribution of predicting the next word. This follows an auto-regressive decoding
process: a language model only observes the words up to position i and predicts the next. By
contrast, in encoder pre-training, the entire sequence can be observed at once, and so it makes no
sense to predict any of the tokens in this sequence.

1.2.2.1 Masked Language Modeling
One of the most popular methods of encoder pre-training is masked language modeling, which
forms the basis of the well-known BERT model [Devlin et al., 2019]. The idea of masked language modeling is to create prediction challenges by masking out some of the tokens in the input
sequence and training a model to predict the masked tokens. In this sense, the conventional language modeling problem, which is sometimes called causal language modeling, is a special case
of masked language modeling: at each position, we mask the tokens in the right-context, and predict the token at this position using its left context. However, in causal language modeling we
only make use of the left-context in word prediction, while the prediction may depend on tokens
in the right-context. By contrast, in masked language modeling, all the unmasked tokens are used
for word prediction, leading to a bidirectional model that makes predictions based on both left and
right-contexts.

Pre-training

10

More formally, for an input sequence x = x0 ...xm , suppose that we mask the tokens at positions A(x) = {i1 , ..., iu }. Hence we obtain a masked token sequence x̄ where the token at
each position in A(x) is replaced with a special symbol [MASK]. For example, for the following
sequence
The early bird catches the worm
we may have a masked token sequence like this
The [MASK] bird catches the [MASK]
where we mask the tokens early and worm (i.e., i1 = 2 and i2 = 6).
Now we have two sequences x and x̄. The model is then optimized so that we can correctly
predict x based on x̄. This can be thought of as an autoencoding-like process, and the training objective is to maximize the reconstruction probability Pr(x|x̄). Note that there is a simple
position-wise alignment between x and x̄. Because an unmasked token in x̄ is the same as the token in x at the same position, there is no need to consider the prediction for this unmasked token.
This leads to a simplified training objective which only maximizes the probabilities for masked
tokens. We can express this objective in a maximum likelihood estimation fashion
(Ŵ, θ̂) = arg max
W,θ

X X

log PrW,θ
(xi |x̄)
i

(1.10)

LogCrossEntropy(pW,θ
, pgold
)
i
i

(1.11)

x∈D i∈A(x)

or alternatively express it using the cross-entropy loss
c θ̂) = arg min
(W,
W,θ

X X

x∈D i∈A(x)

where PrW,θ
(xk |x̄) is the probability of the true token xk at position k given the corrupted input
k
W,θ
x, and pk is the probability distribution at position k given the corrupted input x. To illustrate,
consider the above example where two tokens of the sequence “the early bird catches the worm”
are masked. For this example, the objective is to maximize the sum of log-scale probabilities
Loss = log Pr(x2 = early|x̄ = [CLS] The [MASK] bird catches the [MASK]) +
|

{z

x̄2

}

|

{z
x̄6

}

log Pr(x6 = worm|x̄ = [CLS] The [MASK] bird catches the [MASK])
|

{z
x̄2

}

|

{z
x̄6

}

(1.12)

c and θ̂, we can drop W.
c Then, we can further
Once we obtain the optimized parameters W
fine-tune the pre-trained encoder Encoderθ̂ (·) or directly apply it to downstream tasks.

1.2.2.2 Permuted Language Modeling
While masked language modeling is simple and widely applied, it introduces new issues. One
drawback is the use of a special token, [MASK], which is employed only during training but not

1.2 Self-supervised Pre-training Tasks

11

at test time. This leads to a discrepancy between training and inference. Moreover, the autoencoding process overlooks the dependencies between masked tokens. For example, in the above
example, the prediction of x2 (i.e., the first masked token) is made independently of x6 (i.e., the
second masked token), though x6 should be considered in the context of x2 .
These issues can be addressed using the permuted language modeling approach to pretraining [Yang et al., 2019]. Similar to causal language modeling, permuted language modeling
involves making sequential predictions of tokens. However, unlike causal modeling where predictions follow the natural sequence of the text (like left-to-right or right-to-left), permuted language
modeling allows for predictions in any order. The approach is straightforward: we determine an
order for token predictions and then train the model in a standard language modeling manner, as
described in Section 1.2.1. Note that in this approach, the actual order of tokens in the text remains
unchanged, and only the order in which we predict these tokens differs from standard language
modeling. For example, consider a sequence of 5 tokens x0 x1 x2 x3 x4 . Let ei represent the embedding of xi (i.e., combination of the token embedding and positional embedding). In standard
language modeling, we would generate this sequence in the order of x0 → x1 → x2 → x3 → x4 .
The probability of the sequence can be modeled via a generation process.
Pr(x) = Pr(x0 ) · Pr(x1 |x0 ) · Pr(x2 |x0 , x1 ) · Pr(x3 |x0 , x1 , x2 ) ·
Pr(x4 |x0 , x1 , x2 , x3 )

= Pr(x0 ) · Pr(x1 |e0 ) · Pr(x2 |e0 , e1 ) · Pr(x3 |e0 , e1 , e2 ) ·
Pr(x4 |e0 , e1 , e2 , e3 )

(1.13)

Now, let us consider a different order for token prediction: x0 → x4 → x2 → x1 → x3 . The
sequence generation process can then be expressed as follows:
Pr(x) = Pr(x0 ) · Pr(x4 |e0 ) · Pr(x2 |e0 , e4 ) · Pr(x1 |e0 , e4 , e2 ) ·
Pr(x3 |e0 , e4 , e2 , e1 )

(1.14)

This new prediction order allows for the generation of some tokens to be conditioned on a
broader context, rather than being limited to just the preceding tokens as in standard language
models. For example, in generating x3 , the model considers both its left-context (i.e., e0 , e1 , e2 )
and right-context (i.e., e4 ). The embeddings e0 , e1 , e2 , e4 incorporate the positional information
of x0 , x1 , x2 , x4 , preserving the original order of the tokens. As a result, this approach is somewhat
akin to masked language modeling: we mask out x3 and use its surrounding tokens x0 , x1 , x2 , x4
to predict this token.
The implementation of permuted language models is relatively easy for Transformers. Because the self-attention model is insensitive to the order of inputs, we do not need to explicitly
reorder the sequence to have a factorization like Eq. (1.14). Instead, permutation can be done
by setting appropriate masks for self-attention. For example, consider the case of computing
Pr(x1 |e0 , e4 , e2 ). We can place x0 , x1 , x2 , x3 , x4 in order and block the attention from x3 to x1
in self-attention, as illustrated below

Pre-training

12

x0

x1

x2

x3

x4

Masks for Self-attention:
Blue box = valid attention
Gray box = blocked attention

For a more illustrative example, we compare the self-attention masking results of causal language
modeling, masked language modeling and permuted language modeling in Figure 1.3.
1.2.2.3 Pre-training Encoders as Classifiers
Another commonly-used idea to train an encoder is to consider classification tasks. In selfsupervised learning, this is typically done by creating new classification challenges from the unlabeled text. There are many different ways to design the classification tasks. Here we present two
popular tasks.
A simple method, called next sentence prediction (NSP), is presented in BERT’s original
paper [Devlin et al., 2019]. The assumption of NSP is that a good text encoder should capture
the relationship between two sentences. To model such a relationship, in NSP we can use the
output of encoding two consecutive sentences SentA and SentB to determine whether SentB is
the next sentence following SentA . For example, suppose SentA = ’It is raining .’ and SentB =
’I need an umbrella .’. The input sequence of the encoder could be
[CLS] It is raining . [SEP] I need an umbrella . [SEP]
where [CLS] is the start symbol (i.e., x0 ) which is commonly used in encoder pre-training, and
[SEP] is a separator that separates the two sentences. The processing of this sequence follows a
standard procedure of Transformer encoding: we first represent each token xi as its corresponding
embedding ei , and then feed the embedding sequence {e0 , ..., em } into the encoder to obtain the
output sequence {h0 , ..., hm }. Since h0 is generally considered as the representation of the entire
sequence, we add a Softmax layer on top of it to construct a binary classification system. This
process is illustrated as follows
token: [CLS] It
embedding:

encoding:

is raining . [SEP] I need an umbrella

.

[SEP]

e0

e1 e2

e3

e4

e5

e6

e7

e8

e9

e10

e11

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

h0

h1 h2

↓

Softmax
↓

Is Next or Not?

h3

h4

Encoder

h5

h6 h7 h8

h9

h10

h11

1.2 Self-supervised Pre-training Tasks
x0

x1

x2

x3

13
x4

x0

Pr(x0 ) = 1

x1

Pr(x1 |e0 )

x2

Pr(x2 |e0 , e1 )

x3

Pr(x3 |e0 , e1 , e2 )

x4

Pr(x4 |e0 , e1 , e2 , e3 )

(a) Causal Language Modeling (order: x0 → x1 → x2 → x3 → x4 )
x0

masked

x1

x2

masked

x3

x4

x0

1

x1

Pr(x1 |e0 , emask , e2 , emask , e4 )

x2

1

x3

Pr(x3 |e0 , emask , e2 , emask , e4 )

x4

1

masked

masked

(b) Masked Language Modeling (order: x0 , [MASK], x2 , [MASK], x4 → x1 , x3 )
x0

x1

x2

x3

x4

x0

Pr(x0 ) = 1

x1

Pr(x1 |e0 , e4 , e2 )

x2

Pr(x2 |e0 , e4 )

x3

Pr(x3 |e0 , e4 , e2 , e1 )

x4

Pr(x4 |e0 )

(c) Permuted Language Modeling (order: x0 → x4 → x2 → x1 → x3 )
Fig. 1.3: Comparison of self-attention masking results of causal language modeling, masked language modeling and
permuted language modeling. The gray cell denotes the token at position j does not attend to the token at position i.
The blue cell (i, j) denotes that the token at position j attends to the token at position i. emask represents the embedding
of the symbol [MASK], which is a combination of the token embedding and the positional embedding.

In order to generate training samples, we need two sentences each time, one for SentA and
the other for SentB . A simple way to do this is to utilize the natural sequence of two consecutive sentences in the text. For example, we obtain a positive sample by using actual consecutive
sentences, and a negative sample by using randomly sampled sentences. Consequently, training
this model is the same as training a classifier. Typically, NSP is used as an additional training loss

Pre-training

14
function for pre-training based on masked language modeling.

A second example of training Transformer encoders as classifiers is to apply classificationbased supervision signals to each output of an encoder. For example, Clark et al. [2019] in their
ELECTRA model, propose training a Transformer encoder to identify whether each input token
is identical to the original input or has been altered in some manner. The first step of this method
is to generate a new sequence from a given sequence of tokens, where some of the tokens are
altered. To do this, a small masked language model (call it the generator) is applied: we randomly
mask some of the tokens, and train this model to predict the masked tokens. For each training
sample, this masked language model outputs a token at each masked position, which might be
different from the original token. At the same time, we train another Transformer encoder (call it
the discriminator) to determine whether each predicted token is the same as the original token or
altered. More specifically, we use the generator to generate a sequence where some of the tokens
are replaced. Below is an illustration.

original:
masked:

replaced:

[CLS]

The

boy

spent

hours

working

on

toys

.

↓

↓

↓

↓

↓

↓

↓

↓

↓

[CLS]

The

boy

spent

[MASK]

working

on

[MASK]

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

[CLS]

The

Generator (small masked language model)
boy

spent

decades

working

on

.

toys

.

Then, we use the discriminator to label each of these tokens as orginal or replaced, as follows

replaced: [CLS]

spent decades working

The

boy

↓

↓

↓

↓

↓

↓

↓

↓

↓

↓

on

toys

.

↓

↓

↓

↓

↓

↓

↓

↓

Discriminator (the model we want)

label: original original original original replaced original original original original

For training, the generator is optimized as a masked language model with maximum likelihood
estimation, and the discriminator is optimized as a classifier using a classification-based loss. In
ELECTRA, the maximum likelihood-based loss and the classification-based loss are combined for
jointly training both the generator and discriminator. An alternative approach is to use generative
adversarial networks (GANs), that is, the generator is trained to fool the discriminator, and the discriminator is trained to distinguish the output of the generator from the true distribution. However,
GAN-style training complicates the training task and is more difficult to scale up. Nevertheless,
once training is complete, the generator is discarded, and the encoding part of the discriminator is
applied as the pre-trained model for downstream tasks.

1.2 Self-supervised Pre-training Tasks

15

1.2.3 Encoder-Decoder Pre-training
In NLP, encoder-decoder architectures are often used to model sequence-to-sequence problems,
such as machine translation and question answering. In addition to these typical sequence-tosequence problems in NLP, encoder-decoder models can be extended to deal with many other
problems. A simple idea is to consider text as both the input and output of a problem, and so
we can directly apply encoder-decoder models. For example, given a text, we can ask a model to
output a text describing the sentiment of the input text, such as positive, negative, and neutral.
Such an idea allows us to develop a single text-to-text system to address any NLP problem.
We can formulate different problems into the same text-to-text format. We first train an encoderdecoder model to gain general-purpose knowledge of language via self-supervision. This model
is then fine-tuned for specific downstream tasks using targeted text-to-text data.

1.2.3.1 Masked Encoder-Decoder Pre-training
In Raffel et al. [2020]’s T5 model, many different tasks are framed as the same text-to-text task.
Each sample in T5 follows the format
Source Text → Target Text
Here → separates the source text, which consists of a task description or instruction and the input
given to the system, from the target text, which is the response to the input task. As an example,
consider a task of translating from Chinese to English. A training sample can be expressed as
[CLS] Translate from Chinese to English: 你好！ → hsi Hello!
where [CLS] and hsi are the start symbols on the source and target sides, respectively5 .
Likewise, we can express other tasks in the same way. For example
[CLS] Answer: when was Albert Einstein born?
→ hsi He was born on March 14, 1879.
[CLS] Simplify: the professor, who has has published numerous papers in his field,
will be giving a lecture on the topic next week.
→ hsi The experienced professor will give a lecture next week.
[CLS] Score the translation from English to Chinese. English: when in Rome, do as
the Romans do. Chinese: 人 在 罗马 就 像 罗马 人 一样 做事 。
→ hsi 0.81
where instructions are highlighted in gray. An interesting case is that in the last example we
5
We could use the same start symbol for different sequences. Here we use different symbols to distinguish the
sequences on the encoder and decoder-sides.

Pre-training

16

reframe the scoring problem as the text generation problem. Our goal is to generate a text representing the number 0.81, rather than outputting it as a numerical value.
The approach described above provides a new framework of universal language understanding
and generation. Both the task instructions and the problem inputs are provided to the system
in text form. The system then follows the instructions to complete the task. This method puts
different problems together, with the benefit of training a single model that can perform many
tasks simultaneously.
In general, fine-tuning is necessary for adapting the pre-trained model to a specific downstream
task. In this process, one can use different ways to instruct the model for the task, such as using a
short name of the task as the prefix to the actual input sequence or providing a detailed description
of the task. Since the task instructions are expressed in text form and involved as part of the input,
the general knowledge of instruction can be gained through learning the language understanding
models in the pre-training phase. This may help enable zero-shot learning. For example, pretrained models can generalize to address new problems where the task instructions have never
been encountered.
There have been several powerful methods of self-supervised learning for either Transformer
encoders or decoders. Applying these methods to pre-train encoder-decoder models is relatively
straightforward. One common choice is to train encoder-decoder models as language models. For
example, the encoder receives a sequence prefix, while the decoder generates the remaining sequence. However, this differs from standard causal language modeling, where the entire sequence
is autoregressively generated from the first token. In our case, the encoder processes the prefix at
once, and then the decoder predicts subsequent tokens in the manner of causal language modeling.
Put more precisely, this is a prefix language modeling problem: a language model predicts the
subsequent sequence given a prefix, which serves as the context for prediction.
Consider the following example
[CLS] The puppies are frolicking → hsi outside the house .
|

{z

Prefix

}

|

{z

Subsequent Sequence

}

We can directly train an encoder-decoder model using examples like this. Then, the encoder learns
to understand the prefix, and the decoder learns to continue writing based on this understanding.
For large-scale pre-training, it is easy to create a large number of training examples from unlabeled
text.
It is worth noting that for pre-trained encoder-decoder models to be effective in multi-lingual
and cross-lingual tasks, such as machine translation, they should be trained with multi-lingual
data. This typically requires that the vocabulary includes tokens from all the languages. By
doing so, the models can learn shared representations across different languages, thereby enabling
capabilities in both language understanding and generation in a multi-lingual and cross-lingual
context.
A second approach to pre-training encoder-decoder models is masked language modeling. In
this approach, as discussed in Section 1.2.2, tokens in a sequence are randomly replaced with a
mask symbol, and the model is then trained to predict these masked tokens based on the entire
masked sequence.
As an illustration, consider the task of masking and reconstructing the sentence

1.2 Self-supervised Pre-training Tasks

17

The puppies are frolicking outside the house .

By masking two tokens (say, frolicking and the), we have the BERT-style input and output of the
model, as follows
[CLS] The puppies are [MASK] outside [MASK] house .
→ hsi

frolicking

the

Here
denotes the masked position at which we do not make token predictions. By varying the
percentage of the tokens in the text, this approach can be generalized towards either BERT-style
training or language modeling-style training [Song et al., 2019]. For example, if we mask out all
the tokens, then the model is trained to generate the entire sequence
[CLS] [MASK] [MASK] [MASK] [MASK] [MASK] [MASK] [MASK] [MASK]
→ hsi The puppies are frolicking outside the house .
In this case, we train the decoder as a language model.
Note that, in the context of the encoder-decoder architecture, we can use the encoder to read
the masked sequence, and use the decoder to predict the original sequence. With this objective,
we essentially have a denoising autoencoder: the encoder transforms a corrupted input into some
hidden representation, and the decoder reconstructs the uncorrupted input from this hidden representation. Here is an example of input and output for denoising training.

[CLS] The puppies are [MASK] outside [MASK] house .
→ hsi The puppies are frolicking outside the house .
By learning to map from this corrupted sequence to its uncorrupted counterpart, the model gains
the ability to understand on the encoder side and to generate on the decoder side. See Figure 1.4
for an illustration of how an encoder-decoder model is trained with BERT-style and denoising
autoencoding objectives.
As we randomly select tokens for masking, we can certainly mask consecutive tokens [Joshi et al.,
2020]. Here is an example.
[CLS] The puppies are [MASK] outside [MASK] [MASK] .
→ hsi The puppies are frolicking outside the house .
Another way to consider consecutive masked tokens is to represent them as spans. Here we
follow Raffel et al. [2020]’s work, and use [X], [Y] and [Z] to denote sentinel tokens that cover
one or more consecutive masked tokens. Using this notation, we can re-express the above training
example as
[CLS] The puppies are [X] outside [Y] .
→ hsi [X] frolicking [Y] the house [Z]

Pre-training

18
Loss
[M]

[M]

Encoder
[CLS] The puppies are

[M]

the

[M]

[M]

[M] frolicking [M]

the

[M]

[M]frolicking [M]

Decoder
in

[M] house

.

hsi

[M]

[M]

(a) Training an encoder-decoder model with BERT-style masked language modeling

Loss over the sequence
the

house

.

The puppies are frolicking in

the

house

The puppies are frolicking in

Encoder
[CLS] The puppies are

[M]

Decoder
in

[M] house

.

hsi

(b) Training an encoder-decoder model with denoising autoencoding
Fig. 1.4: Training an encoder-decoder model using BERT-style and denoising autoencoding methods. In both methods,
the input to the encoder is a corrupted token sequence where some tokens are masked and replaced with [MASK] (or
[M] for short). The decoder predicts these masked tokens, but in different ways. In BERT-style training, the decoder
only needs to compute the loss for the masked tokens, while the remaining tokens in the sequence can be simply treated
as [MASK] tokens. In denoising autoencoding, the decoder predicts the sequence of all tokens in an autoregressive
manner. As a result, the loss is obtained by accumulating the losses of all these tokens, as in standard language
modeling.

The idea is that we represent the corrupted sequence as a sequence containing placeholder
slots. The training task is to fill these slots with the correct tokens using the surrounding context.
An advantage of this approach is that the sequences used in training would be shorter, making the
training more efficient. Note that masked language modeling provides a very general framework
for training encoder-decoder models. Various settings can be adjusted to have different training
versions, such as altering the percentage of tokens masked and the maximum length of the masked
spans.

1.2.3.2 Denoising Training
If we view the problem of training encoder-decoder models as a problem of training denoising
autoencoders, there will typically be many different methods for introducing input corruption and
reconstructing the input. For instance, beyond randomly masking tokens, we can also alter some
of them or rearrange their order.
Suppose we have an encoder-decoder model that can map an input sequence x to an output

1.2 Self-supervised Pre-training Tasks

19

sequence y
y = Decodeω (Encodeθ (x))
= Modelθ,ω (x)

(1.15)

where θ and ω are the parameters of the encoder and the decoder, respectively. In denoising
autoencoding problems, we add some noise to x to obtain a noisy, corrupted input xnoise . By
feeding xnoise into the encoder, we wish the decoder to output the original input. The training
objective can be defined as
(θ̂, ω̂) = arg max Loss(Modelθ,ω (xnoise ), x)

(1.16)

θ,ω

Here the loss function Loss(Modelθ,ω (xnoise ), x) evaluates how well the model Modelθ,ω (xnoise )
reconstructs the original input x. We can choose the cross-entropy loss as usual.
As the model architecture and the training approach have been developed, the remaining issue
is the corruption of the input. Lewis et al. [2020], in their BART model, propose corrupting the
input sequence in several different ways.
• Token Masking. This is the same masking method that we used in masked language modeling. The tokens in the input sequence are randomly selected and masked.
• Token Deletion. This method is similar to token masking. However, rather than replacing
the selected tokens with a special symbol [MASK], these tokens are removed from the
sequence. See the following example for a comparison of the token masking and token
deletion methods.
Original (x):

The puppies are frolicking outside the house .

Token Masking (xnoise ):

The puppies are [MASK] outside [MASK] house .

Token Deletion (xnoise ):

The puppies are frolicking outside the house .

where the underlined tokens in the original sequence are masked or deleted.
• Span Masking. Non-overlapping spans are randomly sampled over the sequence. Each
span is masked by [MASK]. We also consider spans of length 0, and, in such cases, [MASK]
is simply inserted at a position in the sequence. For example, we can use span masking to
corrupt the above sequence as
Original (x):
Span Masking (xnoise ):

The 0 puppies are frolicking outside the house .
The [MASK] puppies are [MASK] house .

Here the span frolicking outside the is replaced with a single [MASK]. 0 indicates a length0 span, and so we insert an [MASK] between The and puppies. Span masking introduces
new prediction challenges in which the model needs to know how many tokens are generated from a span. This problem is very similar to fertility modeling in machine translation
[Brown et al., 1993].

Pre-training

20

If we consider a sequence consisting of multiple sentences, additional methods of corruption
can be applied. In the BART model, there are two such methods.
• Sentence Reordering. This method randomly permutes the sentences so that the model can
learn to reorder sentences in a document. Consider, for example, two consecutive sentences
Hard work leads to success . Success brings happiness .
We can reorder the two sentences to have a corrupted input sequence
Success brings happiness . Hard work leads to success .
• Document Rotation. The goal of this task is to identify the start token of the sequence.
First, a token is randomly selected from the sequence. Then, the sequence is rotated so that
the selected token is the first token. For example, suppose we select the token leads from
the above sequence. The rotated sequence is
selected

Hard work leads to success . Success brings happiness . Hard work

where the subsequence Hard work before leads is appended to the end of the sequence.
For pre-training, we can apply multiple corruption methods to learn robust models, for example, we randomly choose one of them for each training sample. In practice, the outcome of
encoder-decoder pre-training depends heavily on the input corruption methods used, and so we
typically need to choose appropriate training objectives through careful experimentation.

1.2.4 Comparison of Pre-training Tasks
So far, we have discussed a number of pre-training tasks. Since the same training objective can
apply to different architectures (e.g., using masked language modeling for both encoder-only and
encoder-decoder pre-training), categorizing pre-training tasks based solely on model architecture
does not seem ideal. Instead, we summarize these tasks based on the training objectives.
• Language Modeling. Typically, this approach refers to an auto-regressive generation procedure of sequences. At one time, it predicts the next token based on its previous context.
• Masked Language Modeling. Masked Language Modeling belongs to a general maskpredict framework. It randomly masks tokens in a sequence and predicts these tokens using
the entire masked sequence.

1.3 Example: BERT

21

• Permuted Language Modeling. Permuted language modeling follows a similar idea to
masked language modeling, but considers the order of (masked) token prediction. It reorders
the input sequence and predicts the tokens sequentially. Each prediction is based on some
context tokens that are randomly selected.
• Discriminative Training. In discriminative training, supervision signals are created from
classification tasks. Models for pre-training are integrated into classifiers and trained together with the remaining parts of the classifiers to enhance their classification performance.
• Denoising Autoencoding. This approach is applied to the pre-training of encoder-decoder
models. The input is a corrupted sequence and the encoder-decoder models are trained to
reconstruct the original sequence.
Table 1.1 illustrates these methods and their variants using examples. The use of these examples does not distinguish between models, but we mark the model architectures where the
pre-training tasks can be applied. In each example, the input consists of a token sequence, and the
output is either a token sequence or some probabilities. For generation tasks, such as language
modeling, superscripts are used to indicate the generation order on the target side. If the superscripts are omitted, it indicates that the output sequence can be generated either autoregressively
or simultaneously. On the source side, we assume that the sequence undergoes a standard Transformer encoding process, meaning that each token can see the entire sequence in self-attention.
The only exception is in permuted language modeling, where an autoregressive generation process is implemented by setting attention masks on the encoder side. To simplify the discussion,
we remove the token hsi from the target-side of each example.

While these pre-training tasks are different, it is possible to compare them in the same framework and experimental setup [Dong et al., 2019; Raffel et al., 2020; Lewis et al., 2020]. Note that
we cannot list all the pre-training tasks here as there are many of them. For more discussions on
pre-training tasks, the interested reader may refer to some surveys on this topic [Qiu et al., 2020;
Han et al., 2021].

1.3 Example: BERT
In this section, we introduce BERT models, which are among the most popular and widely used
pre-trained sequence encoding models in NLP.

1.3.1 The Standard Model
The standard BERT model, which is proposed in Devlin et al. [2019]’s work, is a Transformer
encoder trained using both masked language modeling and next sentence prediction tasks. The
loss used in training this model is a sum of the loss of the two tasks.
LossBERT = LossMLM + LossNSP

(1.17)

As is regular in training deep neural networks, we optimize the model parameters by minimizing
this loss. To do this, a number of training samples are collected. During training, a batch of

Pre-training

22
Method Enc Dec E-D Input
Causal LM

MASS-style
BERT-style

The1 kitten2 is3 chasing4 the5 ball6 .7

•

•

•

•

[C] The kitten is

•

•

[C] The kitten [M] chasing the [M] .

is

•

•

[C] The kitten [M] [M] [M] ball .

is chasing the

•

•

[C] The kitten [M] playing the [M] .

kitten is chasing

Prefix LM
Masked LM

Output
chasing1 the2 ball3 .4
ball
ball

Permuted LM

•

[C] The kitten is chasing the ball .

The5 kitten7 is6 chasing1 the4 ball2 .3

Next Sentence

•

[C] The kitten is chasing the ball .

Pr(IsNext | representation-of-[C])

•

Encode a sentence as ha and

•

[C] The kitten is chasing the ball .

Pr(·|The) Pr(·|kitten) ... Pr(·|.)

[C] . kitten the chasing The is ball

The1 kitten2 is3 chasing4 the5 ball6 .7

[C] The kitten is chasing the ball .

The1 kitten2 is3 chasing4 the5 ball6 .7

[C] The kitten [M] is [M] .

The1 kitten2 is3 chasing4 the5 ball6 .7

[C] The kitten [X] the [Y]

[X] is2 chasing3 [Y] ball5 .6

Prediction
Sentence
Comparison
Token Classification
Token Reordering
Token Deletion
Span Masking
Sentinel Masking
Sentence
Reordering
Document
Rotation

Birds eat worms .

Score(ha , hb )

another sentence as hb

•
•
•
•

1

4

•

[C] The ball rolls away swiftly . The The1 kitten2 is3 chasing4 the5 ball6 .7
kitten is chasing the ball .

The8 ball9 rolls10 away11 swiftly12 .13

•

[C] chasing the ball . The ball rolls

The1 kitten2 is3 chasing4 the5 ball6 .7

away swiftly . The kitten is

The8 ball9 rolls10 away11 swiftly12 .13

Table 1.1: Comparison of pre-training tasks, including language modeling, masked language modeling, permuted
language modeling, discriminative training, and denoising autoencoding. [C] = [CLS], [M] = [MASK], [X], [Y] =
sentinel tokens. Enc, Dec and E-D indicate whether the approach can be applied to encoder-only, decoder-only,
encoder-decoder models, respectively. For generation tasks, superscripts are used to represent the order of the tokens.

training samples is randomly selected from this collection at a time, and LossBERT is accumulated
over these training samples. Then, the model parameters are updated via gradient descent or its
variants. This process is repeated many times until some stopping criterion is satisfied, such as
when the training loss converges.

1.3.1.1 Loss Functions
In general, BERT models are used to represent a single sentence or a pair of sentences, and thus
can handle various downstream language understanding problems. In this section we assume that
the input representation is a sequence containing two sentences SentA and SentB , expressed as
[CLS] SentA [SEP] SentB [SEP]
Here we follow the notation in BERT’s paper and use [SEP] to denote the separator.
Given this sequence, we can obtain LossMLM and LossNSP separately. For masked language
modeling, we predict a subset of the tokens in the sequence. Typically, a certain percentage of the

1.3 Example: BERT

23

tokens are randomly selected, for example, in the standard BERT model, 15% of the tokens in
each sequence are selected. Then the sequence is modified in three ways

• Token Masking. 80% of the selected tokens are masked and replaced with the symbol
[MASK]. For example
Original:

[CLS] It is raining . [SEP] I need an umbrella . [SEP]

Masked:

[CLS] It is [MASK] . [SEP] I need [MASK] umbrella . [SEP]

where the selected tokens are underlined. Predicting masked tokens makes the model learn
to represent tokens from their surrounding context.
• Random Replacement. 10% of the selected tokens are changed to a random token. For
example,
[CLS] It is raining . [SEP] I need an umbrella . [SEP]

Original:

[CLS] It is raining . [SEP] I need an hat . [SEP]

Random Token:

This helps the model learn to recover a token from a noisy input.
• Unchanged. 10% of the selected tokens are kept unchanged. For example,
Original:

[CLS] It is raining . [SEP] I need an umbrella . [SEP]

Unchanged Token:

[CLS] It is raining . [SEP] I need an umbrella . [SEP]

This is not a difficult prediction task, but can guide the model to use easier evidence for
prediction.
Let A(x) be the set of selected positions of a given token sequence x, and x̄ be the modified
sequence of x. The loss function of masked language modeling can be defined as
LossMLM = −

X

i∈A(x)

log Pri (xi |x̄)

(1.18)

where Pri (xi |x̄) is the probability of predicting xi at the position i given x̄. Figure 1.5 shows a
running example of computing LossMLM .
For next sentence prediction, we follow the method described in Section 1.2.2.3. Each training
sample is classified into a label set {IsNext, NotNext}, for example,
Sequence:
Label:
Sequence:
Label:

[CLS] It is raining . [SEP] I need an umbrella . [SEP]
IsNext
[CLS] The cat sleeps on the windowsill . [SEP] Apples grow on trees . [SEP]
NotNext

Pre-training

24
Input: [CLS] It is raining . [SEP] I need an umbrella . [SEP]
Select tokens with a probability of 15%

Token Selection: [CLS] It is raining . [SEP] I need an umbrella . [SEP]
Mask selected tokens with a probability of 80%

Token Masking: [CLS] It is [MASK] . [SEP] I need [MASK] umbrella . [SEP]
Alter selected tokens with a probability of 10%

Token: [CLS] It is [MASK] . [SEP] I need [MASK] hat . [SEP]
Replacement
Keep selected tokens unchanged with a probability of 10%

Unchanged: [CLS] It is [MASK] . [SEP] I need [MASK] hat . [SEP]
Train the Transformer encoder with the modified sequence

training

h0

h1

h2

h3

an umbrella

I

h4

h5

h6

h7

h8

h9

h10

h11

e8

Transformer Encoder
e0

e1

e2

e3

[CLS]

It

is [MASK]

e4

e5

e6

e7

e9

e10

e11

.

[SEP]

I

need [MASK] hat

.

[SEP]

Fig. 1.5: A running example of BERT-style masked language modeling. First, 15% tokens are randomly selected.
These selected tokens are then processed in one of three ways: replaced with a [MASK] token (80% of the time),
replaced with a random token (10% of the time), or kept unchanged (10% of the time). The model is trained to predict
these selected tokens based on the modified sequence. ei represents the embedding of the token at the position i. Gray
boxes represent the Softmax layers.

The output vector of the encoder for the first token [CLS] is viewed as the sequence representation,
denoted by hcls (or h0 ). A classifier is built on top of hcls . Then, we can compute the probability of
a label c given hcls , i.e., Pr(c|hcls ). There are many loss functions one can choose for classification
problems. For example, in maximum likelihood training, we can define LossNSP as
LossNSP = − log Pr(cgold |hcls )

(1.19)

1.3 Example: BERT

25

where cgold is the correct label for this sample.
1.3.1.2 Model Setup
As shown in Figure 1.6, BERT models are based on the standard Transformer encoder architecture.
The input is a sequence of embeddings, each being the sum of the token embedding, the positional
embedding, and the segment embedding.
(1.20)

e = x + epos + eseg

Both the token embedding (x) and positional embedding (epos ) are regular, as in Transformer
models. The segment embedding (eseg ) is a new type of embedding that indicates whether a token
belongs to SentA or SentB . This can be illustrated by the following example.
Token [CLS]

It

is

raining

.

[SEP]

I

need

an

umbrella

.

[SEP]

x

x1

x2

x3

x4

x5

x6

x7

x8

x9

x10

x11

x0

epos PE(0) PE(1) PE(2) PE(3) PE(4) PE(5) PE(6) PE(7) PE(8) PE(9) PE(10) PE(11)
eseg

eA

eA

eA

eA

eA

eA

eB

eB

eB

eB

eB

eB

The main part of BERT models is a multi-layer Transformer network. A Transformer layer
consists of a self-attention sub-layer and an FFN sub-layer. Both of them follow the post-norm
architecture: output = LNorm(F (input) + input), where F (·) is the core function of the sublayer (either a self-attention model or an FFN), and LNorm(·) is the layer normalization unit.
Typically, a number of Transformer layers are stacked to form a deep network. At each position of
the sequence, the output representation is a real-valued vector which is produced by the last layer
of the network.
There are several aspects one may consider in developing BERT models.
• Vocabulary Size (|V |). In Transformers, each input token is represented as an entry in a
vocabulary V . Large vocabularies can cover more surface form variants of words, but may
lead to increased storage requirements.
• Embedding Size (de ). Every token is represented as a de -dimensional real-valued vector.
As presented above, this vector is the sum of the token embedding, positional embedding,
and segment embedding, all of which are also de -dimensional real-valued vectors.
• Hidden Size (d). The input and output of a sub-layer are of d dimensions. Besides, most
of the hidden states of a sub-layer are d-dimensional vectors. In general, d can be roughly
viewed as the width of the network.
• Number of Heads (nhead ). In self-attention sub-layers, one needs to specify the number
of heads used in multi-head self-attention. The larger this number is, the more sub-spaces
attention is performed. In practical systems, we often set nhead ≥ 4.
• FFN Hidden Size (dffn ). The size of the hidden layer of the FFNs used in Transformers is
typically larger than d. For example, a typical setting is dffn = 4d. For larger Transformers,
such as recent large models, dffn may be set to a very large value.

Pre-training

26
...

Output Layer
h0 h1 ...hm

Encoder Output
hi ∈ Rd is the contextual
representation of xi

Layer Normalization

FFN

FFN Sub-layer
hidden size: d
FFN hidden size: dffn

layers
Layer Normalization

Self-attention

Self-attention Sub-layer
hidden size: d
number of heads: nhead
Embedding

Token

Position Segment

x0 x1 ...xm

e = x + epos + eseg ∈ Rde
Input
xi corresponds to an entry of V

Fig. 1.6: The model Architecture of BERT (Transformer encoder). The input tokens are first represented as embeddings, each of which is the sum of the corresponding token embedding, positional embedding and segment embedding.
Then, the embedding sequence is processed by a stack of Transformer layers. Each layer in this stack includes a selfattention sub-layer and a FFN sub-layer. The output of the BERT model is a sequence of vectors produced by the final
Transformer layer.

• Model Depth (L). Using deep networks is an effective way to improve the expressive power
of Transformers. For BERT models, L is typically set to 12 or 24. However, networks with
even greater depth are also feasible and can be applied for further enhancements.
Different settings of these hyper-parameters lead to different model sizes. There are two
widely-used BERT models.
• BERTbase: d = 768, L = 12, nhead = 12, total number of parameters = 110M.
• BERTlarge: d = 1, 024, L = 14, nhead = 16, total number of parameters = 340M.
Training BERT models follows the standard training process of Transformers. Training larger
models such as BERTlarge requires more training effort and time. This is a common problem
for pre-training, especially when a model is trained on a very large amount of data. In practice,

1.3 Example: BERT

27

there are often considerations of training efficiency. For example, a practice is to first train a
BERT model on relatively short sequences for a large number of training steps, and then continue
training it on full-length sequences for the remaining training steps.

1.3.2 More Training and Larger Models
BERT is a milestone model in NLP, sparking many subsequent efforts to improve it. One direction
is to scale up the model itself, including increasing training data and developing larger models.
RoBERTa, an extension of the standard BERT model, is an example of such efforts [Liu et al.,
2019]. It introduces two major improvements. First, simply using more training data and more
compute can improve BERT models without need of changing the model architectures. Second,
removing the NSP loss does not decrease the performance on downstream tasks if the training is
scaled up. These findings suggest exploring a general direction of pre-training: we can continue
to improve pre-training by scaling it up on simple pre-training tasks.
A second approach to improving BERT models is to increase the number of model parameters. For example, in He et al. [2021]’s work, a 1.5 billion-parameter BERT-like model is built by
increasing both the model depth and hidden size. However, scaling up BERT and various other
pre-trained models introduces new challenges in training, for example, training very large models
often becomes unstable and difficult to converge. This makes the problem more complicated, and
requires careful consideration of various aspects, including model architecture, parallel computation, parameter initialization, and so on. In another example, Shoeybi et al. [2019] successfully
trained a 3.9 billion-parameter BERT-like model, where hundreds of GPUs were used to manage
the increased computational demands.

1.3.3 More Efficient Models
Compared to its predecessors, BERT is a relatively large model for the time it was proposed.
This increase in model size results in larger memory requirements and a consequent slowdown in
system performance. Developing smaller and faster BERT models is part of the broader challenge
of building efficient Transformers, which has been extensively discussed in Tay et al. [2020]’s
work and Xiao and Zhu [2023]’s work. However, a deeper discussion of this general topic is
beyond the scope of our current discussion. Here we instead consider a few efficient variants of
BERT.
Several threads of research are of interest to NLP researchers in developing efficient BERT
models. First, work on knowledge distillation, such as training student models with the output
of well-trained teacher models, shows that smaller BERT models can be obtained by transferring
knowledge from larger BERT models. Given that BERT models are multi-layer networks with
several different types of layers, knowledge distillation can be applied at different levels of representation. For example, beyond distilling knowledge from the output layers, it is also possible to
incorporate training loss that measures the difference in output of hidden layers between teacher
models and student models [Sun et al., 2020; Jiao et al., 2020]. Indeed, knowledge distillation has
been one of the most widely-used techniques for learning small pre-trained models.
Second, conventional model compression methods can be directly applied to compress BERT
models. One common approach is to use general-purpose pruning methods to prune the Transformer encoding networks [Gale et al., 2019]. This generally involves removing entire layers

Pre-training

28

[Fan et al., 2019] or a certain percentage of parameters in the networks [Sanh et al., 2020; Chen et al.,
2020]. Pruning is also applicable to multi-head attention models. For example, Michel et al.
[2019] show that removing some of the heads does not significantly decrease the performance
of BERT models, but speeds up the inference of these models. Another approach to compressing BERT models is quantization [Shen et al., 2020]. By representing model parameters as lowprecision numbers, the models can be greatly compressed. While this method is not specific to
BERT models, it proves effective for large Transformer-based architectures.
Third, considering that BERT models are relatively deep and large networks, another thread
of research uses dynamic networks to adapt these models for efficient inference. An idea in this
paradigm is to dynamically choose the layers for processing a token, for example, in depthadaptive models we exit at some optimal depth and thus skip the rest of the layers in the layer
stack [Xin et al., 2020; Zhou et al., 2020]. Similarly, we can develop length-adaptive models in
which the length of the input sequence is dynamically adjusted. For example, we can skip some of
the tokens in the input sequence so that the model can reduce computational load on less important
tokens, enhancing overall efficiency.
Fourth, it is also possible to share parameters across layers to reduce the size of BERT models.
A simple way to do this is to share the parameters of a whole Transformer layer across the layer
stack [Dehghani et al., 2018; Lan et al., 2020]. In addition to the reduced number of parameters,
this enables reuse of the same layer in a multi-layer Transformer network, leading to savings of
memory footprint at test time.

1.3.4 Multi-lingual Models
The initial BERT model was primarily focused on English. Soon after this model was proposed,
it was extended to many languages. One simple way to do this is to develop a separate model
for each language. Another approach, which has become more popular in recent work on large
language models, is to train multi-lingual models directly on data from all the languages. In
response, multi-lingual BERT (mBERT) models were developed by training them on text from
104 languages 6 . The primary difference from monolingual BERT models is that mBERT models
use larger vocabularies to cover tokens from multiple languages. As a result, the representations
of tokens from different languages are mapped into the same space, allowing for the sharing of
knowledge across languages via this universal representation model.
One important application of multi-lingual pre-trained models is cross-lingual learning. In the
cross-lingual setting, we learn a model on tasks in one language, and apply it to the same tasks
in another language. In cross-lingual text classification, for example, we fine-tune a multi-lingual
pre-trained model on English annotated documents. Then, we use the fine-tuned model to classify
Chinese documents.
An improvement to multi-lingual pre-trained models like mBERT is to introduce bilingual data
into pre-training. Rather than training solely on monolingual data from multiple languages, bilingual training explicitly models the relationship between tokens in two languages. The resulting
model will have innate cross-lingual transfer abilities, and thus can be easily adapted to different
languages. Lample and Conneau [2019] propose an approach to pre-training cross-lingual language models (XLMs). In their work, a cross-lingual language model can be trained in either the
causal language modeling or masked language modeling manner. For masked language modeling
6

https://github.com/google-research/bert/

1.3 Example: BERT

29

pre-training, the model is treated as an encoder. The training objective is the same as BERT: we
maximize the probabilities of some randomly selected tokens which are either masked, replaced
with random tokens, or kept unchanged in the input. If we consider bilingual data in pre-training,
we sample a pair of aligned sentences each time. Then, the two sentences are packed together to
form a single sequence used for training. For example, consider an English-Chinese sentence pair
鲸鱼 是 哺乳 动物 。 ↔ Whales are mammals .
We can pack them to obtain a sequence, like this
[CLS] 鲸鱼 是 哺乳 动物 。 [SEP] Whales are mammals . [SEP]
We then select a certain percentage of the tokens and replace them with [MASK].
[CLS] [MASK] 是 [MASK] 动物 。 [SEP] Whales [MASK] [MASK] . [SEP]
The goal of pre-training is to maximize the product of the probabilities of the masked tokens given
the above sequence. By performing training in this way, the model can learn to represent both the
English and Chinese sequences, as well as to capture the correspondences between tokens in the
two languages. For example, predicting the Chinese token 鲸鱼 may require the information
from the English token Whales. Aligning the representations of the two languages essentially
transforms the model into a “translation” model. So this training objective is also called translation language modeling. Figure 1.7 shows an illustration of this approach.
A benefit of multi-lingual pre-trained models is their inherent capability of handling codeswitching. In NLP and linguistics, code-switching refers to switching among languages in a text.
For example, the following is a mixed language text containing both Chinese and English:
周末 我们 打算 去 做 hiking ， 你 想 一起 来 吗 ？
(We plan to go hiking this weekend, would you like to join us?)

For multi-lingual pre-trained models, we do not need to identify whether a token is Chinese or
English. Instead, every token is just an entry of the shared vocabulary. This can be imagined as
creating a “new” language that encompasses all the languages we want to process.
The result of multi-lingual pre-training is influenced by several factors. Given that the model
architecture is fixed, one needs to specify the size of the shared vocabulary, the number (or percentage) of samples in each language, the size of the model, and so on. Conneau et al. [2020]
point out several interesting issues regarding large-scale multi-lingual pre-training for XLM-like
models. First, as the number of supported languages increases, a larger model is needed to handle
these languages. Second, a larger shared vocabulary is helpful for modeling the increased diversity in languages. Third, low-resource languages more easily benefit from cross-lingual transfer
from high-resource languages, particularly when similar high-resource languages are involved in
pre-training. However, interference may occur if the model is trained for an extended period,

Pre-training

30
哺乳

鲸鱼

h0

h1

h2

are mammals

h3

h4

h5

h6

h7

h8

h9

h10

h11

e8

e9

e10

e11

Transformer Encoder
e0

e1

e2

e3

e4

[CLS] [MASK]

是

[MASK]

动物

。

[SEP] Whales [MASK] [MASK]

.

[SEP]

(zh)

(zh)

(zh)

(zh)

(zh)

(zh)

(en)

(en)

(zh)

e5

e6

e7

(en)

(en)

(en)

Fig. 1.7: An illustration of translation language modeling. For ease of understanding, we present a simple example
where all the selected tokens are masked. The model is trained to predict these masked tokens. As the sequence
contains tokens in two languages, predicting a token in one language allows access to tokens in the other language,
thereby enabling cross-lingual modeling. In Lample and Conneau [2019]’s work, an input embedding (i.e., ei ) is the
sum of the token embedding, positional embedding, and language embedding. This requires that each token is assigned
with a language label. Thus we can distinguish tokens in different languages. In multi-lingual pre-training, particularly
in work using shared vocabularies, specifying the language to which a token belongs is not necessary. The use of
language embeddings in turn makes it difficult to handle code-switching. Therefore, we assume here that all token
representations are language-independent.

meaning the overall performance of the pre-trained model starts decreasing at a certain point during pre-training. Thus, in practical systems, one may need to stop the pre-training early to prevent
interference.

1.4 Applying BERT Models
Once a BERT model is pre-trained, it can then be used to solve NLP problems. But BERT models
are not immediately ready for performing specific downstream tasks. In general, additional finetuning work is required to make them adapt. As a first step, we need a predictor to align the
output of the model with the problem of interest. Let BERTθ̂ (·) be a BERT model with pretrained parameters θ̂, and Predictω (·) be a prediction network with parameters ω. By integrating
the prediction network with the output of the BERT model, we develop a model to tackle the
downstream tasks. This model can be expressed as
y = Predictω (BERTθ̂ (x))

(1.21)

where x is the input and y is the output that fits the problem. For example, in classification
problems, the model outputs a probability distribution over labels.
Then, we collect a set of labeled samples D, and fine-tune the model by
(ω̃, θ̃) = arg min
ω,θ̂ +

X

(x,ygold )∈D

Loss(yω,θ̂+ , ygold )

(1.22)

1.4 Applying BERT Models

31

where (x, ygold ) represents a tuple of an input and its corresponding output. The notation of this
equation seems a bit complicated, but the training/tuning process is standard. We optimize the
model by minimizing the loss over the tuning samples. The outcome is the optimized parameters
ω̃ and θ̃. The optimization starts with the pre-trained parameters θ̂. Here we use θ̂ + to indicate that
the parameters are initialized with θ̂, and use yω,θ̂+ to denote the model output computed using
the parameters ω and θ̂ + .
With the fine-tuned parameters ω̃ and θ̃, we can apply the model Predictω̃ (BERTθ̃ (·)) to new
data of the same tasks for which the model was fine-tuned. The form of the downstream tasks
determines the input and output formats of the model, as well as the architecture of the prediction
network. In the following we list some tasks to which BERT models are generally suited.

• Classification (Single Text). One of the most widely-used applications of BERT models is
text classification. In this task, a BERT model receives a sequence of tokens and encodes
it as a sequence of vectors. The first output vector hcls (or h0 ) is typically used as the
representation of the entire text. The prediction network takes hcls as input to produce a
distribution of labels. Let [CLS]x1 x2 ...xm be an input text. See below for an illustration of
BERT-based text classification.

Class

hcls

h1

h2

...

hm

hm+1

BERT
ecls

e1

e2

...

em

em+1

[CLS]

x1

x2

...

xm

[SEP]

Here the gray box denotes the prediction network. Many NLP problems can be categorized
as text classification tasks, and there have been several text classification benchmarks for
evaluating pre-trained models. For example, we can classify texts by their grammatical correctness (grammaticality) or emotional tone (sentiment) [Socher et al., 2013; Warstadt et al.,
2019]. Note that the prediction network could be any classification model, such as a deep
neural network or a more traditional classification model. The entire model can then be
trained or fine-tuned in the manner of a standard classification model. For example, the prediction network can be simply a Softmax layer and the model parameters can be optimized
by maximizing the probabilities of the correct labels.

• Classification (Pair of Texts). Classification can also be performed on a pair of texts. Suppose we have two texts, x1 ...xm and y1 ...yn . We can concatenate these texts to form a single
sequence with a length len. Then, we predict a label for this combined text sequence based
on the hcls vector, as follows

Pre-training

32
Class

hcls

h1

h2

...

hm

hm+1 hm+2 hm+3

...

hlen−1 hlen

elen−1 elen

BERT
ecls

e1

e2

...

em

em+1 em+2 em+3

...

[CLS]

x1

x2

...

xm

[SEP]

y1

...

Text 1

y2

yn

[SEP]

Text 2

where len = n + m + 2. Text pair classification covers several problems, including semantic equivalence judgement (determine whether two texts are semantically equivalent)
[Dolan and Brockett, 2005], text entailment judgement (determine whether a hypothesis
can be logically inferred or entailed from a premise) [Bentivogli and Giampiccolo, 2011;
Williams et al., 2018], grounded commonsense inference (determine whether an event is
likely to happen given its context) [Zellers et al., 2018], and question-answering inference
(determine whether an answer corresponds to a given question).
• Regression. Instead of generating a label distribution, we can have the prediction network
output a real-valued score. For example, by adding a Sigmoid layer to the prediction network, the system can be employed to compute the similarity between two given sentences.
The architecture is the same as that of BERT-based classification systems, with only the
change of the output layer.
Number (similarity, evaluation score, etc.)

hcls

h1

h2

...

hm

hm+1 hm+2 hm+3

...

hlen−1 hlen

elen−1 elen

BERT
ecls

e1

e2

...

em

em+1 em+2 em+3

...

[CLS]

x1

x2

...

xm

[SEP]

y1

...

Text 1

y2

yn

[SEP]

Text 2

For training or fine-tuning, we can minimize the regression loss of the model output as
usual.
• Sequence Labeling. Sequence labeling is a machine learning approach applicable to a wide
range of NLP problems. This approach assigns a label to each token in an input sequence,
and some linguistic annotations can then be derived from this sequence of labels. An example of sequence labeling in NLP is part-of-speech (POS) tagging. We label each word
in a sentence with its corresponding POS tag. Another example is named entity recognition
(NER) in which we label each word with an NER tag, and named entities are identified
using these tags. See below for an illustration of the model architecture for NER.

1.4 Applying BERT Models

33
Tag

Tag

Tag

{B, I, O}{B, I, O}

hcls

h1

h2

{B, I, O}

...

hm

hm+1

BERT
ecls

e1

e2

...

em

em+1

[CLS]

x1

x2

...

xm

[SEP]

Here {B, I, O} is the tag set of NER. For example, B-ORG means the beginning of an
organization, I-ORG means the word is inside an organization, and O means the word does
not belong to any named entity. This NER model can output a distribution over the tag set
at each position, denoted as pi . The training or fine-tuning of the model can be performed
over these distributions {p1 , ..., pm }. For example, suppose pi (tagi ) is the probability of
the correct tag at position i. The training loss can be defined to be the negative likelihood
Loss = −

m
1 X
log pi (tagi )
m i=1

(1.23)

Finding the best label sequence given a trained NER model is a well-studied issue in NLP.
This is often achieved via dynamic programming, which, in the context of path finding over
a lattice, has linear complexity [Huang, 2009].
• Span Prediction. Some NLP tasks require predicting a span in a text. A common example
is reading comprehension. In this task, we are given a query x1 ...xm and a context text
y1 ...yn . The goal is to identify a continuous span in y1 ...yn that best answers the query.
This problem can be framed as a sequence labeling-like task in which we predict a label for
each yj to indicate the beginning or ending of the span. Following Seo et al. [2017], we add
two networks on top of the BERT output for yj : one for generating the probability of yj
being the beginning of the span (denoted by pbeg
j ), and one for generating the probability
of yj being the ending of the span (denoted by pend
j ). The resulting model architecture is
shown as follows
End

End

beg

hcls

h1

h2

...

hm

(pend
n )

Beg

Beg
(p1

End

(pend
)
2

(pend
)
1

)

beg

(p2

Beg
beg

)

hm+1 hm+2 hm+3

(pn

)

...

hlen−1 hlen

elen−1 elen

BERT
ecls

e1

e2

...

em

em+1 em+2 em+3

...

[CLS]

x1

x2

...

xm

[SEP]

y1

...

Query

y2

Context Text

yn

[SEP]

Pre-training

34

We pack the query and context text together to obtain the input sequence. The prediction
networks are only applied to outputs for the context text, generating the probabilities pbeg
j
at each position. The loss can be computed by summing the log likelihoods of the
and pend
j
two models across the entire context text.
Loss = −

n


1X
log pbeg
+ log pend
j
j
n j=1

(1.24)

At test time, we search for the best span by
(ĵ1 , ĵ2 ) =

arg max
1≤j1 ≤j2 ≤n

end
log pbeg
j1 + log pj2




(1.25)

• Encoding for Encoder-decoder Models. While our focus in this section has been primarily
on language understanding problems, it is worth noting that BERT models can be applied
to a broader range of NLP tasks. In fact, BERT models can be used in all the scenarios
where we need to encode a piece of text. One application that we have not mentioned is
text generation which includes a range of tasks such as machine translation, summarization,
question answering, and dialogue generation. These tasks can be formulated as sequenceto-sequence problems: we use an encoder to represent the source text, and a decoder to
generate the corresponding target text. A straightforward method to apply BERT models
is to consider them as encoders. Before fine-tuning, we can initialize the parameters of the
encoder with those from a pre-trained BERT model. Then, the encoder-decoder model can
be fine-tuned on pairs of texts as usual. The following shows the architecture of a neural
machine translation system where a BERT model is applied on the source side.
Target Text

Adapter

y1

y2

BERT (Encoder)

y3

...

yn

Decoder

excls

ex1

...

exm

exm+1

ey0

ey1

ey2

...

eyn−1

[CLS]

x1

...

xm

[SEP]

hsi

y1

y2

...

yn−1

Source Text

Here x1 ...xm denotes the source sequence, y1 ...yn denotes the target sequence, ex1 ...exm
denotes the embedding sequence of x1 ...xm , and ey1 ...eyn denotes the embedding sequence
of y1 ...yn . The adapter, which is optional, maps the output of the BERT model to the form
that is better suited to the decoder.
Fine-tuning BERT models is a complicated engineering problem, influenced by many factors,
such as the amount of fine-tuning data, the model size, and the optimizer used in fine-tuning.
In general, we wish to fine-tune these models sufficiently so that they can perform well in the
downstream tasks. However, fine-tuning BERT models for specific tasks may lead to overfitting,

1.5 Summary

35

which in turn reduces their ability to generalize to other tasks. For example, suppose we have a
BERT model that performs well on a particular task. If we then fine-tune it for new tasks, this
may decrease its performance on the original task. This problem is related to the catastrophic
forgetting problem in continual training, where a neural network forgets previously learned information when updated on new samples. In practical applications, a common way to alleviate
catastrophic forgetting is to add some old data into fine-tuning and train the model with more
diverse data. Also, one may use methods specialized to catastrophic forgetting, such as experience replay [Rolnick et al., 2019] and elastic weight consolidation [Kirkpatrick et al., 2017]. The
interested reader can refer to some surveys for more detailed discussions of this issue in continual
learning [Parisi et al., 2019; Wang et al., 2023a;e].

1.5 Summary
In this chapter we have discussed the general idea of pre-training in NLP. In particular, we have discussed self-supervised pre-training and its application to encode-only, decoder-only, and encoderdecoder architectures. Moreover, we have presented and compared a variety of pre-training tasks
for these architectures. As an example, BERT is used to illustrate how sequence models are pretrained via masked language modeling and applied to different downstream tasks.
Recent years have shown remarkable progress in NLP, led by the large-scale use of selfsupervised pre-training. And sweeping advances are being made across many tasks, not only
in NLP but also in computer vision and other areas of AI. One idea behind these advances is that a
significant amount of knowledge about the world can be learned by simply training these AI systems on huge amounts of unlabeled data. For example, a language model can learn some general
knowledge of a language by repeatedly predicting masked words in large-scale text. As a result,
this pre-trained language model can serve as a foundation model, which can be easily adapted to
address specific downstream NLP tasks. This paradigm shift in NLP has enabled the development
of incredibly powerful systems for language understanding, generation, and reasoning [Manning,
2022]. However, it is important to recognize that we are still in the early stages of creating truly intelligent systems, and there is a long way to go. Nevertheless, large-scale pre-training has opened
a door to intelligent systems that researchers have long aspired to develop, though several key research areas remain open for exploration, such as learning intelligence efficiently using reasonably
small-sized data and acquiring complex reasoning and planning abilities.
Note that this chapter is mostly introductory and cannot cover all aspects of pre-training. For
example, there are many methods to fine-tune a pre-trained model, offering different ways to better
adapt the model to diverse situations. Moreover, large language models, which are considered one
of the most significant achievements in AI in recent years, are skipped in this section. We leave
the discussion of these topics to the following chapters.

C HAPTER 2

Generative Models
One of the most significant advances in NLP in recent years might be the development of large
language models (LLMs). This has helped create systems that can understand and generate natural languages like humans. These systems have even been found to be able to reason, which
is considered a very challenging AI problem. With these achievements, NLP made big strides
and entered a new era of research in which difficult problems are being solved, such as building
conversational systems that can communicate with humans smoothly.
The concept of language modeling or probabilistic language modeling dates back to early experiments conducted by Shannon [1951]. In his work, a language model was designed to estimate
the predictability of English — how well can the next letter of a text be predicted when the preceding N letters are known. Although Shannon’s experiments were preliminary, the fundamental
goals and methods of language modeling have remained largely unchanged over the decades since
then. For quite a long period, particularly before 2010, the dominant approach to language modeling was the n-gram approach [Jurafsky and Martin, 2008]. In n-gram language modeling, we
estimate the probability of a word given its preceding n − 1 words, and thus the probability of a
sequence can be approximated by the product of a series of n-gram probabilities. These probabilities are typically estimated by collecting smoothed relative counts of n-grams in text. While
such an approach is straightforward and simple, it has been extensively used in NLP. For example,
the success of modern statistical speech recognition and machine translation systems has largely
depended on the utilization of n-gram language models [Jelinek, 1998; Koehn, 2010].
Applying neural networks to language modeling has long been attractive, but a real breakthrough appeared as deep learning techniques advanced. A widely cited study is Bengio et al.
[2003]’s work where n-gram probabilities are modeled via a feed-forward network and learned
by training the network in an end-to-end fashion. A by-product of this neural language model
is the distributed representations of words, known as word embeddings. Rather than representing words as discrete variables, word embeddings map words into low-dimensional real-valued
vectors, making it possible to compute the meanings of words and word n-grams in a continuous representation space. As a result, language models are no longer burdened with the curse of
dimensionality, but can represent exponentially many n-grams via a compact and dense neural
model.
The idea of learning word representations through neural language models inspired subsequent
research in representation learning in NLP. However, this approach did not attract significant interest in developing NLP systems in the first few years after its proposal. Starting in about 2012,
though, advances were made in learning word embeddings from large-scale text via simple word
prediction tasks. Several methods, such as Word2Vec, were proposed to effectively learn such
embeddings, which were then successfully applied in a variety of NLP systems [Mikolov et al.,
2013a;b]. As a result of these advances, researchers began to think of learning representations of
sequences using more powerful language models, such as LSTM-based models [Sutskever et al.,
2014; Peters et al., 2018]. And further progress and interest in sequence representation exploded
after Transformer was proposed. Alongside the rise of Transformer, the concept of language modeling was generalized to encompass models that learn to predict words in various ways. Many

36

2.1 A Brief Introduction to LLMs

37

powerful Transformer-based models were pre-trained using these word prediction tasks, and successfully applied to a variety of downstream tasks [Devlin et al., 2019].
Indeed, training language models on large-scale data has led NLP research to exciting times.
While language modeling has long been seen as a foundational technique with no direct link to
the goals of artificial intelligence that researchers had hoped for, it helps us see the emergence of
intelligent systems that can learn a certain degree of general knowledge from repeatedly predicting
words in text. Recent research demonstrates that a single, well-trained LLM can handle a large
number of tasks and generalize to perform new tasks with a small adaptation effort [Bubeck et al.,
2023]. This suggests a step towards more advanced forms of artificial intelligence, and inspires
further exploration into developing more powerful language models as foundation models.
In this chapter, we consider the basic concepts of generative LLMs. For simplicity, we use the
terms large language models or LLMs to refer to generative models like GPT, though this term
can broadly cover other types of models like BERT. We begin by giving a general introduction
to LLMs, including the key steps of building such models. We then discuss two scaling issues of
LLMs: how LLMs are trained at scale, and how LLMs can be improved to handle very long texts.
Finally, we give a summary of these discussions.

2.1 A Brief Introduction to LLMs
In this section we give an introduction to the basic ideas of LLMs as required for the rest of this
chapter and the following chapters. We will use terms word and token interchangeably. Both
of them refer to the basic units used in language modeling, though their original meanings are
different.
Before presenting details, let us first consider how language models work. The goal of language modeling is to predict the probability of a sequence of tokens occurring. Let {x0 , x1 , ..., xm }
be a sequence of tokens, where x0 is the start symbol hsi (or hSOSi)1 . The probability of this sequence can be defined using the chain rule
Pr(x0 , ..., xm ) = Pr(x0 ) · Pr(x1 |x0 ) · Pr(x2 |x0 , x1 ) · · · Pr(xm |x0 , ..., xm−1 )
=

m
Y

i=0

(2.1)

Pr(xi |x0 , ..., xi−1 )

or alternatively in a logarithmic form

log Pr(x0 , ..., xm ) =

m
X
i=0

log Pr(xi |x0 , ..., xi−1 )

(2.2)

Here Pr(xi |x0 , ..., xi−1 ) is the probability of the token xi given all its previous tokens {x0 , ..., xi−1 }
2 . In the era of deep learning, a typical approach to language modeling is to estimate this
1

The start symbol can also be [CLS] following BERT models.
We assume that when i = 0, Pr(xi |x0 , ..., xi−1 ) =
Pr(x0 ) Pr(x1 , ..., xm |x0 ) = Pr(x1 , ..., xm |x0 ).
2

Pr(x0 )

=

1.

Hence Pr(x0 , ..., xm )

=

Generative Models

38
Context

Predict

Decision Rule

Sequence Probability

hsi a

b

hsi a b

c

arg maxx2 ∈V Pr(x2 |hsi a)

Pr(hsi) · Pr(a|hsi)· Pr(b|hsi a)

hsi a b c

d

arg maxx4 ∈V Pr(x4 |hsi a b c)

Pr(hsi) · Pr(a|hsi) · Pr(b|hsi a)·
Pr(c|hsi a b)· Pr(d|hsi a b c)

arg maxx3 ∈V Pr(x3 |hsi a b)

Pr(hsi) · Pr(a|hsi) · Pr(b|hsi a)·
Pr(c|hsi a b)

Table 2.1: Illustration of generating the three tokens b c d given the prefix hsi a via a language model. In each step,
the model picks a token xi from V so that Pr(xi |x0 , ..., xi−1 ) is maximized. This token is then appended to the end
of the context sequence. In the next step, we repeat the same process, but based on the new context.

probability using a deep neural network. Neural networks trained to accomplish this task receive a sequence of tokens x0 , ..., xi−1 and produce a distribution over the vocabulary V (denoted by Pr(·|x0 , ..., xi−1 )). The probability Pr(xi |x0 , ..., xi−1 ) is the value of the i-th entry of
Pr(·|x0 , ..., xi−1 ).
When applying a trained language model, a common task is to find the most likely token given
its previous context tokens. This token prediction task can be described as
x̂i = arg max Pr(xi |x0 , ..., xi−1 )

(2.3)

xi ∈V

We can perform word prediction multiple times to generate a continuous text: each time we
predict the best token x̂i , and then add this predicted token to the context for predicting the next
token x̂i+1 . This results in a left-to-right generation process implementing Eqs. (2.1) and (2.2). To
illustrate, consider the generation of the following three words given the prefix ‘hsi a’, as shown
in Table 2.1. Now we discuss how LLMs are constructed, trained, and applied.

2.1.1 Decoder-only Transformers
As is standard practice, the input of a language model is a sequence of tokens (denoted by
{x0 , ..., xm−1 }). For each step, an output token is generated, shifting the sequence one position forward for the next prediction. To do this, the language model outputs a distribution
Pr(·|x0 , ..., xi−1 ) at each position i, and the token xi is selected according to this distribution.
P
3
This model is trained by maximizing the log likelihood m
i=1 log Pr(xi |x0 , ..., xi−1 ) .

Here, we focus on the decoder-only Transformer architecture, as it is one of the most popular
model architectures used in LLMs. The input sequence of tokens is represented by a sequence
of de -dimensional vectors {e0 , ..., em−1 }. ei is the sum of the token embedding of xi and the
positional embedding of i. The major body of the model is a stack of Transformer blocks (or
layers). Each Transformer block has two stacked sub-layers, one for self-attention modeling and
one for FFN modeling. These sub-layers can be defined using the post-norm architecture
output = LNorm(F (input) + input)
3

Note that

Pm

i=1

log Pr(xi |x0 , ..., xi−1 ) =

Pm

i=0

log Pr(xi |x0 , ..., xi−1 ) since log Pr(x0 ) = 0.

(2.4)

2.1 A Brief Introduction to LLMs

39

or the pre-norm architecture
output = LNorm(F (input)) + input

(2.5)

where input and output denote the input and output, both being an m × d matrix. The i-th rows
of input and output can be seen as contextual representations of the i-th token in the sequence.
F (·) is the core function of a sub-layer. For FFN sub-layers, F (·) is a multi-layer FFN. For
self-attention sub-layers, F (·) is a multi-head self-attention function. In general, self-attention is
expressed in a form of QKV attention
QKT
Attqkv (Q, K, V) = Softmax( √ + Mask)V
d

(2.6)

where Q, K and V ∈ Rm×d are the queries, keys, and values, respectively. It is important to
note that only previous tokens are considered when predicting a token. So a masking variable
Mask ∈ Rm×m is incorporated into self-attention to achieve this. The entry (i, k) of Mask has
a value of 0 if i ≤ k, and a value of − inf otherwise.
Given a representation H ∈ Rm×d , the multi-head self-attention function can be defined as
F (H) = Merge(head1 , ..., headτ )Whead

(2.7)

where Merge(·) representees a concatenation of its inputs, and Whead ∈ Rd×d represents a parameter matrix. headj is the output of QKV attention on a sub-space of representation
headj = Attqkv (Q[j] , K[j] , V[j] )

(2.8)

Q[j],K[j] ,and V[j] are the queries, keys, and values projected onto the j-th sub-space via linear
transformations
Q[j] = HWqj
K

[j]

=

[j]

=

V

HWkj
HWvj

(2.9)
(2.10)
(2.11)

d

where Wqj , Wkj , and Wvj ∈ Rd× τ are the parameter matrices of the transformations.

Suppose we have L Transformer blocks. A Softmax layer is built on top of the output of the
last block. The Softmax layer outputs a sequence of m distributions over the vocabulary, like this




Pr(·|x0 , ..., xm−1 )


..


.




Pr(·|x0 , x1 )
Pr(·|x0 )

 =



Softmax(HL Wo )

(2.12)

where HL is the output of the last Transformer block, and Wo ∈ Rd×|V | is the parameter matrix.

Figure 2.1 shows the Transformer architecture for language modeling. Applying this language

Generative Models

40

Post-norm or Pre-norm
FFN

x2

...

xm

L Blocks

x1

Pr(xm |x0 x1 ...xm−1 )
Pr(x2 |x0 x1 )
Pr(x1 |x0 )

Post-norm or Pre-norm

...
hL
0

hL
1

...

Self-attention

hL
m−1

Language Model
e0

e1

...

em−1

x0

x1

...

xm−1

z0

z1

...

zm−1

Fig. 2.1: The Transformer-decoder architecture for language modeling. The central components are L stacked Transformer blocks, each comprising a self-attention sub-layer and an FFN sub-layer. To prevent the model from accessing
the right-context, a masking variable is incorporated into self-attention. The output layer uses a Softmax function to
generate a probability distribution for the next token, given the sequence of previous tokens. During inference, the
model takes the previously predicted token to predict the next one, repeating this process until the end of the sequence
L
is reached. {z0 , ..., zm−1 } denote the inputs of a Transformer block, and {hL
0 , ..., hm−1 } denote the outputs of the
last Transformer block.

model follows an autoregressive process. Each time the language model takes a token xi−1 as
input and predicts a token xi that maximizes the probability Pr(xi |x0 , ..., xi−1 ). It is important
to note that, despite different implementation details, many LLMs share the same architecture
described above. These models are called large because both their depth and width are significant.
Table 2.2 shows the model sizes for a few LLMs, as well as their model setups.

2.1.2 Training LLMs
Now suppose that we are given a training set D comprising K sequences. The log-likelihood of
each sequence x = x0 ...xm in D can be calculated using a language model
Lθ (x) =

m
X
i=1

log Prθ (xi |x0 , ..., xi−1 )

(2.13)

Here the subscript θ affixed to L(·) and Pr(·) denotes the parameters of the language model. Then,
the objective of maximum likelihood training is defined as
θ̂ = arg max
θ

X

x∈D

Lθ (x)

(2.14)

Training Transformer-based language models with the above objective is commonly viewed
as a standard optimization process for neural networks. This can be achieved using gradient descent algorithms, which are widely supported by off-the-shelf deep learning toolkits. Somewhat

2.1 A Brief Introduction to LLMs

41
# of Parameters

Depth L

Width d

# of Heads
(Q/KV)

0.117B
1.5B
175B

12
48
96

768
1,600
12,288

12/12
25/25
96/96

LLaMA2 [Touvron et al., 2023b]

7B
13B
70B

32
40
80

4,096
5,120
8,192

32/32
40/40
64/64

LLaMA3/3.1 [Dubey et al., 2024]

8B
70B
405B

32
80
126

4,096
8,192
16,384

32/8
64/8
128/8

Gemma2 [Team et al., 2024]

2B
9B
37B

26
42
46

2,304
3,584
4,608

8/4
16/8
32/16

Qwen2.5 [Yang et al., 2024]

0.5B
7B
72B

24
28
80

896
3,584
8,192

14/2
28/4
64/8

DeepSeek-V3 [Liu et al., 2024a]

671B

61

7,168

128/128

Falcon [Penedo et al., 2023]

7B
40B
180B

32
60
80

4,544
8,192
14,848

71/71
128/128
232/232

Mistral [Jiang et al., 2023a]

7B

32

4,096

32/32

LLM
GPT-1 [Radford et al., 2018]
GPT-2 [Radford et al., 2019]
GPT-3 [Brown et al., 2020]

Table 2.2: Comparison of some LLMs in terms of model size, model depth, model width, and number of heads (a/b
means a heads for queries and b heads for both keys and values).

surprisingly, better results were continuously yielded as language models were evolved into more
computationally intensive models and trained on larger datasets [Kaplan et al., 2020]. These successes have led NLP researchers to continue increasing both the training data and model size in
order to build more powerful language models.
However, as language models become larger, we confront new training challenges, which
significantly change the problem compared to training relatively small models. One of these
challenges arises from the need for large-scale distributed systems to manage the data, model
parameters, training routines, and so on. Developing and maintaining such systems requires a
significant amount of work in both software and hardware engineering, as well as expertise in deep
learning. A related issue is that when the training is scaled up, we need more computing resources
to ensure the training process can be completed in an acceptable time. For example, it generally
requires hundreds or thousands of GPUs to train an LLM with tens of billions of parameters
from scratch. This requirement drastically increases the cost of training such models, especially
considering that many training runs are needed as these models are developed. Also, from the
perspective of deep learning, the training process can become unstable if the neural networks are
very deep and/or the model size is very large. In response, we typically need to modify the model
architecture to adapt LLMs to large-scale training. In Section 2.2 we will present more discussions
on these issues.

Generative Models

42

2.1.3 Fine-tuning LLMs
Once we have pre-trained an LLM, we can then apply it to perform various NLP tasks. Traditionally language models are used as components of other systems, for example, they are widely
applied to score translations in statistical machine translation systems. By contrast, in generative
AI, LLMs are considered complete systems and are employed to address NLP problems by making use of their generation nature. A common approach is to describe the task we want to address
in text and then prompt LLMs to generate text based on this description. This is a standard text
generation task where we continue or complete the text starting from a given context.
More formally, let x = x0 ...xm denote a token sequence of context given by users, and
y = y1 ...yn denote a token sequence following the context. Then, the inference of LLMs can be
defined as a problem of finding the most likely sequence y based on x:
ŷ = arg max log Pr(y|x)
y

= arg max
y

n
X
i=1

log Pr(yi |x0 , ..., xm , y1 , ..., yi−1 )

(2.15)

P

Here ni=1 log Pr(yi |x0 , ..., xm , y1 , ..., yi−1 ) essentially expresses the same thing as the righthand side of Eq. (2.2). It models the log probability of predicting tokens from position m + 1,
rather than position 0. Throughout this chapter and subsequent ones, we will employ separate
variables x and y to distinguish the input and output of an LLM, though they can be seen as subsequences from the same sequence. By adopting such notation, we see that the form of the above
equation closely resembles those used in other text generation models in NLP, such as neural
machine translation models.
To illustrate how LLMs are applied, consider the problem of determining the grammaticality
for a given sentence. We can define a template like this
{*sentence*}
Question: Is this sentence grammatically correct?
Answer:
Here
represents the text we intend to generate. {*sentence*} is a placeholder variable that
will be replaced by the actual sentence provided by the users. For example, suppose we have a
sentence “John seems happy today.”. We can replace the {*sentence*} in the template with this
sentence to have an input to the language model
John seems happy today.
Question: Is this sentence grammatically correct?
Answer:
To perform the task, the language model is given the context x =“John seems happy today .\n
Question : Is this sentence grammatically correct?\n Answer :”4 . It then generates the following
4

\n is a special character used for line breaks.

2.1 A Brief Introduction to LLMs

43

text as the answer, based on the context. For example, the language model may output “Yes” (i.e.,
y = “Yes”) if this text is the one with the maximum probability of prediction given this context.
Likewise, we can define more templates to address other tasks. For example, we can translate
an English sentence into Chinese using the following template

{*sentence*}
Question: What is the Chinese translation of this English sentence?
Answer:

or using an instruction-like template

{*sentence*}
Translate this sentence from English into Chinese.

or using a code-like template.

[src-lang] = English [tgt-lang] = Chinese [input] = {*sentence*}
[output] =

The above templates provide a simple but effective method to “prompt” a single LLM to perform various tasks without adapting the structure of the model. However, this approach requires
that the LLM can recognize and follow the instructions or questions. One way to do this is to incorporate training samples with instructions and their corresponding responses into the pre-training
dataset. While this method is straightforward, building and training LLMs from scratch is computationally expensive. Moreover, making instruction-following data effective for pre-training
requires a significant amount of such data, but collecting large-scale labeled data for all tasks of
interest is very difficult.
A second method, which has been a de facto standard in recent research, is to adapt LLMs
via fine-tuning. As such, the token prediction ability learned in the pre-training phase can be
generalized to accomplish new tasks. The idea behind fine-tuning is that some general knowledge
of language has been acquired in pre-training, but we need a mechanism to activate this knowledge
for applying it to new tasks. To achieve this, we can slightly fine-tune the model parameters using
instruction-following data. This approach is called instruction fine-tuning.
An instruction fine-tuning sample, which is represented by a sequence of tokens, can be seen
as a tuple consisting of an input and the desired output. Here, the input includes instructions,
system information (or system prefix), and any other user-provided information5 . To illustrate,
consider the following examples (blue text = input and underlined text = output).
5
System information refers to a sequence of tokens added at the beginning of an input in order to guide the behavior
of an LLM, such as, you are a helpful assistant and should not output toxic content.

Generative Models

44

Read the following sentence, and answer whether it is grammatically correct.
LLMs are powerful models but are expensive to build.
Yes

LLMs are powerful models but are expensive to build.
Does this sentence make sense grammatically?
Answer Option
- Yes
- No
Yes

Yes or No? Yes = the following sentence is grammatically correct. No = it contains grammatical errors.
LLMs are powerful models but are expensive to build.
Yes

All these samples describe the same binary classification task, but with different instructions.
To increase the diversity in the fine-tuning data and make LLMs generalize, we can define more
tasks using instructions.

Write a short story about three characters.
Character 1: A boy whose name is Andy.
Character 2: Andy’s dog.
Character 3: Andy’s friend, Jane.
In the heart of a bustling city, Andy, his loyal dog Rocket, and his friend
Jane discovered a hidden garden beneath the glow of twilight. Surrounded by
ancient trees and whispering winds, they shared dreams and secrets, forging an
unbreakable bond. As the night deepened, they made a pact under the stars to
always find magic in the mundane. Together, they realized that true adventure lay
in the moments spent with friends, and every day was a treasure waiting to be
discovered.

2.1 A Brief Introduction to LLMs

45

Show me a recipe for making ice cream.
Ingredients:
2 cups heavy cream; 1 cup milk; 3/4 cup sugar; 2 tsp vanilla extract;
Pinch of salt.
Steps:
1. Whisk all ingredients until sugar dissolves.
2. Chill mixture for 2 hours in the fridge.
3. Freeze in a shallow dish, stirring every 30 minutes, 3-4 times.
4. Freeze solid for 2 hours.
5. Serve your homemade ice cream.

If you buy 5 apples and each apple costs $1.20, how much do you spend in total?
$6.00

Write a Python program to calculate the sum of squares of the following numbers.
1 , 2 , 10 , -9 , 78
numbers = [1,2,10,-9 ,78]
sum_of_squares = sum(x**2 for x in numbers)
print(sum_of_squares)
To acquire instruction-following abilities, a certain amount of fine-tuning data is required. This
data may include diverse instructions and possible responses. It has been found that scaling the
number of fine-tuning tasks is beneficial for improving the performance of LLMs [Chung et al.,
2022]. Note that although more fine-tuning data is favorable, the amount of this data is generally
orders of magnitude smaller than that of the pre-training data. For example, LLMs can be finetuned with tens or hundreds of thousands of samples, or even fewer if these samples are of high
quality [Zhou et al., 2023a; Chen et al., 2023b], whereas pre-training such models may require
billions or trillions of tokens, resulting in significantly larger computational demands and longer
training times [Touvron et al., 2023a].
It is also worth noting that we should not expect the fine-tuning data to cover all the downstream tasks to which we intend to apply LLMs. A common understanding of how the pre-training
+ fine-tuning approach works is that LLMs have gained knowledge for understanding instructions
and generating responses in the pre-training phase. However, these abilities are not fully activated
until we introduce some form of supervision. The general instruction-following behavior emerges
as we fine-tune the models with a relatively small amount of labeled data. As a result, we can
achieve some level of zero-shot learning: the fine-tuned models can handle new tasks that they
have not been explicitly trained or fine-tuned for [Sanh et al., 2022; Wei et al., 2022a]. This zeroshot learning ability distinguishes generative LLMs from earlier pre-trained models like BERT,
which are primarily fine-tuned for specific tasks.
Once we have prepared a collection of instruction-described data, the fine-tuning process is
relatively simple. This process can be viewed as a standard training process as pre-training, but on
a much smaller training dataset. Let Dtune be the fine-tuning dataset and θ̂ be the model parameters

Generative Models

46

optimized via pre-training. We can modify Eq. (2.14) to obtain the objective of fine-tuning
θ̃ = arg max
θ̂ +

X

sample∈Dtune

(2.16)

Lθ̂+ (sample)

Here θ̃ denotes the optimal parameters. The use of notation θ̂ + means that the fine-tuning starts
with the pre-trained parameters θ̂.
For each sample ∈ Dtune , we divide it into an input segment xsample and an output segment
ysample , that is,
(2.17)

sample = [ysample , xsample ]
We then define the loss function to be
Lθ̂+ (sample) = − log Prθ̂+ (ysample |xsample )

(2.18)

In other words, we compute the loss over the sub-sequence ysample , rather than the entire sequence.
In a practical implementation of back-propagation for this equation, the sequence [ysample , xsample ]
is constructed in the forward pass as usual. However, in the backward pass, error gradients are
propagated back only through the parts of the network that correspond to ysample , leaving the rest
of the network unchanged. As an example, consider a sequence
hsi Square this number . 2 . The result is 4 .
|

{z

Context (Input)

} |

{z

}

Prediction (Output)

The loss is calculated and back propagated only for The result is 4 ..

Instruction fine-tuning also requires substantial engineering work. In order to achieve satisfactory results, one may experiment with different settings of the learning rate, batch size, number
of fine-tuning steps, and so on. This typically requires many fine-tuning runs and evaluations. The
cost and experimental effort of fine-tuning remain critical and should not be overlooked, though
they are much lower than those of the pre-training phase.
While we focus on instruction fine-tuning for an illustrative example here, fine-tuning techniques play an important role in developing various LLMs and are more widely used. Examples
include fine-tuning LLMs as chatbots using dialog data, and adapting these models to handle very
long sequences. The wide application of fine-tuning has led researchers to improve these techniques, such as designing more efficient fine-tuning algorithms. While the research on fine-tuning
is fruitful, in this section we just give a flavour of the key steps involved. We will see more detailed
discussions on this topic in the following chapters.

2.1.4 Aligning LLMs with the World
Instruction fine-tuning provides a simple way to adapt LLMs to tasks that can be well defined. This
problem can broadly be categorized as an alignment problem. Here, alignment is referred to as a
process of guiding LLMs to behave in ways that align with human intentions. The guidance can
come from labeled data, human feedback, or any other form of human preferences. For example,

2.1 A Brief Introduction to LLMs

47

we want LLMs not only to be accurate in following instructions, but also to be unbiased, truthful,
and harmless. So we need to supervise the models towards human values and expectations. A
common example is that when we ask an LLM how to build a weapon, it may provide a list of key
steps to do so if it is not carefully aligned. However, a responsible model should recognize and
avoid responding to requests for harmful or illegal information. Alignment in this case is crucial
for ensuring that LLMs act responsibly and in accordance with ethical guidelines.
A related concept to alignment is AI safety. One ultimate goal of AI is to build intelligent
systems that are safe and socially beneficial. To achieve this goal we should keep these systems
robust, secure, and subjective, in any conditions of real-world use, even in conditions of misuse
or adverse use. For LLMs, the safety can be increased by aligning them with appropriate human
guidance, such as human labeled data and interactions with users during application.
Alignment is difficult as human values and expectations are diverse and shifting. Sometimes,
it is hard to describe precisely what humans want, unless we see the response of LLMs to user
requests. This makes alignment no longer a problem of tuning LLMs on predefined tasks, but a
bigger problem of training them with the interactions with the real world.
As a result of the concerns with controlling AI systems, there has been a surge in research
on the alignment issue for LLMs. Typically, two alignment steps are adopted after LLMs are
pre-trained on large-scale unlabeled data.
• Supervised Fine-tuning (SFT). This involves continuing the training of pre-trained LLMs
on new, task-oriented, labelled data. A commonly used SFT technique is instruction finetuning. As described in the previous subsection, by learning from instruction-response
annotated data, LLMs can align with the intended behaviors for following instructions,
thereby becoming capable of performing various instruction-described tasks. Supervised
fine-tuning can be seen as following the pre-training + fine-tuning paradigm, and offers a
relatively straightforward method to adapt LLMs.
• Learning from Human Feedback. After an LLM finishes pre-training and supervised finetuning, it can be used to respond to user requests if appropriately prompted. But this model
may generate content that is unfactual, biased, or harmful. To make the LLM more aligned
with the users, one simple approach is to directly learn from human feedback. For example,
given some instructions and inputs provided by the users, experts are asked to evaluate how
well the model responds in accordance with their preferences and interests. This feedback
is then used to further train the LLM for better alignment.
A typical method for learning from human feedback is to consider it as a reinforcement learning (RL) problem, known as reinforcement learning from human feedback (RLHF) [Ouyang et al.,
2022]. The RLHF method was initially proposed to address general sequential decision-making
problems [Christiano et al., 2017], and was later successfully employed in the development of
the GPT series models [Stiennon et al., 2020]. As a reinforcement learning approach, the goal of
RLHF is to learn a policy by maximizing some reward from the environment. Specifically, two
components are built in RLHF:
• Agent. An agent, also called an LM agent, is the LLM that we want to train. This agent
operates by interacting with its environment: it receives a text from the environment and

Generative Models

48

outputs another text that is sent back to the environment. The policy of the agent is the
function defined by the LLM, that is, Pr(y|x).
• Reward Model. A reward model is a proxy of the environment. Each time the agent
produces an output sequence, the reward model assigns this output sequence a numerical
score (i.e., the reward). This score tells the agent how good the output sequence is.

In RLHF, we need to perform two learning tasks: 1) reward model learning, which involves
training a reward model using human feedback on the output of the agent, and 2) policy learning,
which involves optimizing a policy guided by the reward model using reinforcement learning
algorithms. Here is a brief outline of the key steps involved in RLHF.
• Build an initial policy using pre-training and instruction fine-tuning.
• Use the policy to generate multiple outputs for each input, and then collect human feedback
on these outputs (e.g., comparisons of the outputs).
• Learn a reward model from the human feedback.
• Fine-tune the policy with the supervision from the reward model.
Figure 2.2 shows an overview of RLHF. Given that this section serves only as a brief introduction to concepts of LLMs, a detailed discussion of RLHF techniques will not be included. We
instead illustrate the basic ideas behind RLHF using a simple example.
Suppose we have trained an LLM via pre-training and instruction fine-tuning. This LLM is
deployed to respond to requests from users. For example, a user may input
How can I live a more environmentally friendly life?
We use the LLM to generate 4 different outputs (denoted by {y1 , ..., y4 }) by sampling the
output space
Output 1 (y1 ): Consider switching to an electric vehicle or bicycle instead of
traditional cars to reduce carbon emissions and protect our planet.
Output 2 (y2 ): Adopt a minimalist lifestyle. Own fewer possessions to reduce
consumption and the environmental impact of manufacturing and
disposal.
Output 3 (y3 ): Go off-grid. Generate your own renewable energy and collect
rainwater to become completely self-sufficient and reduce reliance
on non-renewable resources.
Output 4 (y4 ): Support local farm products to reduce the carbon footprint of
transporting food, while enjoying fresh, healthy food.

2.1 A Brief Introduction to LLMs

49
Comparisons
y1 ≻ y4 ≻ y2 ≻ y3

SFT Data
Write a poem about the
weather in London . ...

Annotating Data with Human Preferences

Pre-training Data

Model Output

How can I get there? ...
I love the food here! ...

1. ............
3. ............

Pre-training &
Supervised fine-tuning

2. ............
4. ............

Predicting
User Input

LLM

LLM

How can I live more
environmentally friendly?

(b) Annotating Data with Human Preferences

(a) Learning an Initial LLM

Reward Scores
{r(x, y)}
Evaluate the Input-output Pairs

Comparison Data
{(x, yk1 ≻ yk2 )}

RL Fine-tuning

Reward Model

Input-output Pairs
{x, y}

Training

Reward Model
(c) Training the Reward Model

Sampling y via the Policy Pr(y|x)

LLM

(Policy)

Dataset D
x∼D

(d) Training/Fine-tuning the Policy

Fig. 2.2: An overview of RLHF. There are 4 key steps involved: a) training an initial LLM (i.e., policy) using pretraining and supervised fine-tuning; b) collecting human preference data by ranking the outputs of the LLM; c) training
a reward model using the ranking results; d) RL fine-tuning of the policy based on the reward model. Double line
arrows mean training or fine-tuning.

We then ask annotators to evaluate these outputs. One straightforward way is to assign a rating
score to each output. In this case, the reward model learning problem can be framed as a task of
training a regression model. But giving numerical scores to LLM outputs is not an easy task for
annotators. It is usually difficult to design an annotation standard that all annotators can agree on
and easily follow. An alternative method, which is more popular in the development of LLMs, is
to rank these outputs. For example, a possible ranking of the above outputs is
y1 ≻ y4 ≻ y2 ≻ y3

Generative Models

50

A reward model is then trained using this ranking result. In general, a reward model in RLHF
is a language model that shares the same architecture as the target LLM, but with a smaller model
size. Given the input x and output yk , we concatenate them to form a sequence seq k = [x, yk ].
This sequence is processed from left to right using forced decoding. Since each position can
only access its left context in language modeling, the output of the top-most Transformer layer at
the first position cannot be used as the representation of the sequence. Instead, a special symbol
(e.g., h\si) is added to the end of the sequence, and the corresponding output of the Transformer
layer stack is considered as the representation of the entire sequence. An output layer, such as a
linear transformation layer, is built on top of this representation to generate the reward, denoted
by R(seq k ) or R(x, yk ).
We train this reward model using ranking loss. For example, a pair-wise ranking loss function
can be written in the form
Lossω (Dr ) = −E(x,yk1 ,yk2 )∼Dr log(Sigmoid(Rω (x, yk1 ) − Rω (x, yk2 )))

(2.19)

where ω represents the parameters of the reward model, and Dr represents a set of tuples of an
input and a pair of outputs. (x, yk1 , yk2 ) ∼ Dr is a sampling operation which draws a sample
(x, yk1 , yk2 ) from Dr with some probability. As an example, suppose we first draw a model
input x with a uniform distribution and then draw a pair of model outputs with a probability of
yk1 ≻ yk2 given x (denoted by Pr(yk1 ≻ yk2 |x)). The corresponding loss function is given by
Lossω (Dr )
= −
= −

X

Pr(x) · Pr(yk1 ≻ yk2 |x) · log(Sigmoid(Rω (x, yk1 ) − Rω (x, yk2 )))

1 X

K

Pr(yk1 ≻ yk2 |x) · log(Sigmoid(Rω (x, yk1 ) − Rω (x, yk2 )))

(2.20)

where K represents the number of model inputs involved in sampling. While the form of these
functions may seem complex, their idea is simple: we penalize the model if the predicted ranking
of two outputs differs from the human-labeled ranking. By contrast, the model receives a bonus,
if the predicted ranking matches the human-labeled ranking.
We can train the reward model by minimizing the above ranking loss
ω̂ = arg min Lossω (Dr )
ω

(2.21)

The resulting model Rω̂ (·) can be employed to evaluate any given pair of input and output. Note
that although the reward model is trained using a ranking-based objective, it is used for scoring.
This allows it to provide continuous supervision signals, which is very beneficial for training other
models.
We now turn to the policy learning problem. A commonly adopted objective is to maximize
the reward on a set of input-output pairs. Following an analogous form of Eq. (2.16), we obtain a
simple training objective for RL fine-tuning
θ̃ = arg max E(x,yθ̂+ )∼Drlft Rω̂ (x, yθ̂+ )

(2.22)

θ̂ +

where the optimal parameters θ̃ are obtained by fine-tuning the pre-trained parameters θ̂. Drlft is

2.1 A Brief Introduction to LLMs

51

the RL fine-tuning dataset. For each sample (x, yθ̂+ ), x is sampled from a prepared dataset of
input sequences, and yθ̂+ is sampled from the distribution Prθ̂+ (y|x) given by the policy.
In practice, more advanced reinforcement learning algorithms, such as proximal policy optimization (PPO), are often used for achieving more stable training, as well as better performance.
We leave the detailed discussion of reinforcement learning algorithms to the following parts of
this book where RLHF is extensively used for alignment.
An interesting question arises here: why not consider learning from human preferences as
a standard supervised learning problem? This question is closely related to our aforementioned
discussion on the difficulty of data annotation. Often, describing human values and goals is challenging, and it is even more difficult for humans to provide outputs that are well aligned. As an
alternative, annotating the preferences of a given list of model outputs offers a simpler task. By
doing so, we can create a model that understands human preferences, which can then be used as
a reward model for training policies. From the perspective of machine learning, RLHF is particularly useful for scenarios where the desired behavior of an agent is difficult to demonstrate
but can be easily recognized by humans. Another advantage of RLHF is its ability to explore the
sample space. By employing sampling techniques, models trained with reinforcement learning
can venture beyond the annotated data set to explore additional samples. This exploratory ability
allows RLHF to discover potentially beneficial policies that are not immediately apparent from
the labeled data alone.

2.1.5 Prompting LLMs
We have so far shown that LLMs can be used to perform various tasks by giving them appropriate
prompts. There are no restrictions on these prompts, which can include any information we wish
to ask or communicate with LLMs, such as natural language instructions and the context of conversations. Since this approach requires no additional training or tuning, adapting LLMs becomes
highly efficient once they are developed. This somewhat influences the paradigms in NLP: we
no longer need to develop specific systems for individual tasks but can instead use a single, welltrained LLM to perform different tasks by prompting it. An appealing aspect of LLM prompting
arises as a result: users can easily have “customized” systems by designing their own prompts
for LLMs. Given the important role played by prompting in LLMs, prompting engineering has
become a very active area of research in NLP.
The term prompt is used in many different ways in the literature. In this chapter, this term
refers to the entire input to LLMs, and so we use the terms prompt and model input interchangeably. Before discussing prompting further, let us first see a few examples where the prompts
are more complex than those presented in the previous subsections. Note that this subsection is
not aimed at writing high-quality prompts but rather at highlighting some interesting issues in
prompting LLMs.
One of the popular ways to use LLMs is to assign them a “role” played in generating responses.
For example, LLMs can act as a psychologist when answering questions.

Generative Models

52

Please explain what delayed gratification is?
Note, you are a researcher with a deep background in psychology. Your own
field of study is developmental psychology in children, with a focus on aspects
of control and intelligence. You are very familiar with the concept of delayed
gratification and have applied it to your own research. You have published papers
on delayed gratification and conducted related experiments.
Delayed gratification is the process of resisting an immediate reward in
anticipation of receiving a more valuable reward in the future. It is a significant
concept in developmental psychology, especially in relation to how children learn
self-control and develop cognitive abilities. Research shows that the ability to
delay gratification is associated with positive outcomes in academic performance,
social competence, and emotional well-being. Studies and experiments in this
area highlight the importance of cultivating this skill early in life to support
long-term success and resilience.

Another example is the use of LLMs in detecting and correcting errors such as syntactic or
semantic mistakes in text. For an LLM which is trained on both code and natural language data,
we may use it for code debugging6 .

Fix the bugs in this C language program.
#include <stdio.h>
int main() {
printg("Hello, World!")
return 0;
}
1. The function name ’printg’ is incorrect . It should be ’printf’.
2. There is a missing semicolon at the end of the ’printf’ function call.
Here is the corrected version of the program:
#include <stdio.h>
int main() {
printf("Hello, World!");
return 0;
}

Also, LLMs can be fine-tuned to act as a conversation partner 7 .

6

In this example, the code is not tokenized for easier reading.
To fine-tune an LLM for multi-turn dialogue, one needs to consider conversation history in the context for predicting the response in the current round of conversation. This makes the actual prompt used in response generation
relatively longer than that used in single-turn dialogue.
7

2.1 A Brief Introduction to LLMs

53

Hi! I’m an assistant. How can I help you?
Who won the FIFA World Cup 2022?
Argentina won the FIFA World Cup 2022.
Where was it held?
The 2022 FIFA World Cup was held in Qatar.
How many times has Argentina won the World Cup?
Argentina has won the FIFA World Cup three times.
...
These examples and previous ones have shown that appropriate responses can be generated
via prompts involving clear instructions and questions. However, when problem solving requires
knowledge that is not explicitly specified, LLMs may make mistakes, even though the instructions
are sufficiently clear and precise. A family of challenging tasks for LLMs involves arithmetic
reasoning and commonsense reasoning. For example, we can ask an LLM to solve primary school
math problems presented in natural language.
Jack has 7 apples. He ate 2 of them for dinner, but then his mom gave him 5 more
apples. The next day, Jack gave 3 apples to his friend John. How many apples
does Jack have left in the end?
The answer is 10.
The correct answer should be 7, so the model output is incorrect.
One approach to addressing such issues is to incorporate learning into prompts, called incontext learning or (ICL). The idea of ICL is to demonstrate the ways to solve problems in
prompts, and condition predictions on these demonstrations. Here is an example where a similar
problem and the corresponding answer are presented in the prompt (green = demonstrations).
Tom has 12 marbles. He wins 7 more marbles in a game with his friend but then
loses 5 marbles the next day. His brother gives him another 3 marbles as a gift.
How many marbles does Tom have now?
The answer is 17.
Jack has 7 apples. He ate 2 of them for dinner, but then his mom gave him 5 more
apples. The next day, Jack gave 3 apples to his friend John. How many apples
does Jack have left in the end?
The answer is 12.
But the LLM still made mistakes this time. A reason for this might be that solving math
problems does not only involve problem-answer mappings but also, to a larger extent, the underlying logical inference in multiple steps. A method to improve the inference abilities of LLMs
is chain-of-thought prompting (COT prompting) [Wei et al., 2022c]. In COT prompting, we
decompose complex reasoning problems into multiple problem-solving intermediate steps. These

Generative Models

54

steps are demonstrated in prompts so that LLMs can be prompted to learn to mimic the underlying
thought processes in solving the problems. COT prompting has been found to be very useful for
achieving good results on challenging tasks, such as the GSM8K mathematical reasoning benchmark.
Consider the above demonstration example Tom has 12 marbles. He wins 7 more marbles in
a game with his friend but then loses 5 marbles the next day. His brother gives him another 3
marbles as a gift. How many marbles does Tom have now ?. The corresponding logical reasoning
steps can be structured:
1. Tom starts with 12 marbles.
2. He wins 7 more marbles in a game, so we add those to his initial count: 12 marbles + 7
marbles = 19 marbles.
3. The next day, Tom loses 5 marbles, so we subtract those from his current total: 19 marbles
- 5 marbles = 14 marbles.
4. His brother gives him 3 more marbles, so we add those to his current total: 14 marbles + 3
marbles = 17 marbles.
5. After all these events, Tom has 17 marbles.
We can add these reasoning steps into the prompt, and get the correct answer.
Tom has 12 marbles. He wins 7 more marbles in a game with his friend but then
loses 5 marbles the next day. His brother gives him another 3 marbles as a gift.
How many marbles does Tom have now?
Tom initially has 12 marbles. After winning a game, he gains 7 more, increasing
his total to 19 marbles. He loses 5 marbles the next day, bringing his total down
to 14 marbles. His brother gifts him 3 more marbles, increasing his total to 17
marbles. Therefore, Tom now has 17 marbles. So the answer is 17.
Jack has 7 apples. He ate 2 of them for dinner, but then his mom gave him 5 more
apples. The next day, Jack gave 3 apples to his friend John. How many apples
does Jack have left in the end?
Jack starts with 7 apples. He eats 2 apples for dinner, so we subtract 2 from 7,
leaving him with 5 apples. His mom gives him 5 more apples, adding 5 to his
current total. Now he has 5 + 5 = 10 apples. The next day, Jack gives 3 apples to
his friend John. Subtracting 3 from 10, Jack is left with 7 apples. Therefore, Jack
has 7 apples left. So the answer is 7.
This approach is also called one-shot COT prompting. Likewise, we can add more demonstrations to have a few-shot COT prompting approach.
Furthermore, a special kind of COT, zero-shot COT, introduces instructions that provoke
LLMs to think, without presenting any intermediate reasoning steps in the prompts [Kojima et al.,

2.1 A Brief Introduction to LLMs

55

2022]. An interesting example is that by appending a simple sentence “Let’s think step by step.”
to the end of a prompt, LLMs can generate reasoning steps to reach the final answer.

Jack has 7 apples. He ate 2 of them for dinner, but then his mom gave him 5 more
apples. The next day, Jack gave 3 apples to his friend John. How many apples
does Jack have left in the end?
Let’s think step by step.
1. Initial Quantity: Jack starts with 7 apples.
2. After Dinner: He eats 2 apples, so 7 − 2 = 5 apples remain.
3. His Mom Gives More: His mom gives him 5 more apples, so 5 + 5 = 10
apples are now with him.
4. Giving to John: The next day, Jack gives 3 apples to his friend John,
so 10 − 3 = 7 apples are left.
In the end, Jack has 7 apples left.

Zero-shot, one-shot, and few-shot learning are common concepts in the area of in-context
learning for LLMs and are not restricted to COT prompting. Broadly speaking, any prompting
that involves only simple instructions without any demonstrations can be considered a form of
zero-shot learning. This zero-shot learning ability emerges as LLMs are pre-trained and/or finetuned. Also, one-shot and few-shot learning methods are more often considered when LLMs do
not acquire the corresponding zero-shot learning ability. These methods are therefore important
for in-context learning when addressing new tasks. Examples include those for performing various
NLP tasks by demonstrating task-formatted samples. See the following examples for sentiment
sentence classification and phrase translation via few-shot learning.

Given the following text snippets, classify their sentiment as Positive, Negative,
or Neutral.
Example 1: “I had an amazing day at the park!”
Sentiment: Positive
Example 2: “The service at the restaurant was terrible.”
Sentiment: Negative
Example 3: “I think it’s going to rain today.”
Sentiment: Neutral
Text: “This movie was a fantastic journey through imagination.”
Sentiment: Positive

Generative Models

56

Translate the following Chinese phrases into English.
Example 1: “你好”
Translation: “Hello”
Example 2: “谢谢你”
Translation: “Thank you”
Phrase to translate: “早上好”
Translation: “Good Morning”
Above, we have presented examples to illustrate the fundamental in-context learning capabilities of prompting LLMs. This section, however, does not include more advanced prompting
techniques in order to keep the content concise and compact. More discussions on prompting can
be found in Chapter 3.

2.2 Training at Scale
As a first step in developing LLMs, we need to train these models on large amounts of data.
The training task is itself standard: the objective is to maximize the likelihood, which can be
achieved via gradient descent. However, as we scale up both the model size and the amount
of data, the problem becomes very challenging, for example, large models generally make the
training unstable. In this section, we discuss several issues of large-scale training for LLMs,
including data preparation, model modification, and distributed training. We also discuss the
scaling laws for LLMs, which help us understand their training efficiency and effectiveness.

2.2.1 Data Preparation
The importance of data cannot be overstated in NLP. As larger neural networks are developed,
the demand for data continues to increase. For example, developing LLMs may require trillions
of tokens in pre-training (see Table 2.3), orders of magnitude larger than those used in training
conventional NLP models. In general, we may want to gather as much training data as possible.
However, larger training datasets do not mean better training results, and the development of
LLMs raises new issues in creating or collecting these datasets.
A first issue is the quality of data. High-quality data has long been seen as crucial for training
data-driven NLP systems. Directly using raw text from various sources is in general undesirable.
For example, a significant portion of the data used to train recent LLMs comes from web scraping,
which may contain errors and inappropriate content, such as toxic information and fabricated
facts. Also, the internet is flooded with machine-generated content due to the widespread use of
AI, presenting further challenges for processing and using web-scraped data. Researchers have
found that training LLMs on unfiltered data is harmful [Raffel et al., 2020]. Improving data quality
typically involves incorporating filtering and cleaning steps in the data processing workflow. For
example, Penedo et al. [2023] show that by adopting a number of data processing techniques, 90%
of their web-scraped data can be removed for LLM training. In addition to large-scale web-scraped
data, LLM training data often includes books, papers, user-generated data on social media, and
so on. Most of the latest LLMs are trained on such combined datasets, which are found to be

2.2 Training at Scale

57
# of Tokens

LLM

Data

GPT3-175B [Brown et al., 2020]

0.5T

Webpages, Books, Wikipedia

Falcon-180B [Almazrouei et al., 2023]

3.5T

Webpages, Books, Conversations,
Code, Technical Articles

LLaMA2-65B [Touvron et al., 2023a]

1.0T ∼ 1.4T

PaLM-450B [Chowdhery et al., 2022]

0.78T

Webpages, Books, Conversations,
Code, Wikipedia, News

6T

Webpages, Mathematics, Code

Gemma-7B [Gemma Team, 2024]

Webpages, Code, Wikipedia,
Books, Papers, Q&As

Table 2.3: Amounts of training data used in some LLMs in terms of the number of tokens.

important for the strong performance of the resulting models.
A second issue is the diversity of data. We want the training data to cover as many types of
data as possible, so that the trained models can adapt to different downstream tasks easily. It has
been widely recognized that the quality and diversity of training data both play very important
roles in LLMs. An interesting example is that incorporating programming code into training data
has been found to be beneficial for LLMs. The benefits are demonstrated not only in enhancing the
programming abilities of LLMs, but also in improving reasoning for complex problems, especially
those requiring COT prompting. The concept “diversity” can be extended to include language
diversity as well. For example, many LLMs are trained on multi-lingual data, and therefore we
can handle multiple languages using a single model. While this approach shows strong abilities
in multi-lingual and cross-lingual tasks, its performance on specific languages largely depends on
the volume and quality of the data for those languages. It has been shown in some cases to provide
poor results for low-resource languages.
A third issue is the bias in training data. This is not a problem that is specific to LLMs but
exists in many NLP systems. A common example is gender bias, where LLMs show a preference
for one gender over another. This can partly be attributed to class imbalance in the training data,
for example, the term nurses is more often associated with women. In order to debias the data,
it is common practice to balance the categories of different language phenomena, such as gender,
ethnicity, and dialects. The bias in data is also related to the diversity issue mentioned above.
For example, since many LLMs are trained and aligned with English-centric data, they are biased towards the cultural values and perspectives prevalent among English-speaking populations.
Increasing language diversity in training data can somewhat mitigate the bias.
Another issue with collecting large-scale data is the privacy concern. If LLMs are trained
on data from extensive sources, this potentially leads to risks regarding the exposure of sensitive
information, such as intellectual property and personal data. This is particularly concerning given
the capacity of LLMs to represent patterns from the data they are trained on, which might inadvertently involve memorizing and reproducing specific details. A simple approach to privacy
protection is to remove or anonymize sensitive information. For example, anonymization techniques can be applied to remove personally identifiable information from training data to prevent
LLMs from learning from such data. However, in practice, erasing or redacting all sensitive data
is difficult. Therefore, many LLMs, particularly those launched for public service, typically work
with systems that can detect the potential exposure of sensitive data, or are fine-tuned to reject

Generative Models

58
certain requests that could lead to information leakage.

2.2.2 Model Modifications
Training LLMs is difficult. A commonly encountered problem is that the training process becomes more unstable as LLMs get bigger. For example, one needs to choose a small learning rate
to achieve stable training with gradient descent, but this in turn results in much longer training
times. Sometimes, even when the training configuration is carefully designed, training may diverge at certain points during optimization. The training of LLMs is generally influenced by many
factors, such as parameter initialization, batching, and regularization. Here, we focus on common
modifications and improvements to the standard Transformer architecture, which are considered
important in developing trainable LLMs.

2.2.2.1 Layer Normalization with Residual Connections
Layer normalization is used to stabilize training for deep neural networks. It is a process of
subtracting the mean and dividing by the standard deviation. By normalizing layer output in
this way, we can effectively reduce the covariate shift problem and improve the training stability.
In Transformers, layer normalization is typically used together with residual connections. As
described in Section 2.1.1, a sub-layer can be based on either the post-norm architecture, in which
layer normalization is performed right after a residual block, or the pre-norm architecture, in
which layer normalization is performed inside a residual block. While both of these architectures
are widely used in Transformer-based systems [Wang et al., 2019], the pre-norm architecture has
proven to be especially useful in training deep Transformers. Given this, most LLMs are based on
the pre-norm architecture, expressed as output = LNorm(F (input)) + input.
A widely-used form of the layer normalization function is given by
LNorm(h) = α ·

h−µ
+β
σ+ǫ

(2.23)

where h is a d-dimensional real-valued vector, µ is the mean of all the entries of h, and σ is the
corresponding standard deviation. ǫ is introduced for the sake of numerical stability. α ∈ Rd and
β ∈ Rd are the gain and bias terms.

A variant of layer normalization, called root mean square (RMS) layer normalization, only
re-scales the input vector but does not re-center it [Zhang and Sennrich, 2019]. The RMS layer
normalization function is given by
LNorm(h) = α ·

h
σrms + ǫ

where σrms is the root mean square of h, that is, σrms = ( 1d
function is used in LLMs like the LLaMA series.

+β

Pd

(2.24)

2 1
k=1 hk ) 2 . This layer normalization

2.2 Training at Scale

59

2.2.2.2 Activation Functions in FFNs
In Transformers, FFN sub-layers are designed to introduce non-linearities into representation
learning, and are found to be useful for preventing the representations learned by self-attention
from degeneration8 [Dong et al., 2021]. A standard form of the FFNs used in these sub-layers can
be expressed as
FFN(h) = σ(hWh + bh )Wf + bf

(2.25)

where Wh ∈ Rd×dh , bh ∈ Rdh , Wf ∈ Rdh ×d , and bf ∈ Rd are the parameters, and dh is the
hidden size. σ(·) is the activation function of the hidden layer. A common choice for σ(·) is the
rectified linear unit (ReLU), given by
σrelu (h) = max(0, h)

(2.26)

In practical implementations, increasing dh is helpful and thus it is often set to a larger number
in LLMs. But a very large hidden size poses challenges for both training and deployment. In this
case, the design of the activation function plays a relatively more important role in wide FFNs.
There are several alternatives to the ReLU in LLMs. One of these is the gaussian error linear
unit (GeLU) which can be seen as a smoothed version of the ReLU. Rather than controlling the
output by the sign of the input, the GeLU function weights its input by the percentile Pr(h ≤ h).
Here h is a d-dimensional vector whose entries are drawn from the standard normal distribution
Gaussian(0, 1)9 . Specifically, the GeLU function is defined to be
σgelu (h) = h Pr(h ≤ h)
= hΦ(h)

(2.27)

where Φ(h) is the cumulative distribution function of Gaussian(0, 1), which can be implemented
in convenient ways [Hendrycks and Gimpel, 2016]. The GeLU function has been adopted in
several LLMs, such as BERT, GPT-3, and BLOOM.
Another family of activation functions which is popular in LLMs is gated linear unit (GLU)based functions. The basic form of GLUs is given by
σglu (h) = σ(hW1 + b1 ) ⊙ (W2 + b2 )

(2.28)

where W1 ∈ Rd×d , b1 ∈ Rd , W2 ∈ Rd×d , and b2 ∈ Rd are model parameters. Different choices
of σ(·) result in different versions of GLU functions. For example, if σ(·) is defined to be the
GeLU function, we will have the GeGLU function
σgeglu (h) = σgelu (hW1 + b1 ) ⊙ (W2 + b2 )

(2.29)

This activation function has been successfully applied in LLMs like Gemma.
As another example, consider σ(·) to be the Swish function σswish (h) = h ⊙ Sigmoid(ch)
8

Here degeneration refers to the phenomenon in which the rank of a matrix is reduced after some processing.
Pr(h ≤ h) is an informal notation. It refers to a vector, with each entry representing the percentile for the
corresponding entry of h.
9

Generative Models

60
[Ramachandran et al., 2017]. Then, the SwiGLU function is given by
σswiglu (h) = σswish (hW1 + b1 ) ⊙ (W2 + b2 )

(2.30)

Both the PaLM and LLaMA series are based on the SwiGLU function. For more discussions of
GLUs, the reader can refer to Shazeer [2020]’s work.

2.2.2.3 Removing Bias Terms
Another popular model design is to remove the bias terms in affine transformations used in LLMs.
This treatment can be applied to layer normalization, transformations of the inputs to QKV attention, and FFNs. For example, we can modify Eq. (2.25) to obtain an FFN with no bias terms
FFN(h) = σ(hWh )Wf

(2.31)

Chowdhery et al. [2022] report that removing bias terms helps improve the training stability
of LLMs. This method has been used in several recent LLMs, such as LLaMA and Gemma.

2.2.2.4 Other Issues
Many LLMs also involve modifications to their positional embedding models. For example, one
can replace sinusoidal positional encodings with rotary position embeddings so that the learned
LLMs can handle long sequences better. These models will be discussed in Section 2.3.
Note that while model modifications are common in training LLMs, the stability of training
can be improved in many different ways. For example, increasing the batch size as the training
proceeds has been found to be useful for some LLMs. In general, achieving stable and efficient
large-scale LLM training requires carefully designed setups, including learning schedules, optimizer choices, training parallelism, mixed precision training, and so on. Some of these issues are
highly engineered, and therefore, we typically need a number of training runs to obtain satisfactory
LLMs.

2.2.3 Distributed Training
Training LLMs requires significant amounts of computational resources. A common approach to
improving training efficiency is to use large-scale distributed systems. Fortunately, alongside the
rise of neural networks in AI, deep learning-oriented software and hardware have been developed,
making it easier to implement LLMs and perform computations. For example, one can now easily
fine-tune an LLM using deep learning software frameworks and a machine with multiple GPUs.
However, scaling up the training of LLMs is still challenging, and requires significant efforts in
developing hardware and software systems for stable and efficient distributed training.
An important consideration of distributed training is parallelism. There are several forms
of parallelism: data parallelism, model parallelism, tensor parallelism, and pipeline parallelism.
Despite different ways to distribute computations across devices, these parallelism methods are
based on a similar idea: the training problem can be divided into smaller tasks that can be executed simultaneously. The issue of parallelism in training LLMs has been extensively studied

2.2 Training at Scale

61

[Narayanan et al., 2021; Fedus et al., 2022]. Here we sketch the basic concepts.
• Data Parallelism. This method is one of the most widely used parallelism methods for
training neural networks. To illustrate, consider the simplest case where the standard delta
rule is used in gradient descent
θt+1 = θt − lr ·

∂Lθt (Dmini )
∂θt

(2.32)

where the new parameters θt+1 is obtained by updating the latest parameters θt with a small
∂Lθt (Dmini )
is the gradient of the loss
step lr in the direction of the negative loss gradient.
∂θt
with respect to the parameters θt , and is computed on a minibatch of training sample Dmini .
In data parallelism, we divide Dmini into N smaller batches, denoted by {D 1 , ..., D N }.
Then, we distribute these batches to N workers, each with a corresponding batch. Once
the data is distributed, these workers can work at the same time. The gradient of the entire
minibatch is obtained by aggregating the gradients computed by the workers, like this
∂Lθt (Dmini )
∂θt

=

∂Lθt (D 1 ) ∂Lθt (D 2 )
∂Lθt (D N )
+
+··· +
∂θt
∂θt
∂θt

|

{z

worker 1

}

|

{z

worker 2

|

}

{z

worker N

(2.33)

}

In ideal cases where the workers coordinate well and the communication overhead is small,
data parallelism can achieve nearly an N -fold speed-up for training.
• Model Parallelism. Although data parallelism is simple and effective, it requires each
worker to run the entire LLM and perform the complete forward and backward process.
As LLMs grow larger, it sometimes becomes unfeasible to load and execute an LLM on a
single device. In this case, we can decouple the LLM into smaller components and run these
components on different devices. One simple way to do this is to group consecutive layers
in the layer stack and assign each group to a worker. The workers operate in the order of
the layers in the stack, that is, in the forward pass we process the input from lower-level to
upper-level layers, and in the backward pass we propagate the error gradients from upperlevel to lower-level layers. Consider, for example, a Transformer decoder with L stacked
blocks. To distribute the computation load, each block is assigned to a worker. See the
following illustration for a single run of the forward and backward passes of this model.
BL (↑) BL (↓)

Worker L
...
Worker 2
Worker 1 B1 (↑)

...
B2 (↑)

...
B2 (↓)
B1 (↓)

Here Bl denotes the computation of block l, and the symbols ↑ and ↓ denote the forward and
backward passes, respectively. Note that this parallelism method forces the workers to run
in sequence, so a worker has to wait for the previous worker to finish their job. This results
in the devices being idle for most of the time. In practical systems, model parallelism is
generally used together with other parallelism mechanisms to maximize the use of devices.

Generative Models

62

• Tensor Parallelism. Parallelism can also be performed in a single computation step. A
common example is splitting a large parameter matrix into chunks, multiplying an input
tensor with each of these chunks separately, and then concatenating the results of these
multiplications to form the output. For example, consider the multiplication of the representation h ∈ Rd with the parameter matrix Wh ∈ Rd×dh in an FFN sub-layer (see Eq.
(2.25)). We can slice the matrix Wh ∈ Rd×dh vertically to a sequence of M sub-matrices
h

W1h W2h ... WM
h

Wh =

i

(2.34)

where each sub-matrix Wkh has a shape of d × dMh . The multiplication of h with Wh can be
expressed as
h

hWh = h W1h W2h ... WM
h
=

h

i

hW1h hW2h ... hWM
h

i

(2.35)

We can perform matrix multiplications {hW1h , hW2h , ..., hWM
h } on M devices separately.
As a result, we distribute a large matrix multiplication across multiple devices, each of
which may have relatively small memory. From the perspective of the design of modern
GPUs, tensor parallelism over GPUs provides a two-level, tile-based approach to parallel
computing. First, at a higher level, we decompose a matrix multiplication into sub-matrix
multiplications that can directly fit into the memory of GPUs. Then, at a lower level, we
execute these sub-matrix multiplications on GPUs using tile-based parallel algorithms that
are specifically optimized for GPUs.
• Pipeline Parallelism. Above, in model parallelism, we have described a simple approach
to spreading groups of model components across multiple devices. But this method is inefficient because only one device is activated at a time during processing. Pipeline parallelism addresses this issue by introducing overlaps between computations on different
devices [Harlap et al., 2018; Huang et al., 2019]. To do this, a batch of samples is divided
into a number of micro-batches, and then these micro-batches are processed by each worker
as usual. Once a micro-batch is processed by a worker and passed to the next one, the
following micro-batch immediately occupies the same worker. In other words, we create
a pipeline in which different computation steps can overlap if multiple jobs are given to
the pipeline. The following shows an illustration of pipeline parallelism for processing 3
micro-batches.

Worker L

BL,1

...

...

Worker 2
Worker 1

B1,1

B2,1

B2,2

B1,2

B1,3

B2,3

BL,2

BL,3

BL,1

BL,2

BL,3

...
B2,1

B2,2

B2,3

B1,1

B1,2

B1,3

Here Bl,k represents the processing of the k-th micro-batch by the l-th worker. Ideally we
would like to maximize the number of micro-batches, and thus minimize the idle time of the

2.2 Training at Scale

63

workers. However, in practice, using small micro-batches often reduces GPU utilization and
increases task-switching costs. This may, in turn, decrease the overall system throughput.
The ultimate goal of parallel processing is to achieve linear growth in efficiency, that is, the
number of samples that can be processed per unit of time increases linearly with the number of
devices. However, distributed training is complicated, and influenced by many factors in addition
to the parallelism method we choose. One problem, which is often associated with distributed
systems, is the cost of communication. We can think of a distributed system as a group of networked nodes. Each of these nodes can perform local computation or pass data to other nodes. If
there are a large number of such nodes, it will be expensive to distribute and collect data across
them. Sometimes, the time savings brought about by parallelism are offset by the communication overhead of a large network. Another problem with large-scale distributed systems is that
the synchronization of nodes introduces additional costs. As is often the case, some nodes may
take longer to work, causing others to wait for the slowest ones. While we can use asynchronous
training to handle heterogeneity in computational resources, this may lead to stale gradients and
non-guaranteed convergence. Moreover, as more nodes are added to the network, there is more
chance to have crashed nodes during training. In this case, we need to ensure that the whole
system is fault tolerant. In many practical settings, to increase scalability, one needs to take into
account additional issues, including architecture design, data transfer and computation overlap,
load balancing, memory bandwidth and so on.
Training LLMs is so computationally expensive that, even though distributed training is already in use, researchers and engineers often still employ various model compression and speedup methods to improve training efficiency [Weng, 2021]. One example is mixed precision training,
in which low precision data (such as FP16 and FP8 data) is used for gradient computation on each
individual node, and single or double precision data (such as FP32/FP64 data) is used for updating
the model [Micikevicius et al., 2018]. A key operation in this approach is gradient accumulation
where gradients need to be accumulated and synchronized across nodes. However, due to the
non-associativity of floating-point addition, this can lead to slight numerical differences in accumulated gradients on different nodes, which may affect model convergence and final performance.
This problem is more obvious if there are a large number of nodes involved in distributed training,
especially given that low-precision numerical computations may encounter overflow and underflow issues, as well as inconsistencies across different hardware devices. Therefore, the design of
distributed systems needs to consider these numerical computation issues to ensure satisfactory
results and convergence.

2.2.4 Scaling Laws
The success of LLMs reveals that training larger language models using more resources can lead
to improved model performance. Researchers have explained this as scaling laws of LLMs. More
specifically, scaling laws describe the relationships between the performance of LLMs and the
attributes of LLM training, such as the model size, the amount of computation used for training,
and the amount of training data. For example, Hestness et al. [2017] show that the performance of
deep neural networks is a power-law-like function of the training data size. In the beginning, when
the amount of training data is not large, the performance of the model improves slowly. Afterward,
when more training data is used, the model enters a phase of rapid performance improvement, and
the performance curve resembles a power-law curve. Ultimately, the improvement in performance

Generative Models

Number of Test Errors (Log-scale)

64

Slow Reduction
Phase

Power-law Reduction
Phase

Convergence
Phase
(Irreducible Error)

Training Dataset Size (Log-scale)

Fig. 2.3: A scaling law of test error against a variable of interest (e.g., training dataset size) [Hestness et al., 2017]. The
curve of the scaling law can be divided into three phases. At the beginning, the number of test errors decreases slowly
when more training data is used, but this only lasts for a short period. In the second phase, the number of test errors
decreases drastically, and the curve becomes a power law curve. After that, the error reduction slows down again in the
third phase. Note that there are irreducible errors that cannot be eliminated, regardless of the amount of training data.

becomes slow again, and more data does not lead to significant gains. Figure 2.3 shows an example
of such curves.
In NLP, a traditional view holds that the performance gains will disappear at a certain point
as the training is scaled up. However, recent results show that, if we consider the problem on
a larger scale, scaling up training is still a very effective method for obtaining stronger LLMs.
For example, both closed-source and open-source LLMs can benefit from more data, even though
trillions of tokens have already been used for training.
With the increase in the scale of model training, LLMs exhibit new capabilities, known as the
emergent abilities of LLMs. For example, Wei et al. [2022b] studied the scaling properties of
LLMs across different model sizes and amounts of computational resources. Their work shows
that some abilities emerge when we scale the model size to certain level. The appearance of
emergent abilities has demonstrated the role of scaled training in enhancing the performance of
LLMs, and it has also, to some extent, motivated researchers to continuously attempt to train larger
models. As larger and stronger LMs continue to appear, our understanding of the scaling laws
continues to mature. This helps researchers predict the performance of LLMs during training and
estimate the minimal computational resources required to achieve a given level of performance.
To understand how model performance scales with various factors considered during training,
it is common to express the model performance as a function of these factors. For example, in
the simplest case, we can express the loss or error of an LLM as a function of a single variable of
interest. However, there are no universal scaling laws that can describe this relationship. Instead,
different functions are proposed to fit the learning curves of LLMs.
Let x be the variable of interest (such as the number of model parameters) and L(x) be the
loss of the model given x (such as the cross-entropy loss on test data). The simplest form of L(x)
is a power law
L(x) = axb

(2.36)

2.2 Training at Scale

65
4.2

N
−0.076
L(N ) = ( 8.8·10
13 )

5.6

3.9

Test Loss

4.8

Test Loss

D
−0.095
L(D) = ( 5.4·10
13 )

4.0
3.2

3.6
3.3
3

2.4

2.7
5

10

7

9

10

108

10

Number of Parameters

109

Dataset Size

Fig. 2.4: Test loss against model size (N ) and training dataset size (D) (data points are plotted for illustrative purposes).

−0.076
N
We plot test loss as a function of N , which is defined as L(N ) = 8.8×10
, and a function of D, which is
13
defined as L(D) =

D
5.4×1013


−0.095

[Kaplan et al., 2020].

where a and b are parameters that are estimated empirically. Despite its simplicity, this function has successfully interpreted the scaling ability of language models and machine translation systems in terms of model size (denoted by N ) and training dataset size (denoted by D)
[Gordon et al., 2021; Hestness et al., 2017]. For example, Kaplan et al. [2020] found that the performance of their language model improves as a power law of either N or D after an initial

−0.076
N
and L(D) =
transient period, and expressed these relationships using L(N ) = 8.8×10
13

−0.095
D
(see Figure 2.4).
5.4×1013

An improvement to this scaling law is to add an irreducible error term to the power law. The
form of L(x) is then given by
L(x) = axb + ǫ∞

(2.37)

where ǫ∞ is the irreducible error that accounts for the error due to unknown variables, which is
present even as x → ∞. Eq. (2.37) is one of the most widely used forms for designing scaling
laws of LLMs. For example, Rosenfeld et al. [2020] developed a scaling law that involves both
model scaling and dataset scaling, like this
L(N, D) = aN b + cDd + ǫ∞

(2.38)

An example of such formulation is the Chinchilla scaling law. It states that the test loss per
token is the sum of the inverse proportion functions of N and D, with an additional irreducible
error term. Hoffmann et al. [2022] express this scaling law as
L(N, D) =

406.4
0.34
|N{z }

model scaling

+

410.7
0.28
|D{z }

dataset scaling

+

1.69
|{z}

(2.39)

irreducible error

All the scaling laws mentioned above are based on monotonic functions. So they cannot cover
functions with inflection points, such as double descent curves. In response, researchers have
explored more sophisticated functions to fit the learning curves. Examples of such functions can

Generative Models

66
be found in Alabdulmohsin et al. [2022] and Caballero et al. [2023]’s work.

The significance of scaling laws lies in providing directional guidance for LLM research: if
we are still in the region of the power law curve, using more resources to train larger models is a
very promising direction. While this result “forces” big research groups and companies to invest
more in computational resources to train larger models, which is very expensive, scaling laws
continuously push the boundaries of AI further away. On the other hand, understanding scaling
laws helps researchers make decisions in training LLMs. For example, given the computational
resources at hand, the performance of LLMs may be predicted.
One last note on scaling laws in this section. For LLMs, a lower test loss does not always
imply better performance on all downstream tasks. To adapt LLMs, there are several steps such
as fine-tuning and prompting that may influence the final result. Therefore, the scaling laws for
different downstream tasks might be different in practice.

2.3 Long Sequence Modeling
We have already seen that, in large-scale training, larger language models can be developed by using more data and computational resources. However, scaling up can also occur in other directions.
For instance, in many applications, LLMs are adapted to process significantly long sequences. An
interesting example is that we pre-train an LLM on extensive texts of normal length and then apply it to deal with very long token sequences, far beyond the length encountered in pre-training.
Here we use Pr(y|x) to denote the text generation probability where x is the context and y is the
generated text. There are broadly three types of long sequence modeling problems.
• Text generation based on long context (i.e., x is a long sequence). For example, we
generate a short summary for a very long text.
• Long text generation (i.e., y is a long sequence). For example, we generate a long story
based on a few keywords.
• Long text generation based on long context (i.e., both x and y are long sequences). For
example, we translate a long document from Chinese to English.
Recently, NLP researchers have been more interested in applying and evaluating LLMs on
tasks where extremely long input texts are involved. Imagine an LLM, which reads a C++ source
file containing tens of thousands of lines, and outlines the functionality of the program corresponding to the source file. Such models, capable of handling extensive textual contexts, are sometimes
called long-context LLMs. In this section we will restrict ourselves to long-context LLMs, but
the methods discussed here can be applicable to other problems.
For Transformers, dealing with long sequences is computationally expensive, as the computational cost of self-attention grows quadratically with the sequence length. This makes it infeasible
to train and deploy such models for very long inputs. Two strands of research have tried to adapt
Transformers to long-context language modeling.
• The first explores efficient training methods and model architectures to learn self-attention
models from long-sequence data.

2.3 Long Sequence Modeling

67

• The other adapts pre-trained LLMs to handle long sequences with modest or no fine-tuning
efforts.
Here, we will discuss the former briefly since it can be found in general discussions of efficient
Transformer architectures [Tay et al., 2020; Xiao and Zhu, 2023]. We will focus on the latter,
highlighting popular methods in recent LLMs. We will also discuss the strengths and limitations
of these long-sequence models.

2.3.1 Optimization from HPC Perspectives
We begin our discussion by considering improvements to standard Transformer models from the
perspectives of high-performance computing. Most of these improvements, though not specifically designed for LLMs, have been widely applied across various deep learning models [Kim et al.,
2023]. A commonly used approach is to adopt a low-precision implementation of Transformers.
For example, we can use 8-bit or 16-bit fixed-point data types for arithmetic operations, instead
of 32-bit or 64-bit floating-point data types. Using these low-precision data types can increase
the efficiency and memory throughput, so that longer sequences can be processed more easily.
An alternative approach is to improve Transformers by using hardware-aware techniques. For
example, on modern GPUs, the efficiency of Transformers can be improved by using IO-aware
implementations of the self-attention function [Dao et al., 2022; Kwon et al., 2023].
Another way to handle long sequences is through sequence parallelism [Li et al., 2023b;
Korthikanti et al., 2023]. Specifically, consider the general problem of attending the query qi
at the position i to the keys K and values V. We can divide K by rows and obtain a set of submatrices {K[1] , ..., K[nu ] }, each corresponding to a segment of the sequence. Similarly, we can
obtain the sub-matrices of V, denoted by {V[1] , ..., V[nu ] }. Then, we assign each pair of K[u] and
V[u] to a computing node (e.g., a GPU of a GPU cluster). The assigned nodes can run in parallel,
thereby parallelizing the attention operation.
Recall that the output of the self-attention model can be written as
Attqkv (qi , K, V) =

m−1
X

αi,j vj

(2.40)

j=0

where αi,j is the attention weight between positions i and j. In Transformers, αi,j is obtained
by normalizing the rescaled version of the dot product between qi and kj . Let βi,j denote the
attention score between qi and kj . We have
βi,j

=

qi · kj
√ + Mask(i, j)
d

(2.41)

where Mask(i, j) is the masking variable for (i, j). Then, we define the attention weight αi,j to
be

αi,j = Softmax(βi,j )
exp(βi,j )
= P
j ′ exp(βi,j ′ )

(2.42)

Generative Models

68

On each computing node, we need to implement these equations. Given the keys and values
assigned to this node, computing the numerator of the right-hand side of Eq. (2.42) (i.e., exp(βi,j ))
is straightforward, as all the required information is stored on the node. However, computing the
denominator of the right-hand side of Eq. (2.42) involves a sum of exp(βi,j ′ ) over all j ′ s, which
requires transferring data to and from other nodes. To illustrate, suppose that vj and kj are placed
on node u. We can rewrite Eq. (2.42) as
αi,j
=

}|

{

exp(βi,j )

X

kj ′ ∈K[1]

|

node u

z

exp(βi,j ′ ) + · · · +
{z

node 1

kj ′ ∈K[u]

|

}

X

exp(βi,j ′ ) + · · · +
{z

node u

exp(βi,j ′ )

kj ′ ∈K[nu ]

|

}

X

{z

(2.43)

}

node nu

where the notation kj ′ ∈ K[u] represents that kj ′ is a row vector of K[u] . In a straightforward
P
implementation, we first perform the summations { kj′ ∈K[u] exp(βi,j ′ )} separately on the corresponding nodes. Then, we collect these summation results from different nodes to combine them
into a final result. This corresponds to a collective operation in the context of parallel processing.
There are many efficient implementations of such operations, such as the all-reduce algorithms.
Hence the sum of all exp(βi,j ) values can be computed using optimized routines in collective
communication toolkits.
Given the attention weights {αi,j }, we then compute the attention results using Eq. (2.40).
The problem can be re-expressed as
Attqkv (qi , K, V)
=

X

vj ′

|

∈V[1]

αi,j ′ vj ′ + · · · +

{z

node 1

}

X

vj ′

|

∈V[u]

αi,j ′ vj ′ + · · · +

{z

node u

}

X

αi,j ′ vj ′

vj ′ ∈V[nu ]

|

{z

node nu

(2.44)

}

Like Eq. (2.43), Eq. (2.44) can be implemented as a summation program in parallel processing. First, perform the weighted summations of values on different nodes simultaneously. Then,
we collect the results from these nodes via collective operations.
Note that, although this section primarily focuses on long sequence modeling, much of the motivation for sequence parallelism comes from the distributed training methods of deep networks,
as discussed in Section 2.2.3. As a result, the implementation of these methods can be based on
the same parallel processing library.

2.3.2 Efficient Architectures
One difficulty of applying Transformers to long sequences is that self-attention has a quadratic
time complexity with respect to the sequence length. Moreover, a key-value cache (or KV cache
for short) is maintained during inference, and its size increases as more tokens are processed. Although the KV cache grows linearly with the sequence length, for extremely long input sequences,
the memory footprint becomes significant and it is even infeasible to deploy LLMs for such tasks.
As a result, the model architecture of long-context LLMs generally moves away from the standard

2.3 Long Sequence Modeling

69

Transformer, turning instead to the development of more efficient variants and alternatives.
One approach is to use sparse attention instead of standard self-attention. This family of
models is based on the idea that only a small number of tokens are considered important when
attending to a given token, and so most of the attention weights between tokens are close to zero.
As a consequence, we can prune most of the attention weights and represent the attention model
in a compressed form. To illustrate, consider the self-attention model
(2.45)

Attqkv (Q, K, V) = α(Q, K)V
where the attention weight matrix α(Q, K) ∈ Rm×m is obtained by
QKT
α(Q, K) = Softmax( √ + Mask)
d

α0,0
0
0
 α
α
0
1,1
 1,0

α2,0
α2,1
α2,2
= 

..
..
 ..
 .
.
.

αm−1,0 αm−1,1 αm−1,2

h

...
...
...
..
.

0
0
0
..
.

...

αm−1,m−1










(2.46)

i

Each row vector αi,0 ... αi,i 0 ... 0 corresponds to a distribution of attending the i-th
token to every token of the sequence. Since language models predict next tokens only based on
their left-context, we normally write the output of the attention model at position i as

Attqkv (qi , K≤i , V≤i ) =

=

v
i  0

... αi,i  ... 


h

αi,0

i
X





vi

(2.47)

αi,j vj

j=0









v0
k0
 . 
 . 



where K≤i =  ..  and V≤i =  .. 
 are the keys and values up to position i.
ki

vi

h

i

In the original version of self-attention αi,0 ... αi,i is assumed to be dense, that is, most of
h

i

the values are non-zero. In sparse attention, some of the entries of αi,0 ... αi,i are considered
non-zero, and the remaining entries are simply ignored in computation. Suppose G ⊆ {0, ..., i} is
the set of indices of the non-zero entries. For language models, the output of the sparse attention
model at position i is given by
Attsparse (qi , K≤i , V≤i ) =

X

α′i,j vj

(2.48)

j∈G

Here {α′i,j } are normalized over G. Hence their values are different from the original attention
weights (in fact we have α′i,j > αi,j ). The sparsity of the model is determined by how large G is.
Sparse attention models differ in the way we define G. One simple approach is to define G based

Generative Models

70

on heuristically designed patterns. For example, a widely-used pattern involves having G cover a
window of tokens located near position i [Parmar et al., 2018].
While sparse attention reduces the computation through the use of sparse operations, such
models still have significant limitations as we must keep the entire KV cache (i.e., K≤i and V≤i )
during inference. If the sequence is very long, storing this cache will become highly memoryintensive. To address this, we can consider a different form of attention models where the KV
cache is not explicitly retained. Linear attention is one such approach [Katharopoulos et al., 2020].
It uses a kernel function φ(·) to project each query and key onto points qi′ = φ(qi ) and ki′ = φ(ki ),
respectively. By removing the Softmax function under such transformations10 , the form of the
resulting attention model is given by

Attqkv (qi , K≤i , V≤i ) ≈ Attlinear (qi′ , K′≤i , V≤i )
=

qi′ µi
qi′ νi

(2.49)

where µi and νi are variables that are computed in the recurrent forms

T

(2.50)

T
νi−1 + k′ i

(2.51)

µi = µi−1 + k′ i vi
νi =

µi and νi can be seen as representations of the history up to position i. A benefit of this model is
that we need not keep all past queries and values. Instead only the latest representations µi and
νi are used. So the computational cost of each step is a constant, and the model can be easily
extended to deal with long sequences.
In fact, this sequential approach to long sequence modeling arises naturally when we adopt a
viewpoint of recurrent models. Such models read one token (or a small number of tokens) at a
time, update the recurrent state using these inputs, and then discard them before the next token
arrives. The output at each step is generated based only on the recurrent state, rather than on all the
previous states. The memory footprint is determined by the recurrent state which has a fixed size.
Recurrent models can be used in real-time learning scenarios where data arrives in a stream and
predictions can be made at any time step. In NLP, applying recurrent models to language modeling is one of the earliest successful attempts to learn representations of sequences. Although
Transformer has been used as the foundational architecture in LLMs, recurrent models are still
powerful models, especially for developing efficient LLMs. More recently, recurrent models have
started their resurgence in language modeling and have been reconsidered as a promising alternative to Transformers [Gu and Dao, 2023]. Figure 2.5 shows a comparison of the models discussed
in this subsection.

10
In the new space after this transformation, the Softmax normalization can be transformed into the simple scaling
normalization.

2.3 Long Sequence Modeling

71
Attqkv (qi , K≤i , V≤i )

k0

k1

···

ki−2

ki−1

ki

v0

v1

···

vi−2

vi−1

vi

qi

(a) Standard Self-attention

Attqkv (qi , {k1 , ki }, {v1 , vi })

k0

k1

···

ki−2

ki−1

ki

v0

v1

···

vi−2

vi−1

vi

qi

(b) Sparse Attention

T

⇒

µi

νi = νi−1 + k′ i

⇒

νi

µi = µi−1 + k′ i vi
T

q′ µ

Attlinear (qi , K≤i , V≤i ) = qi′ νii
i

k0

k1

···

ki−2

ki−1

ki

v0

v1

···

vi−2

vi−1

vi

qi

(c) Linear Attention

hi = f (hi−1 , inputi )
h0

h1

···

hi−3

hi−2

(d) Recurrent Models

hi−1

hi
inputi

Fig. 2.5: Illustrations of self-attention, sparse attention, linear attention and recurrent models. Blue boxes = cached
states for producing the output at position i. f (·) = a recurrent cell.

2.3.3 Cache and Memory
LLMs based on the standard Transformer architecture are global models. The inference for these
models involves storing the entire left-context in order to make predictions for future tokens. This
requires a KV cache where the representations (i.e., keys and values) of all previously-generated

Generative Models

72

tokens are kept, and the cost of caching grows as the inference proceeds. Above, we have discussed methods for optimizing this cache via efficient attention approaches, such as sparse attention and linear attention. Another idea, which may have overlap with the previous discussion, is
to explicitly encode the context via an additional memory model.

2.3.3.1 Fixed-size KV Cache
A straightforward approach is to represent the keys and values using a fixed-size memory model.
Suppose we have a memory Mem which retains the contextual information. We can write the
attention operation at position i in a general form
(2.52)

Att(qi , Mem) = Attqkv (qi , K≤i , V≤i )

In this model, Mem is simply the KV cache, i.e., Mem = (K≤i , V≤i ). Thus the size of
Mem is determined by i. If we define Mem as a fixed-size variable, then the cost of performing
Att(qi , Mem) will be fixed. There are several alternative ways to design Mem.
• One of the simplest methods is to consider a fixed-size window of previous keys and values.
Mem is therefore given by
(2.53)

Mem = (K[i−nc +1,i] , V[i−nc +1,i] )

where nc denotes the size of the window. The notation K[i−nc +1,i] and V[i−nc +1,i] denote
the keys and values over positions from i − nc + 1 to i.11 This model can be seen as a type
of local attention model.
• It is also possible to define Mem as a pair of summary vectors, which leads to a more
compressed representation of the history. A simple way to summarize the previous keys
and values is to use the moving average of them. For example, Mem can be defined as the
unweighted moving average of the previous nc keys and values
 Pi

j=i−nc +1 kj

Mem =

nc

,

Pi

j=i−nc +1 vj

nc



(2.54)

Alternatively, we can use a weighted version of moving average
Mem =

 Pi

j=i−nc +1 βj−i+nc kj
P nc
,
j=1 βj

Pi

j=i−nc +1 βj−i+nc vj
Pnc
j=1 βj



(2.55)

Here {β1 , ..., βnc } are the coefficients, which can be either learned as model parameters
or determined via heuristics. For example, they can be set to increasing coefficients (i.e.,
β1 < β2 < ... < βnc −1 < βnc ) in order to give larger weight to positions that are closer to








vi−nc +1
ki−nc +1




.
.
11
..
More formally, we write K[i−nc +1,i] = 
 and V[i−nc +1,i] =  .. . Sometimes we denote
vi
ki
K[i−nc +1,i] by {ki−nc +1 , ..., ki } and V[i−nc +1,i] by {vi−nc +1 , ..., vi } for notation simplicity.

2.3 Long Sequence Modeling

73

i. We can extend the moving average to include all the positions up to i. This leads to the
cumulative average of the keys and values, given in the form
Mem =

 Pi

j=0 kj

i+1

,

Pi

j=0 vj

i+1



(2.56)

In general, the cumulative average can be written using a recursive formula
Memi =

(ki , vi ) + i · Memi−1
i+1

(2.57)

where Memi and Memi−1 denote the cumulative averages of the current and previous positions, respectively. An advantage of this model is that we only need to store a single
key-value pair during inference, rather than storing all the key-value pairs. Note that the
above memory models are related to recurrent models, and more advanced techniques have
been used to develop alternatives to self-attention mechanisms in Transformers [Ma et al.,
2023].
• The memory Mem can also be a neural network. At each step, it takes both the previous
output of the memory and the current states of the model as input, and produces the new
output of the memory. This neural network can be formulated as the function
Mem = Update(Skv , Mempre )

(2.58)

Here Mem and Mempre represent the outputs of the memory at the current step and the
previous step, respectively. Skv is a set of key-value pairs, representing the recent states of
the model. This formulation is general and allows us to develop various memory models by
selecting different Update(·) and Skv configurations. For example, if Skv only contains the
latest key-value pair (ki , vi ) and Update(·) is defined as a recurrent cell, then Eq. (2.58)
can be expressed as an RNN-like model
Mem = f ((ki , vi ), Mempre )

(2.59)

where f (·) is a recurrent cell. Recurrence can also be applied to segment-level modeling
for efficiency consideration. A simple approach is that we can divide the sequence into
segments, and treat Skv as a segment. Applying recurrent models to Update(·) will result in
memory models that operate on segments. A special example is that we define Update(·) as
an FIFO function that adds Skv into the memory and removes the oldest key-value segment
from the memory, given by
Mem = FIFO(Skv , Mempre )

(2.60)

Consider a memory which includes two segments, one for current segment, and one for the
previous segment. In the attention operation, each position can access the history key-value
pairs in two closest consecutive segments. This essentially defines a local memory, but it
and its variants have been widely used segment-level recurrent models [Dai et al., 2019;
Hutchins et al., 2022; Bulatov et al., 2022].
• The above memory models can be extended to involve multiple memories. An example

Generative Models

74

of this approach is compressive Transformer [Rae et al., 2019]. It employs two distinct
fixed-size memories: one for modeling local context (denoted by Mem), and the other for
modeling and compressing long-term history (denoted by CMem). The KV cache in this
model is the combination of Mem and CMem. The attention function can be written as
Attcom (qi , Mem, CMem) = Attqkv (qi , [Mem, CMem])

(2.61)

where [Mem, CMem] is a combined memory of Mem and CMem. As with other segmentlevel models, the compressive Transformer model operates on segments of the sequence.
k as the key-value
Each segment is a sequence of ns consecutive tokens, and we denote Skv
pairs corresponding to the tokens of the k-th segment. When a new segment arrives, Mem
k to Mem, and then
is updated in an FIFO fashion: we append the nc key-value pairs in Skv
pop the ns oldest key-value pairs from Mem, which is given by
k
Mem = FIFO(Skv
, Mempre )

(2.62)

The popped key-value pairs are then used to update the compressive memory CMem. These
ns key-value pairs are compressed into ncs key-value pairs via a compression network.
CMem is an FIFO which appends the compressed ncs key-value pairs to the tail of the
queue, and drops the first ncs key-value pairs of the queue. It is given by
k
CMem = FIFO(Ckv
, CMempre )

(2.63)

k represents the set of compressed key-value pairs. Implicit in the compressive
where Ckv
Transformer model is that local context should be represented explicitly with minimal information loss, while long-range context can be more compressed.

• We have already seen that both global and local contexts are useful and can be modeled using attention models. This view motivates the extension to attention models for
combining both local and long-term memories [Ainslie et al., 2020; Zaheer et al., 2020;
Gupta and Berant, 2020]. A simple but widely-used approach is to involve the first few tokens of the sequence in attention, serving as global tokens. This approach is usually applied
along with other sparse attention models. An advantage of incorporating global tokens of
the sequence is that it helps smooth the output distribution of the Softmax function used in
attention weight computation, and thus stabilizes model performance when the context size
is very large [Xiao et al., 2024]. One drawback, however, is that using a fixed-size global
memory may result in information loss. When dealing with long sequences, we need to
enlarge the KV cache for sufficient representations of the context, but this in turn increases
the computational cost.

Figure 2.6 shows illustrations of the above approaches. Note that, while we focus on optimization of the KV cache here, this issue is closely related to those discussed in the previous section.
All of the methods we have mentioned so far can broadly be categorized as efficient attention
approaches, which are widely used in various Transformer variants.

2.3 Long Sequence Modeling

75
Memory
Size = 4 × 2

···

Keys

···

Values
i−7

i−6

i−5

i−4

i−3

i−2

i−1

i

(a) Window-based Cache

ki−3 +ki−2 +ki−1 +ki
4

⇒

vi−3 +vi−2 +vi−1 +vi
4

⇒

Memory
Size = 1 × 2

···

Keys

···

Values
i−7

i−6

i−5

i−4

i−3

i−2

i−1

i

(b) Moving Average-based Cache

Mem = Update( Skv , Mempre )

Memory
Size = 1 × 2

⇒

···

Keys

···

Values
i−7

i−6

i−5

i−4

i−3

i−2

i−1

i

(c) Recurrent Network as Cache

Compressed
Memory
Size = 2 × 2

Memory
Size = 4 × 2

···

Keys

···

Values
i−7

i−6

i−5

i−4

i−3

i−2

i−1

i

(d) Hybrid Cache (Compressed Memory + Local Memory)

Fig. 2.6: Illustrations of fixed-size KV caches in LLMs. Blue boxes represent the keys and values generated during
LLM inference, green boxes represent the keys and values stored or encoded in the primary memory, and orange boxes
represent the keys and values stored or encoded in the compressed memory.

Generative Models

76
2.3.3.2 Memory-based Models

The modeling of memories discussed above was based on updates to the KV cache, and the resulting models are typically referred to as internal memories. We now consider another family
of models, called external memories, which operate as independent models to access large-scale
contexts for LLMs. Many such models are based on memory-based methods which have been
extensively discussed in machine learning [Bishop, 2006]. A common example is nearest neighbor algorithms: we store context representations in a datastore, and try to find the most similar
stored representations to match a given query. The retrieved context representations are then used
to improve attention for this query.
Here, we consider the k-nearest neighbors (k-NN) method which is one of the most popular
memory-based methods. Since our focus is language modeling in this section, we define a sample
in the datastore as a key-value pair corresponding to some context state. Note that “context” is a
broad concept here, not just a sequence prefix in text generation. One might, for example, view
the entire dataset as the context for predicting tokens. This allows us to retrieve the closest context
situation in a set of sequences, rather than a given sequence prefix. Although we will restrict
ourselves to context modeling for a single sequence, in this subsection, we discuss a relatively
more general case.
Suppose we have a set of keys {kj } with corresponding values {vj }, and suppose we store
these key-value pairs in a vector database12 . For each query qi , we find its k nearest neighbours by
growing the radius of the sphere centered as qi until it contains k data points in {kj }. This results
in a set of k keys along with their corresponding values, denoted by Memknn . As before, we
denote Mem as the local memory for the query, such as the KV cache of neighboring tokens. Our
goal is to attend query qi to both the local memory Mem and the long-term memory Memknn .
There are, of course, several ways to incorporate Mem and Memknn into the attention model.
For example, we might simply combine them to form a single KV cache [Mem, Memknn ], and
attend qi to [Mem, Memknn ] via standard QKV attention. Or we might use Mem and Memknn
in separate attention steps. An example of such approaches is the model developed by Wu et al.
[2021]. It linearly combines the two types of attention, given by
Att(qi , Mem, Memknn ) = g ⊙ Attlocal + (1 − g) ⊙ Attknn

(2.64)

Attlocal = Att(qi , Mem)

(2.65)

Attknn = Att(qi , Memknn )

(2.66)

Here g ∈ Rd is the coefficient vector, which can be the output of a learned gate.

Given the k-NN-based memory model described above, the remaining task is to determine
which key-value pairs are retained in the datastore. For standard language modeling tasks, we
consider the previously seen tokens in a sequence as the context, so we can add the keys and
values of all these tokens into the datastore. In this case, the resulting k-NN-based attention
model is essentially equivalent to a sparse attention model [Gupta et al., 2021].
Alternatively, we can extend the context from one sequence to a collection of sequences.
For example, we might collect all key-value pairs across the sequences in a training dataset and
add them to the datastore to model a larger context. Thus, LLMs can predict tokens based on a
12
A vector database, or vector store, is a database that provides highly optimized retrieval interfaces for finding stored
vectors that closely match a query vector.

2.3 Long Sequence Modeling

77

generalized context. A problem with this approach is that the computational cost would be large
if many sequences are involved. Since these sequences are part of our training data, we can build
and optimize an index for the vectors in the datastore before running the LLMs. As a result, the
retrieval of similar vectors can be very efficient, as in most vector databases.
In fact, all the above-mentioned methods can be viewed as instances of a retrieval-based approach. Instead of using retrieval results to improve attention, we can apply this approach in other
ways as well. One application of k-NN-based search is k-NN language modeling (or k-NN LM)
[Khandelwal et al., 2020]. The idea is that, although it is attempting to extend the context used
in self-attention by incorporating nearest neighbors in representation learning, in practice, similar
hidden states in Transformers are often highly predictive of similar tokens in subsequent positions.
In k-NN LM, each item in the datastore is a key-value tuple (z, w), where z represents a hidden
state of the LLM at a position, and w represents the corresponding prediction. A typical way to
create the datastore is to collect the output vector of the Transformer layer stack and the corresponding next token for each position of each sequence in a training dataset. During inference,
we have a representation hi given a prefix. Given this representation, we first search the datastore
for k closest matching data items {(z1 , w1 ), ..., (zk , wk )}. Here {w1 , ..., wk } are thought of as
reference tokens for prediction, and thus can be used to guide the token prediction based on hi .
One common way to make use of reference tokens is to define a distribution over the vocabulary
V,
h

i

Prknn (·|hi ) = Softmax( −d0 · · · −d|V | )

(2.67)

where dv equals the distance between hi and zj if wj equals the v-th entry of V , and equals 0
otherwise. We use a linear function with a coefficient λ that interpolates between the retrievalbased distribution Prknn (·|hi ) and the LLM output distribution Prlm (·|hi )
Pr(·|hi ) = λ · Prknn (·|hi ) + (1 − λ) · Prlm (·|hi )

(2.68)

Then, as usual, we can choose the next token y by maximizing the probability Pr(y|hi ).
As with information retrieval (IR) systems, the datastore can also manage texts and provide
access to relevant texts for a query. For example, we can store a collection of text documents
in a search engine with full-text indexing, and then search it for documents that match a given
text-based query. Applying IR techniques to LLMs leads to a general framework called retrievalaugmented generation (RAG). The RAG framework works as follows. We use the context x as
the query and find the k most relevant document pieces {c1 , ..., ck } from the datastore via efficient
IR techniques13 . These search results are combined with the original context via a prompting

13
In piratical applications, queries are typically generated using a query generation system, which may expand it
with variations of tokens and query intent.

Generative Models

78
template g(·)14 , resulting in an augmented input for the LLM
x′ = g(c1 , ..., ck , x)

(2.69)

Then, we use x′ as the context and predict the following text using the model Pr(y|x′ ). One
advantage of RAG is that we need not modify the architecture of LLMs, but instead augment the
input to LLMs via an additional IR system. Figure 2.7 shows a comparison of the use of different
external memories in LLMs.

2.3.3.3 Memory Capacity
A memory model in LLMs, in the form of a simple key-value cache or a datastore, can broadly
be seen as an encoder of contextual information. Ideally, before we say that a memory model
is representative of the entire context in token prediction, we need to make sure that the model
can accurately represent any part of the context. The standard KV cache is one such model that
completely stores all past history. In this case, the model is said to have adequate capacity for
memorizing the context. In many practical applications, however, complete memorization is not
required. Instead, the goal is to enable LLMs to access important contextual information. As a
result, efficient and compressed memory models are developed, as described in this section. Note
that, the longer the sequence, the more difficult it becomes for a low-capacity memory model to
capture important contextual information. It is therefore common practice to simply increase the
model capacity when processing long contexts.
While high-capacity models are generally favorable, they are difficult to train and deploy. A
challenging scenario is that the tokens arrive in a stream and the context continuously grows.
Developing LLMs for such tasks is difficult as we need to train Transformers on extremely long
sequences. A possible way to address this difficulty is to use non-parametric methods, such as
retrieval-based methods. For example, as discussed above, we can use a vector database to store
previously generated key-value pairs, and thus represent the context by this external memory
model. Although this approach side-steps the challenge of representing long context in Transformers, building and updating external memory models are computationally expensive. These
models are more often used in problems where the context is given in advance and fixed during
inference, and hence unsuitable for streaming context modeling.
In cases where the size of the context continuously grows, applying fixed-size memory models
is a commonly used approach. For example, in recurrent models, a sequence of arbitrary length
can be summarized into a set of hidden states by which we have a fixed computational cost per
step. While recurrent models were initially found to be not very good at handling long-distance
dependencies in sequence modeling in early applications of deep learning to NLP, recent advancements have shown that their variants are now effective in modeling extremely long sequences.
[Bulatov et al., 2022; Hutchins et al., 2022; Munkhdalai et al., 2024; Ma et al., 2024].
14

For example, the template could be:
message = {*c1 *} ... {*ck *}
input: {*x*}
output:

2.3 Long Sequence Modeling

79
g ⊙ Att(qi , Mem) + (1 − g) ⊙ Att(qi , Memknn )

Att(qi , Memknn )

Att(qi , Mem)
···
···

qi

KV Cache

k Nearest
Neighbors

Keys/values in LLM

Datastore

Search

Keys/values in Datastore

(a) k-NN Search Augmented Attention

Output Distribution

Distribution Pr(·)
Distribution Prknn (·)

···
Att(qi , Mem)

Att(qi , Mem)
···
···

qi

KV Cache

k Nearest
Neighbors

Keys/values in LLM

Datastore

Keys in Datastore

Search

Predicted Tokens

(b) k-NN Language Modeling

LLM

c1 = Deep network is ...

Message: deep network ... machine learning ...

c2 = Machine learning is ...

What is deep learning?

···
k Nearest
Neighbors

Datastore

Search

Input Context:
x = What is deep learning?

(c) Retrieval-augmented Generation

Fig. 2.7: Illustrations of external memories (or datastores) for language modeling.

Generative Models

80

There is no general definition of memory capacity in LLMs. A simple approach might consider
how much storage is used to retain contextual information. For example, memory capacity could
be defined by the size of the KV cache in Transformers or the vector database used in retrievalbased methods. A related concept is model complexity. In machine learning, there are several
ways to define the model complexity of a model. One of the simplest methods is by counting the
number of parameters. However, it should be emphasized that the memory models discussed here
primarily serve to store information, rather than add trainable parameters. Therefore, a model with
a large memory capacity is not necessarily more complex. Nevertheless, in practice determining
the capacity of a memory model is not straightforward. In general, we need to control the trade-off
between maximizing the performance and controlling the memory footprint.

2.3.4 Sharing across Heads and Layers
In Transformers, the KV cache is a data structure that can be dynamically adjusted along multiple
dimensions, such as heads, layers, and sequence length. For example, consider an LLM with L
layers. Each layer has τ attention heads, and each head produces a dh -dimensional output. During
inference, we store the keys and values for up to m tokens. The space complexity of this caching
mechanism is O(L · τ · dh · m). As we have seen previously, this complexity can be reduced by
caching the keys and values for fewer tokens. For example, in sliding window attention, a fixedsize window is used to cache the keys and values in local context. And this model has a space
complexity of O(L · τ · dh · mw ), with mw being the size of the window.

In addition to reducing m, we can also decrease the size of the KV cache along other dimensions. A widely-used approach is to enable sharing across heads in multi-head self-attention.
Recall from Section 2.1.1 that multi-head self-attention uses multiple sets of queries, keys, and
values (each set is called a head), each performing the QKV attention mechanism as usual. This
can be expressed as
Output = Merge(head1 , ..., headτ )Whead

(2.70)

where headj ∈ Rdh is computed using the standard QKV attention function
[j]

[j]

[j]

headj = Attqkv (qi , K≤i , V≤i )
[j]

[j]

(2.71)

[j]

Here, qi , K≤i , and V≤i are the query, keys, and values that are projected onto the j-th feature
sub-space. So this model can be interpreted as performing attention on a group of feature subspaces in parallel (see Figure 2.8 (b)). The KV cache needs to retain the keys and values for all
[1]
[1]
[τ ]
[τ ]
these heads, that is, {(K≤i , V≤i ), ..., (K≤i , V≤i )}.

One refinement to the multi-head attention model, called multi-query attention (MQA), is to
share keys and values across heads, while allowing queries to be unique for each head [Shazeer,
2019]. In MQA, there is a single set of keys and values (K≤i , V≤i ). In addition, there are τ
[1]
[τ ]
queries {qi , ..., qi }, each corresponding to a different head. For each head, we have
[j]

headj = Attqkv (qi , K≤i , V≤i )

(2.72)

Figure 2.8 (c) illustrates this model. By sharing keys and values, the size of the KV cache would

2.3 Long Sequence Modeling
value

key

81
query

(b) Multi-head Attention

(a) Single-head Attention

value

key

query

query

key

value

(c) Multi-query Attention

value

query

key

value

(d) Grouped Query Attention

query

key

Layer l

Sharing

Layer l − 1
(e) Cross-layer Multi-head Attention

Fig. 2.8: Illustration of QKV attention based on different multi-head and sharing mechanisms. (a) = single-head
attention, and (b-e) = attention with multiple heads.

be O(L · dh · m).

Grouped query attention (GQA) is a natural extension to multi-head attention and MQA
[Ainslie et al., 2023]. In GQA, heads are divided into ng groups, each corresponding to a shared
[n ]
[n ]
[1]
[1]
set of keys and values. Hence we have ng sets of keys and values {(K≤i , V≤i ), ..., (K≤ig , V≤ig )}.
See Figure 2.8 (d) for an illustration. Let g(j) be the group id for the j-th head. The GQA model
can be expressed as
[j]

[g(j)]

[g(j)]

headj = Attqkv (qi , K≤i , V≤i )

(2.73)

The size of the KV cache of GQA is O(L·ng ·dh ·m). One benefit of GQA is that we can trade-off
between computational efficiency and model expressiveness by adjusting ng . When ng = τ , the
model becomes the standard multi-head attention model. By contrast, when ng = 1, it becomes

Generative Models

82
the GQA model.

Sharing can also be performed across layers. Such a method falls into the family of shared
weight and shared activation methods, which have been extensively used in Transformers [Dehghani et al.,
2018; Lan et al., 2020]. For example, one can share KV activations or attention weights across
layers to reduce both computation and memory footprints [Xiao et al., 2019; Brandon et al., 2024].
Figure 2.8 (e) shows an illustration of this method, where a query in a layer directly accesses the
KV cache of a lower-level layer.

2.3.5 Position Extrapolation and Interpolation
Since Transformer layers are order-insensitive to input, we need some way to encode positional
information in the input tokens. To do this, it is common to add positional embeddings to token
embeddings, and then feed these combined embeddings into the Transformer layer stack as input.
In this case, the embedding at position i can be expressed as
ei = xi + PE(i)

(2.74)

where xi ∈ Rd denotes the token embedding, and PE(i) ∈ Rd denotes the positional embedding.
In general, the token embedding xi is a position-independent vector, and so the positional embedding PE(i) is used to encode the positional context. A straightforward approach is to treat PE(i)
as a learnable variable and train it alongside other model parameters. In this way, we can learn
a unique representation for each position, and thus distinguish the tokens appearing at different
positions of a sequence.
Representations of positions using learned vectors can work well in tasks where the sequences
at training and test times are of similar lengths. In practice, however, we often impose length
restrictions on sequences during training to prevent excessive computational costs, but wish to
apply the trained models to much longer sequences during inference. In this case, using learned
positional embeddings has obvious drawbacks, as there are no trained embeddings for positions
that are not observed in the training phase.
An alternative approach to modeling positional information is to develop positional embeddings that can generalize: once trained, the embedding model can be used to handle longer sequences. Suppose that we train a positional embedding model on sequences with a maximum
length of ml , and we wish to apply the trained model to a sequence of length m (m >> ml ). If
the embedding model is limited in the range of positions that we can observe from training data,
then this model will simply fail to deal with new data outside that range. See Figure 2.9 (a) for
an illustration where the learned embedding model cannot model data points outside the training
domain if it lacks the ability to extrapolate.
There are several approaches to making positional embedding models generalize. They can
be grouped into two classes.
• Extrapolation. The model learned on observed data points (i.e., positions) can be directly
employed to assign meaningful values to data points beyond the original range. For example, suppose we have a series of numbers 1, 2, ..., 10, and we want to understand the
meaning of a new number, 15. Knowing that these numbers are natural numbers used for
ordering, we can easily infer that 15 is a number that follows 10, even though 15 has not

2.3 Long Sequence Modeling

83

Value

1
0

−1

0

1,024
Sequence Length
(a) Encoding with No Generalization

2,048

0

1,024
Sequence Length
(b) Extrapolation

2,048

0

1,024
Sequence Length
(c) Interpolation

2,048

Value

1
0

−1

Value

1
0

−1

Fig. 2.9: Illustrations of different positional embedding methods for a range of positions. Blue points represent the
positions that have been observed during training, and red points represent the positions that are newly observed at test
time. In sub-figure (a), the encoding model only memorizes the points seen during training, and cannot generalize. In
sub-figures (b) and (c), the model can generalize through extrapolation and interpolation.

been observed before. Figure 2.9 (b) shows an example of this approach, where a function
is learned to fit the data points within a specific range and then applied to estimate the values
of data points outside that range.
• Interpolation. This approach maps a larger range of data points into the original observation range. For example, suppose we have a model designed for numbers in the range
[1, 10]. When given a new range of [1, 20], we can scale this down by dividing every number by 2, thereby fitting all numbers into [1, 10]. This scaling allows us to use the model
trained on the range [1, 10] to describe data points in the expanded range of [1, 20]. See
Figure 2.9 (c) for an illustration of this approach.

In fact, positional embeddings in many systems have achieved some level of generalization.
For example, sinusoidal encoding, the most common positional embedding method, employs sine
and cosine functions that can naturally extend to sequences of any length. Although this approach
might seem direct and simple, it does not perform well when we significantly extend the sequences
for processing. In this subsection, we will discuss several alternative methods based on either
extrapolation or interpolation.

Generative Models

84
2.3.5.1 Attention with Learnable Biases

One problem with Eq. (2.74) is that the embedding model treats each token independently and
therefore ignores the distance between different tokens. A common improvement to this model,
called relative positional embedding, is to consider the pairwise relationship between tokens
[Shaw et al., 2018]. The general idea behind this is to obtain the offset between any pair of positions and incorporate it into the self-attention model. One of the simplest forms of self-attention
with relative positional embedding is given by
Attqkv (qi , K≤i , V≤i ) =

i
X

α(i, j)vj

(2.75)

j=0

α(i, j) = Softmax(

qi kjT + PE(i, j)
√
+ Mask(i, j))
d

(2.76)

The only difference between this model and the original self-attention model is that a bias term
PE(i, j) is added to the query-key product in this new model. Intuitively, PE(i, j) can be interpreted as a distance penalty for the pair of positions i and j. As i moves away from j, the value of
PE(i, j) decreases.
PE(i, j) can be defined in several different ways. Here, we consider the T5 version of relative
positional embedding, called the T5 bias [Raffel et al., 2020]. For each pair of query qi and key
kj , the offset between them is defined to be15
d(i, j) = i − j

(2.77)

A simple design for the bias PE(i, j) is to share the same learnable variable for all query-key
pairs with the same offset, i.e., PE(i, j) = ui−j , where ui−j is the variable corresponding to
the offset i − j. However, simply assigning a unique value to each offset will restrict this model
to observed offsets. When i − j is larger than the maximum trained offset, the model cannot
generalize.
The T5 bias instead adopts a generalization of this model. Rather than assigning each querykey offset a unique bias term, it groups difference offsets into “buckets”, each corresponding to
one learnable parameter. More specifically, the bias terms for nb + 1 buckets are given as follows.
• For buckets 0 to nb2+1 − 1, each bucket corresponds to one offset, that is, bucket 0 ↔ offset
0, bucket 1 ↔ offset 1, bucket 2 ↔ offset 2, and so on. We express this as b(i − j) = i − j.
• For buckets nb2+1 to nb , the size of each bucket increases logarithmically. For example, the
bucket number for a given offset i − j ≥ nb2+1 can be defined as
b(i − j) =

log(i − j) − log( nb2+1 ) nb + 1
nb + 1
+⌊
⌋
·
2
2
log(distmax ) − log( nb2+1 )

(2.78)

where the parameter distmax is typically set to a relatively large number to indicate the
15

For language modeling, a query is only allowed to attend to its left-context, and so we have i − j ≥ 0. In the more
general case of self-attention, where a token can attend to all tokens in the sequence, we may have negative offsets
when i < j.

2.3 Long Sequence Modeling

85
logarithmically increased bucket size

fixed bucket size

Bucket

0

1

2

3

Offset

0

1

2

3

···

14

15

16

17

14 15 16 ∼ 20 21 ∼ 26

18

27 ∼ 33

···

32

802 ∼ ∞

(i − j)

Fig. 2.10: Illustration of distributing query-key offsets into buckets in the T5 model (nb = 32 and distmax = 1024).
Boxes represent buckets. In the first half of the buckets, we use a fixed bucket size. In the second half of the buckets,
we increase the bucket size logarithmically. The last bucket contains all the query-key offsets that are not covered by
previous buckets.

maximum offset we may encounter.

• When i − j > distmax , we place i − j in the last bucket. In other words, bucket nb contains
all the offsets that are not assigned to the previous buckets.

Together, these can be expressed as the function
b(i − j)

=



i − j

n +1

min(nb , b2 + ⌊

n +1
log(i−j)−log( b2 )
n +1
log(distmax )−log( b2 )

· nb2+1 ⌋)

0 ≤ i − j < nb2+1
i − j ≥ nb2+1

(2.79)

Figure 2.10 shows an illustration of these buckets. We see that in the first half of the buckets,
each bucket is associated with only one value of i − j, while in the second half, the bucket size
increases as i − j grows. The last bucket is designed to handle sequences of arbitrarily long
lengths.
All PE(i, j)s in a bucket share the same bias term ub(i−j) . Substituting PE(i, j) = ub(i−j)
into Eq. (2.76), the attention weight for qi and kj becomes16
α(i, j) = Softmax(

qi kjT + ub(i−j)
√
+ Mask(i, j))
d

(2.81)

The parameters {u0 , ..., unb } are learned as common parameters during training. It should
be emphasized that this model can generalize to long sequences. This is because PE(i, j)s with
similar query-key offsets share the same parameter, and this sharing strategy is particularly important for achieving good generalization, given that large query-key offsets are rare in training.
In practice, we often set nb to a moderate number, and thus it can help control the overfitting of
positional embedding models.
16

Note that, in Raffel et al. [2020]’s T5 model, the rescaling operation for the query-key product is removed. The
attention weight α(i, j) is then given by
α(i, j)

=

Softmax(qi kjT + ub(i−j) + Mask(i, j))

(2.80)

Generative Models

86
2.3.5.2 Attention with Non-learned Biases

Relative positional embedding models are based on a set of learned biases for the query-key product in self-attention. An alternative approach is to give these biases fixed values via heuristics,
rather than training them on a particular dataset. One benefit of this heuristics-based approach is
that it does not rely on a training process and thus can be directly applied to any sequences once
the biases are set.
One example of such an approach is Press et al. [2022]’s approach, called attention with
linear biases or ALiBi for short. In the ALiBi approach, the bias term is defined as the negative
scaled query-key offset
PE(i, j) = −β · (i − j)
= β · (j − i)

(2.82)

where β is the scaling factor. Adding this term to the query-key product, we obtain a new form of
attention weights
α(i, j) = Softmax(

qi kjT + β · (j − i)
√
+ Mask(i, j))
d

(2.83)

This model can be interpreted as adding a fixed penalty to qi kjT whenever j moves one step
away from i. So we do not need to adapt it to a range of sequence lengths, and can employ it to
model arbitrarily long sequences. See Figure 2.11 for a comparison of the T5 bias and the ALiBi
bias.
In general, the scalar β should be tuned on a validation dataset. However, Press et al. [2022]
found that setting β to values decreasing geometrically by a factor of 21a for multi-head attention
performs well on a variety of tasks. Specifically, for a self-attention sub-layer involving nhead
heads, the scalar for the k-th head is given by
βk =

1
8

2k

(2.84)

The ALiBi approach provides a simple form of relative positional embeddings. There are
other similar methods for designing query-key biases using the offset i − j. Table 2.4 shows a
comparison of such biases. As an aside it is worth noting that the form of the right-hand side
of Eq. (2.82) is very similar to length features used in conventional feature-based systems. For
example, in statistical machine translation systems, such features are widely used to model word
reordering problems, resulting in models that can generalize well across different translation tasks
[Koehn, 2010].

2.3.5.3 Rotary Positional Embedding
As with sinusoidal embeddings, rotary positional embeddings are based on hard-coded values for
all dimensions of an embedding [Su et al., 2024]. Recall that in the sinusoidal embedding model,
positions are represented as combinations of sine and cosine functions with different frequencies.
These embeddings are then added to token embeddings to form the inputs to the Transformer

2.3 Long Sequence Modeling

87

qi kjT

Bias (ub(i−j) )

q0 kT
0

u0

T
q1 kT
0 q1 k1

u1

u0

T
T
q2 kT
0 q2 k1 q2 k2

u2

u1

u0

u2

u2

u1

u0

T
T
T
T
q4 kT
0 q4 k1 q4 k2 q4 k3 q4 k4

u3

u2

u2

u1

u0

T
T
T
T
T
q5 kT
0 q5 k1 q5 k2 q5 k3 q5 k4 q5 k5

u3

u3

u2

u2

u1

u0

T
T
T
T
T
T
q6 kT
0 q6 k1 q6 k2 q6 k3 q6 k4 q6 k5 q6 k6

u3

u3

u3

u2

u2

u1

+

T
T
T
q3 kT
0 q3 k1 q3 k2 q3 k3

u0

(a) The T5 bias (nb = 3 and distmax = 5)

qi kjT

Bias (−β(i − j))

q0 kT
0

0

T
q1 kT
0 q1 k1

−1β

T
T
q2 kT
0 q2 k1 q2 k2

−2β −1β

+

T
T
T
q3 kT
0 q3 k1 q3 k2 q3 k3

0
0

−3β −2β −1β

0

T
T
T
T
q4 kT
0 q4 k1 q4 k2 q4 k3 q4 k4

−4β −3β −2β −1β

0

T
T
T
T
T
q5 kT
0 q5 k1 q5 k2 q5 k3 q5 k4 q5 k5

−5β −4β −3β −2β

−β

0

T
T
T
T
T
T
q6 kT
0 q6 k1 q6 k2 q6 k3 q6 k4 q6 k5 q6 k6

−6β −5β −4β −3β −2β

−β

0

(b) The ALiBi bias

Fig. 2.11: Query-key products with biases (above = the T5 bias and below = the ALiBi bias). The color scale of the
biases ranges from light blue denoting small absolute values to deep blue denoting large absolute values.

layer stack. Rotary positional embeddings instead model positional context as rotations to token
embeddings in a complex space. This leads to a model expressed in the form of multiplicative
embeddings
ei = xi R(i)

(2.85)

where R(i) ∈ Rd×d is the rotation matrix representing the rotations performed on the token
embedding xi ∈ Rd .

For simplicity, we will first consider embeddings with only two dimensions and return to a
discussion hof the more
i general formulation later. Suppose we have a 2-dimensional token embedding x = x1 x2 . We can represent it as a vector in a plane, originating at the origin (0, 0)
and terminating at (x1 , x2 ). A counterclockwise rotation of this vector refers to an operation of

Generative Models

88
Entry

Query-Key Bias (PE(i, j))

T5 [Raffel et al., 2020]
ALiBi [Press et al., 2022]

ub(i−j)
−β · ( i − j )

−β1 ( i − j )β2

Kerple [Chi et al., 2022]

(power)

−β1 log(1 + β2 ( i − j ))
¯
Pd/2

Sandwich [Chi et al., 2023]

k=1 cos

FIRE [Li et al., 2024]

(logarithmic)



¯
( i − j )/100002k/d




f ψ( i − j )/ψ(max(mlen , i))

¯ and mlen are hyper-parameters. In the T5
Table 2.4: Query-key biases as relative positional embeddings. β, β1 , β2 , d,
model, b(i − j) denotes the bucket assigned to i − j. In the FIRE model, ψ(·) is a monotonically increasing function
such as ψ(x) = log(cx + 1), and f (·) is an FFN.

moving the vector around the origin while maintaining its magnitude, as shown in Figure 2.12 (a).
The degree of rotation is usually defined by a specific angle, denoted by θ. The rotation can be
expressed mathematically in the form
Ro(x, θ) = xRθ
h

=

x1 x2

h

=
"

i

"

cos θ sin θ
− sin θ cos θ

#

cos θ · x1 − sin θ · x2 sin θ · x1 + cos θ · x2

#

i

(2.86)

cos θ sin θ
is the rotation matrix. If two or more rotations are performed on the
where Rθ =
− sin θ cos θ
same vector, we can rotate the vector further. This follows from the fact that the composition of
successive rotations is itself a rotation. More formally, rotating a vector by an angle θ for t times
can be expressed as
Ro(x, tθ) = xRtθ
=

h

cos tθ · x1 − sin tθ · x2 sin tθ · x1 + cos tθ · x2

i

(2.87)

If we interpret t as the position of a token represented by x in a sequence, then we will find
that the above equation defines a simple positional embedding model. As shown in Figure 2.12
(b), we start moving the token from position 0. Each time we move one step forward, the vector
is rotated by the angle θ. Upon arriving at the position t, the representation of the token with
positional context is given by Ro(x, iθ). As the rotations do not change the magnitude of the
embedding, the original “meaning” of the token is retained. The positional information is injected
into the embedding, when it gets rotated.
A popular way to understand
h vectori rotation is to define it in complex spaces. It is easy
to transform each vector x = x1 x2 in the 2D Euclidean space R2 to a complex number
x′ = x1 + ix2 in the complex space C via a bijective linear map. Then, the rotation of x with the
angle tθ corresponds to the multiplication by eitθ . Given that eitθ = cos tθ + i sin tθ, the rotation

2.3 Long Sequence Modeling

89

x2

x2
vector x

rotated vector
xRθ

x

xRθ

θ

θ
θ

x1
xR2θ

x1

θ
xR3θ

(b) Multi-step Rotation

(a) Single-step Rotation

The1 cat2 is3 sleeping4 peacefully5
in6 the7 warm8 sunlight9 .10

x2

sleeping4

7θ cat2
7θ

x1

sleeping11

Every1 afternoon2 ,3 you4 ’ll5 find6 that7
the8 cat9 is10 sleeping11 on12 my13 bed14 .15

cat9

(c) Angles between embeddings of two tokens at different positions

Fig. 2.12: Illustrations of vector rotations in a plane. Sub-figures (a) and (b) show rotations of a vector in a single
step and multiple steps, respectively. Sub-figure (c) shows the embeddings of tokens cat and sleeping in two different
sentences. We show these sentences with a subscript affixed to each token to indicate its position. If we represent
tokens as vectors, we can add positional information by rotating these vectors. This rotation preserves the “distances”
between the vectors. For example, given that the distance between cat and sleeping is the same in both sentences, the
angle between their embeddings also remains the same during rotation.

operation can be re-expressed in the form
xRtθ 7→ x′ eitθ
=

(x1 + ix2 )(cos tθ + i sin tθ)

=

cos tθ · x1 − sin tθ · x2 + i(sin tθ · x1 + cos tθ · x2 )

(2.88)

Here we denote the token representation x′ eitθ by C(x, tθ). The inner product of the representations of the tokens at positions t and s can be written as
hC(x, tθ), C(y, sθ)i = (x′ y′ )ei(t−s)θ

(2.89)

where y′ is the complex conjugate of y′ . As can be seen, the result of this inner product involves
a term t − s, and so it can model the offset between the two tokens.

Generative Models

90

Now we go back to representations in the 2D Euclidean space. The dot-product of Ro(x, tθ)
and Ro(y, sθ) is can be written as a function of (t − s)θ
Ro(x, tθ)[Ro(y, sθ)]T = xRtθ [yRsθ ]T
= xRtθ [Rsθ ]T yT
= xR(t−s)θ yT

(2.90)

Given this result, if we consider Ro(x, tθ) and Ro(y, sθ) as the query and the key, then the selfattention operation will implicitly involve the modeling of relative positional context.
This rotary positional embedding canh be extended to imulti-dimensional embeddings. For
a d-dimensional token embedding x = x1 x2 ... xd , we can treat it as a d2 -dimensional
h

i

h

i

complex vector x′ = x′1 x′2 ... x′d/2 = x1 + ix2 x3 + ix4 ... xd−1 + ixd , where
each consecutive pair of items forms a complex number. Then, the rotary positional embedding in
the complex space is given by
C(x, tθ) =

d/2
X

x′k eitθk ~ek

(2.91)

k=1

where ~ek is the standard basis vector with a single non-zero value in the k-th coordinate and 0’s
elsewhere [Biderman et al., 2021].
Although this formula involves a complicated expression, its equivalent form in the d-dimensional
Euclidean space is relatively easy to understand. We can write it as

Ro(x, tθ) =



i


h

Rtθ2

x1 x2 ... xd 


"



Rtθ1



..

.
Rtθd/2







(2.92)

#

i
h
cos tθk sin tθk
where Rtθk =
. θ = θ1 , ..., θd/2 are the parameters for controlling the an− sin tθk cos tθk

gles of rotations in different dimensions. Typically, θk is set to 10000−
to the setting in sinusoidal embeddings.

2(k−1)
d

, which is analogous

In a practical implementation, Eq. (2.92) can be rewritten into a form that relies solely on the
element-wise product and addition of vectors.
T





T



T



T

sin tθ1
−x2
cos tθ1
x1
 sin tθ 
 x 
 cos tθ 
 x 
1 
1 

 1 

 2 


 . 


 . 
.
.








.
.
.
.
Ro(x, tθ) =  .  ⊙ 
.
.

 + .  ⊙


xd−1 

xd



cos tθd/2 

cos tθd/2



 −xd 

xd−1



sin tθd/2 

sin tθd/2

Finally, we rewrite Eq. (2.85) to obtain the form of the embedding at position i

(2.93)

2.3 Long Sequence Modeling

91

(2.94)

ei = Ro(xi , iθ)

2.3.5.4 Position Interpolation
In position interpolation, our goal is to map the positions in the new sequence to match the observed range in training. Suppose the sequence length for training ranges from 0 to ml . When
m > ml at test time, we represent the positions in [0, m] such that our representations fit [0, ml ].
To illustrate, consider the rotary positional embedding model described
above.
The embedding
i
h
of each token is described by a model Ro(xi , iθ) in which θ = θ1 , ..., θd/2 are the parameters.
Ro(xi , iθ) can be cast in the form of a linear combination of two periodic functions (see Eq.
(2.93))
cos iθ =
sin iθ =

h

cos iθ1 ... cos iθd/2

h

sin iθ1 ... sin iθd/2

i

i

(2.95)
(2.96)

θk is a exponential function of k and takes the form
θ k = b−

2(k−1)
d

(2.97)

where b is the base. The period of cos iθk and sin iθk is
Tk = 2π · b

2(k−1)
d

(2.98)

The key idea behind position interpolation is to adjust this period so that the new positions can
m
be encoded within the range [0, ml ]. One way to achieve this is to scale up Tk by m
, given by
l
Tk′ =

2(k−1)
m
· 2π · b d
ml

(2.99)

Hence all points in [0, m] are compressed into [0, ml ]. This linear scaling can be easily realized
by modifying the input to the embedding model [Chen et al., 2023c]. The new model with linear
positional interpolation is given by
Ro′ (xi , iθ) = Ro(xi ,

ml
iθ)
m

(2.100)

Another method of positional interpolation is to scale the base17 . Suppose that the base b is
scaled by λ. We wish the period of this new model in the last dimension of θ (i.e., dimension d2 )
to be equal to that of the linear positional interpolation model. This can be expressed as
2π · (λb)

2( d
2 −1)
d

=

2( d
m
2 −1)
· 2π · b d
ml

(2.101)

17
This method was first proposed in https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/
ntkaware_scaled_rope_allows_llama_models_to_have/

Generative Models

92
Solving this equation, we obtain
λ =
=

m 
 2( dd−1)
2
ml
d
m 
 d−2
ml

(2.102)

This gives an embedding model
Ro′ (xi , iθ) = Ro(xi , iθ ′ )

(2.103)

where
h

0

2

d−2

θ ′ = (λb)− d , (λb)− d , ..., (λb)− d

i

(2.104)

Note that scaling the base provides a non-uniform method for scaling the periods across different dimensions of θ. This method has been found to be helpful for extending LLMs to longer
sequences, and several improvements have been developed [Peng et al., 2024; Ding et al., 2024].

2.3.6 Remarks
In this section, we have presented a variety of methods for long-context language modeling. We
close this section by discussing some interesting issues related to these methods.

2.3.6.1 Need for Long Context
One of the ultimate goals of long-context LLMs is that these models can precisely encode infinite
context. The so-called infinite context refers more to the fact that an LLM can continuously read
words. This motivates LLMs that can handle extremely long context or stream data. As discussed
in Section 2.3.3, it is common to use fixed-size memory models to process continuously expanding
context. Many such systems are based on recurrent architectures or their variants, because they
are inherently suited to model time series problems where the effects of past inputs continue
indefinitely. Another way to achieve infinite memory is to develop alternatives to self-attention
models, for example, one can use continuous-space attention models to encode context, which
removes the dependency on context length [Martins et al., 2022].
When studying long-context LLMs, it is natural to wonder what mechanisms may explain the
use of long context in language modeling. Can we compress the representation of infinite context
into a relatively small-sized model? Are all context tokens useful for predicting next tokens? How
do LLMs prepare for token prediction when they see the context? Can we know in advance which
contextual information will be critical for prediction? General answers to all these questions
are not obvious, but they inspire follow-on research of explainable models, and some interesting
results have been found. For example, Deletang et al. [2024] conducted extensive experiments
to show that LLMs are powerful in-context compressors. Although viewing predictive models
as compression models has long been studied in machine learning, it also provides insights into
our understanding of the LLM scaling laws. Pal et al. [2023] and Wu et al. [2024] investigated
whether the features learned up to the current step, though not intentionally, are already sufficient

2.3 Long Sequence Modeling

93

for predicting tokens at the following steps. Note that the need for long-context in language
modeling is highly dependent on the problem that we address. A related issue is where to apply
LLMs and how to evaluate them. For example, in summarization tasks we may only need to distill
and focus on a few key aspects of the text, while in retrieval-like tasks we need to “memorize”
the entire context so that the relevant information can be accessed. We will discuss the evaluation
issue later in this subsection.

2.3.6.2 Pre-training or Adapting LLMs?
Training LLMs requires significant computational costs. Although it is straightforward to train
LLMs on long sequence data, the training becomes computationally unwieldy for large data sets. It
is common practice to pre-train LLMs on general datasets, and then adapt them with modest finetuning effort. For example, LLMs with relative or rotary positional embeddings can be directly
trained on large-scale data in the pre-training phase. While the resulting models may exhibit some
abilities to extrapolate lengths in the inference phase, it may be more effective to fine-tune them
on longer sequences.
Ideally, we would like to pre-train LLMs with standard Transformer architectures and adapt
them to new tasks. This allows us to use many off-the-shelf LLMs and efficiently adapt them to
handle long sequences. However, when new architectures are adopted, it seems inevitable that
we need to train these models from scratch. This poses practical difficulties for developing longcontext LLMs, as we cannot leverage well-developed, pre-trained models and must instead train
them ourselves. On the other hand, fine-tuning is still an effective way to adapt LLMs with certain
architectures that are different from those in pre-training. An example is models augmented with
external memories. In these models, the pre-trained LLMs are fixed, and the focus is on how
to make these LLMs collaborate with the memory models. In RAG, for instance, it is common
to fine-tune LLMs to improve their use of retrieval-augmented inputs. Another example of finetuning LLMs for long-context modeling is that we train an LLM with full attention models, and
then replace them with sparse attention models in the fine-tuning phase. The pre-trained LLM
provides initial values of model parameters used in a different model, and this model is then finetuned as usual.

2.3.6.3 Evaluating Long-context LLMs
Evaluating long-context LLMs is important, but it is a new issue in NLP. The general idea is that,
if we input a long context to an LLM, then we can check from the output of the LLM whether it
understands the entire context and makes use of it in predicting following tokens. In conventional
research of NLP, such evaluations are often aimed at examining the ability of NLP models in
handling long-range dependencies. However, the size of contexts used in recent LLMs is much
larger than that used in NLP systems a few years ago. This motivates researchers to develop new
evaluation benchmarks and metrics for long-context LLMs.
One approach is to use the perplexity metric. However, in spite of its apparent simplicity, this
method tends to reflect more on the LLMs’ ability to make use of local context rather than global
context. It is therefore tempting to develop evaluation methods that are specific to long-context
LLMs. Popular methods include various synthetic tasks where artificially generated or modified

Generative Models

94

data is used to evaluate specific capabilities of long-context LLMs. In needle-in-a-haystack18 and
passkey retrieval tasks [Mohtashami and Jaggi, 2024; Chen et al., 2023c], for instance, LLMs are
required to identify and extract a small, relevant piece of information from a large volume of given
text. The assumption here is that an LLM with sufficient memory should remember earlier parts
of the text as it processes new information. This LLM can thus pick out the relevant details, which
might be sparse and hidden among much irrelevant information, from the text. Alternatively,
in copy memory tasks (or copy tasks for short), LLMs are used to repeat the input text or a
specific segment multiple times. These tasks were initially proposed to test the extent to which
recurrent models can retain and recall previously seen tokens [Hochreiter and Schmidhuber, 1997;
Arjovsky et al., 2016], and have been adopted in evaluating recent LLMs [Bulatov et al., 2022;
Gu and Dao, 2023].
Another approach to evaluating long-context LLMs is to test them on NLP tasks that involve
very long input sequences. Examples include long-document or multi-document summarization,
long-document question answering, code completion, and so on. A benefit of this approach is that
it can align evaluations with user expectations.
Although many methods have been developed, there is still no general way to evaluate longcontext LLMs [Liu et al., 2024c]. One problem is that most of these methods focus on specific
aspects of LLMs, rather than their fundamental ability to model very long contexts. Even though
an LLM can pick out the appropriate piece of text from the input, we cannot say that it truly understands the entire context. Instead, it might just remember some important parts of the context,
or even simply recall the answer via the model learned in pre-training. Moreover, the data used
in many tasks is small-scale and relatively preliminary, leading to discrepancies between evaluation results and actual application performance. A more interesting issue is that the results of
LLMs are influenced by many other factors and experimental setups, for example, using different
prompts can lead to very different outcomes. This makes evaluation even more challenging because improvements may not solely result from better modeling of long contexts, and there is a
risk of overclaiming our results. Nevertheless, many open questions remain in the development
and evaluation of long-context LLMs. For example, these models still suffer from limitations
such as restricted context length and high latency. Studying these issues is likely to prove valuable
future directions.

2.4 Summary
In this chapter, we have discussed the concept of LLMs and related techniques. This can be considered a general, though not comprehensive, introduction to LLMs, laying the foundation for further
discussions on more advanced topics in subsequent chapters. Furthermore, we have explored two
ways to scale up LLMs. The first focuses on the large-scale pre-training of LLMs, which is crucial for developing state-of-the-art models. The second focuses on methods for adapting LLMs to
long inputs, including optimizing attention models, designing more efficient and compressed KV
caches, incorporating memory models, and exploring better positional embeddings.
The strength of LLMs lies in their ability to break the constraints of training NLP models for
a limited number of specific tasks. Instead, LLMs learn from large amounts of text through the
simple task of token prediction — we predict the next token in a sentence given its prior tokens.
18

https://github.com/gkamradt/LLMTest_NeedleInAHaystack

2.4 Summary

95

A general view is that, by repeating this token prediction task a large number of times, LLMs can
acquire some knowledge of the world and language, which can then be applied to new tasks. As a
result, LLMs can be prompted to perform any task by framing it as a task of predicting subsequent
tokens given prompts. This emergent ability in language models comes from several dimensions,
such as scaling up training, model size, and context size. It is undeniable that scaling laws are
currently the fundamental principle adopted in developing large language models, although simply increasing model size has yet to prove sufficient for achieving AGI. These continuously scaled
LLMs have been found to show capabilities in general-purpose language understanding, generation, and reasoning. More recently, it has been found that scaling up the compute at inference time
can also lead to significant improvements in complex reasoning tasks [OpenAI, 2024].
Given their amazing power, LLMs have attracted considerable interest, both in terms of techniques and applications. As a result, the explosion of research interest in LLMs has also led to a
vast number of new techniques and models. However, we do not attempt to provide a comprehensive literature review on all aspects of LLMs, given the rapid evolution of the field. Nevertheless,
one can still gain knowledge about LLMs from general reviews [Zhao et al., 2023; Minaee et al.,
2024] or more focused discussions on specific topics [Ruan et al., 2024].

C HAPTER 3

Prompting
In the context of LLMs, prompting refers to the method of providing an LLM with a specific input
or cue to generate a desired output or perform a task. For example, if we want the LLM to translate
a sentence from English to Chinese, we can prompt it like this
Translate the text from English to Chinese.
Text: The early bird catches the worm.
Translation:
Prompting is crucial for LLMs because it directly influences how effectively these models understand and respond to user queries. A well-crafted prompt can guide an LLM to generate more
accurate, relevant, and contextually appropriate responses. Furthermore, this process can be iteratively refined. By analyzing the responses of the LLM, users can adjust their prompts to align
more closely with their specific needs. Given the importance of prompting in applying LLMs,
prompt design has become an essential skill for users and developers working with LLMs. This
leads to an active research area, called prompt engineering, in which we design effective prompts
to make better use of LLMs and enhance their practical utility in real-world applications.
An important concept related to prompting is in-context learning. When prompting an LLM,
we can add new information to the context, such as demonstrations of problem-solving. This
allows the LLM to learn from this context how to solve the problem. Here is an example of
prompting LLMs with a few demonstrations of how to classify text based on sentiment polarity.
Here are some examples of text classification.
Example 1: We had a delightful dinner together. → Label: Positive
Example 2: I’m frustrated with the delays. → Label: Negative
What is the label for “That comment was quite hurtful.”?
Label:
In-context learning is often seen as an emergent ability of LLMs that arises after pre-training.
Though LLMs can be trained or tuned to perform new tasks, in-context learning provides a very
efficient way to adapt these models without any training or tuning effort. Perhaps this is one of
the most notable features of LLMs: they indeed learn general knowledge about the world and
language during pre-training, which we can easily apply to new challenges. Moreover, in-context
learning reflects the broader trend of making AI systems more generalizable and user-friendly.
Instead of requiring specialized engineers to fine-tune models for every unique task, users can
interact with LLMs in a more intuitive way, simply providing examples or adjusting the context
as needed.
In this chapter, we focus on prompting techniques in LLMs. We begin by considering several
interesting prompt designs commonly used in prompt engineering. Then, we discuss a series of
96

3.1 General Prompt Design

97

refinements to these methods. Finally, we explore approaches for automating prompt design.

3.1 General Prompt Design
This section presents basic concepts in prompt design, along with examples of how to prompt
LLMs for various NLP tasks. Since the effectiveness of prompting is highly dependent on the
LLMs being used, prompts often vary across different LLMs, making it difficult to provide a
comprehensive list of prompts for all LLMs and downstream tasks. Therefore, this discussion is
not focused on any specific LLM. Instead, the goal is to provide guiding principles for prompt
design.

3.1.1 Basics
The term prompt is used in many different ways. In this chapter we define a prompt as the input
text to an LLM, denoted by x. The LLM generates a text y by maximizing the probability Pr(y|x).
In this generation process, the prompt acts as the condition on which we make predictions, and it
can contain any information that helps describe and solve the problem.
A prompt can be obtained using a prompt template (or template for short) [Liu et al., 2023a].
A template is a piece of text containing placeholders or variables, where each placeholder can
be filled with specific information. Here are two templates for asking the LLM for weekend
suggestions.
Please give me some suggestions for a fun weekend.

If {∗premise∗}, what are your suggestions for a fun weekend.

In the first template, we simply instruct the LLM to return some suggestions. So the template
is just a piece of text with no variables. In the second template, the variable {∗premise∗} needs to
be specified by the users to provide a premise for making suggestions. For example, if we input
premise = the weather is nice this weekend
then we can generate a prompt
If the weather is nice this weekend,
what are your suggestions for a fun weekend.

We can also design a template with multiple variables. Here is an example in which we
compare the two sentences in terms of their semantic similarity.

Prompting

98

Here is a sentence
{∗sentence1∗}
Here is another sentence
{∗sentence2∗}

Compute the semantic similarity between the two sentences

A popular way to format prompts is to write each input or output in a “name:content” style.
For example, we can describe a conversation between two people, named John and David, and use
the LLM to continue the conversation. A template of such prompts is given by
John: {∗utterance1∗}
David: {∗utterance2∗}
John: {∗utterance3∗}
David: {∗utterance4∗}
John: {∗utterance5∗}
David: {∗utterance6∗}
John: {∗utterance7∗}
David:
The “name:content” format can be used to define the task that we want the LLM to perform.
For example, given that “Q” and “A” are commonly used abbreviations for “Question” and “Answer”, respectively, we can use the following template to do question-answering.
Q: {∗question∗}
A:
This format can be used to describe more complex tasks. For example, the following is an
example of providing a specification for a translation task
Task: Translation
Source language: English
Target language: Chinese
Style: Formal text
Template: Translate the following sentence: {∗sentence∗}

In practical systems, it is common to represent and store such data in key-value pairs, such as the
JSON format1 .
When the problem is difficult to describe in an attribute-based manner, it is more common
to instruct LLMs with a clear and detailed description. There are many ways to do this. One
1

The JSON representation is

3.1 General Prompt Design

99

example is to assign a role to LLMs and provide sufficient context. The following is a template
that instructs an LLM to act as an expert and answer questions from children.

You are a computer scientist with extensive knowledge in the field
of deep learning.
Please explain the following computer-related concept to a child
around 10 years old, using simple examples whenever possible.
{∗concept∗}

Here the text “You are a computer scientist ... deep learning. ” is sometimes called system
information, and is provided to help the LLM understand the context or constraints of the task it
is being asked to perform.

3.1.2 In-context Learning
Learning can occur during inference. In-context learning is one such method, where prompts
involve demonstrations of problem-solving, and LLMs can learn from these demonstrations how
to solve new problems. Since we do not update model parameters in this process, in-context
learning can be viewed as a way to efficiently activate and reorganize the knowledge learned in
pre-training without additional training or fine-tuning. This enables quick adaptation of LLMs to
new problems, pushing the boundaries of what pre-trained LLMs can achieve without task-specific
adjustments.
In-context learning can be illustrated by comparing three methods: zero-shot learning, oneshot learning and few-shot learning. Zero-shot learning, as its name implies, does not involve a
traditional “learning” process. It instead directly applies LLMs to address new problems that were
not observed during training. In practice, we can repetitively adjust prompts to guide the LLMs in
generating better responses, without demonstrating problem-solving steps or providing examples.
Consider the following example. Suppose we want to use an LLM as an assistant that can help
correct English sentences. A zero-shot learning prompt is given by

{
"Task": "Translation"
"Source language": "English"
"Target language": "Chinese"
"Style": "Formal text"
"Template": "Translate the following sentence: {∗sentence∗}"
}

Prompting

100

SYSTEM
USER

You are a helpful assistant, and are great at grammar correction.
You will be provided with a sentence in English. The task is
to output the correct sentence.
Input: She don’t like going to the park.
Output:

Here the gray words are used to indicate different fields of the prompt.
In one-shot learning, we extend this prompt by adding a demonstration of how to correct
sentences, thereby allowing the LLM to learn from this newly-added experience.

SYSTEM
DEMO

You are a helpful assistant, and are great at grammar correction.
You will be provided with a sentence in English. The task is
to output the correct sentence.
Input: There is many reasons to celebrate.
Output: There are many reasons to celebrate.

USER

You will be provided with a sentence in English. The task is
to output the correct sentence.
Input: She don’t like going to the park.
Output:

Furthermore, we can add more demonstrations to enable few-shot learning.

SYSTEM

You are a helpful assistant, and are great at grammar correction.

DEMO1

You will be provided with a sentence in English. The task is
to output the correct sentence.
Input: There is many reasons to celebrate.
Output: There are many reasons to celebrate.

DEMO2

You will be provided with a sentence in English. The task is
to output the correct sentence.
Input: Me and my friend goes to the gym every day.
Output: My friend and I go to the gym every day.

USER

You will be provided with a sentence in English. The task is
to output the correct sentence.
Input: She don’t like going to the park.
Output:

In few-shot learning, we essentially provide a pattern that maps some inputs to the corresponding outputs. The LLM attempts to follow this pattern in making predictions, provided that
the prompt includes a sufficient number of demonstrations, although generally small. It is also

3.1 General Prompt Design

101

possible to use simpler patterns to achieve this. For example, one can use the following few-shot
learning prompt for translating words from Chinese to English.

DEMO

USER

现在
来
去
男孩
女孩

→
→
→
→
→

now
come
go
boy

If the LLM is powerful enough, few-shot learning can enable it to address complex problems, such as mathematical reasoning. For example, consider the following task of summing two
numbers and then dividing the sum by their product.

DEMO

USER

12 5
3 1
−9 4
15 15
19 73

→
→
→
→
→

(12 + 5)/(12 × 5) = 0.283
(3 + 1)/(3 × 1) = 1.33
(−9 + 4)/(−9 × 4) = 0.138
(15 + 15)/(15 × 15) = 0.133

In many practical applications, the effectiveness of in-context learning relies heavily on the
quality of prompts and the fundamental abilities of pre-trained LLMs. On one hand, we need a
significant prompt engineering effort to develop appropriate prompts that help LLMs learn more
effectively from demonstrations. On the other hand, stronger LLMs can make better use of incontext learning for performing new tasks. For example, suppose we wish to use an LLM to
translate words from Inuktitut to English. If the LLM lacks pre-training on Inuktitut data, its
understanding of Inuktitut will be weak, and it will be difficult for the model to perform well in
translation regardless of how we prompt it. In this case, we need to continue training the LLM
with more Inuktitut data, rather than trying to find better prompts.
It might be interesting to explore how in-context learning emerges during pre-training and
why it works during inference. One simple understanding is that LLMs have gained some knowledge of problem-solving, but there are many possible predictions, which are hard to distinguish
when the models confront new problems. Providing demonstrations can guide the LLMs to follow the “correct” paths. Furthermore, some researchers have tried to interpret in-context learning from several different perspectives, including Bayesian inference [Xie et al., 2022], gradient
decent [Dai et al., 2023; Von Oswald et al., 2023], linear regression [Akyürek et al., 2023], meta
learning [Garg et al., 2022], and so on.

3.1.3 Prompt Engineering Strategies
Designing prompts is highly empirical. In general, there are many ways to prompt an LLM for
performing the same task, and we need to perform a number of trial-and-error runs to find a
satisfactory prompt. To write good prompts more efficiently, one can follow certain strategies.
Examples of common prompting principles include

Prompting

102

• Describing the task as clearly as possible. When we apply an LLM to solve a problem,
we need to provide a precise, specific, and clear description of the problem and instruct the
LLM to perform as we expect. This is particularly important when we want the output of
the LLM to meet certain expectations. For example, suppose we are curious about climate
change. A simple prompt for asking the LLM to provide some information is

Tell me about climate change.

Since this instruction is too general, the LLM may generate a response that addresses any
aspect of climate change, which may not align with our specific interests. In this case, we
can instead use prompts that are specific and detailed. One such example is

Provide a detailed explanation of the causes and effects of climate change,
including the impact on global temperatures, weather patterns, and sea
levels. Also, discuss possible solutions and actions being taken to mitigate
these effects.

Now suppose we intend to explain climate change to a 10-year-old child. We can adjust the
above prompt further.

Explain the causes and effects of climate change to a 10-year-old child.
Talk about how it affects the weather, sea levels, and temperatures. Also,
mention some things people are doing to help. Try to explain in simple
terms and do not exceed 500 words.

• Guiding LLMs to think. LLMs have exhibited surprisingly good capabilities to “think”.
A common example is that well-developed LLMs have achieved impressive performance
in mathematical reasoning tasks, which are considered challenging. In prompt engineering,
the “thinking” ability of LLMs needs to be activated through appropriate prompting, especially for problems that require significant reasoning efforts. In many cases, an LLM that
is instructed to “think” can produce completely different results compared with the same
LLM that is instructed to perform the task straightforwardly. For example, Kojima et al.
[2022] found that simply appending “Let’s think step by step” to the end of each prompt
can improve the performance of LLMs on several reasoning tasks. LLMs can be prompted
to “think” in a number of ways. One method is to instruct LLMs to generate steps for reasoning about the problem before reaching the final answer. For example, consider a task of
solving mathematical problems. See below for a simple prompt for this task.

3.1 General Prompt Design

103

You are a mathematician. You will be provided with a math problem.
Please solve the problem.

Since solving math problems requires a detailed reasoning process, LLMs would probably
make mistakes if they attempted to work out the answer directly. So we can explicitly ask
LLMs to follow a given reasoning process before coming to a conclusion.

You are a mathematician. You will follow these detailed reasoning steps
when solving math problems.
Step 1: Problem Interpretation.
The mathematician carefully listens to your query and understands the intricate details of the mathematical challenge you have presented.
Step 2: Strategy Formulation.
Drawing upon their extensive knowledge, the mathematician chooses the
most effective strategy tailored to the type of math problem, whether it is
algebra, calculus, or geometry.
Step 3: Detailed Calculation.
With precision and expertise, the mathematician performs the necessary
calculations step by step, adhering to all mathematical principles.
Step 4: Solution Review.
Before providing the final answer, the mathematician meticulously checks
the calculations for accuracy and offers a concise explanation or rationale
for the solution.
You will be provided with a math problem. Please solve the problem.
{∗problem∗}

Another method to guide LLMs to “think” is through multiple rounds of interaction with
LLMs. For example, as a first step, we can instruct LLMs to solve the problem directly

You will be provided with a math problem. Please solve the problem.
{∗problem∗}

Now we have an initial answer to the problem. As a second step, we prompt LLMs to
evaluate the correctness of the answer and, if necessary, rework it to find a better solution.

Prompting

104

You will be provided with a math problem, along with a solution. Evaluate
the correctness of this solution, and identify any errors if present. Then,
work out your own solution.
Problem: {∗problem∗}
Solution: {∗solution∗}

The prompts presented here are closely related to a long line of research on reasoning problems in LLMs. It is impossible to provide a complete discussion of all related issues because
this topic covers a large family of methods. But we will see a relatively more detailed discussion on how to improve prompting through more reasoning in Section 3.2.

• Providing reference information. As discussed in the previous section, we can include
demonstrations in prompts and allow LLMs to in-context learn from these demonstrations
how to perform the task. In fact, given the remarkable ability of language understanding of
LLMs, we can add any type of text into the prompts and so these models can predict based
on enriched contexts. In many applications, we have various information that is relevant
to user queries. Instead of using LLMs to make unconstrained predictions, we often want
LLMs to produce outputs that are confined to the relevant text. One such example is RAG,
where the relevant text for the user query is provided by calling an IR system, and we
prompt LLMs to generate responses based on this provided relevant text. The following
prompt shows an example.

You are an expert that can generate answers to input queries. You have now
been provided with a query and the corresponding context information.
Please generate an answer based on this context information. Note that
you need to provide the answer in your own words, not just copy from the
context provided.
Context information: {∗IR-result∗}
Query: {∗query∗}

If the context information is highly reliable, we can even restrict LLMs to answering using
only the provided text. An example prompt is shown as follows

3.1 General Prompt Design

105

You are an expert tasked with generating answers from input queries. You
have been provided with a query and corresponding context information,
organized in a table where each row represents a useful record. Please
generate an answer using only this context information. Ensure that you
provide the answer in your own words.
Context information: {∗table∗}
Query: {∗query∗}

When dealing with real-world problems, we often have prior knowledge and additional
information about the problems that help produce better answers. Considering such information in prompting is generally helpful in improving the result.
• Paying attention to prompt formats. In general, the performance of LLMs is highly
sensitive to the prompts we input. Sometimes a small modification to a prompt can lead to a
big change in model output. An interesting example is that changing the order of sentences
in a prompt may cause LLMs to generate different results. To make prompts easy to read
and reduce ambiguity, it is common to format them in a way that ensures clarity. One
example is that we define several fields for prompts and fill different information in each
field. Another example is we can use code-style prompts for LLMs which can understand
and generate both natural language and code. See the following for a code-style prompt that
performs translation where one demonstration is presented.

[English] = [I have an apple.]
[German] = [Ich habe einen Apfel.]
[English] = [I have an orange.]
[German] =

LLMs can receive text in various formats. This allows us to use control characters, XML
tags, and specific formatting to represent complex data. And it is useful to specify how the
input and output should be formatted or structured. For example, we can delimit sections of
text using quotes and prompt LLMs accordingly (e.g., adding a sentence like “the input text
is delimited by double quotes” to the prompt).

Above, we have discussed only a few strategies for writing good prompts. There are, of course,
many such methods, and one needs to develop their own through practice. Interested readers can
refer to various online documents for more information, such as OpenAI’s manual on the GPT
series models2 .
2
See
https://platform.openai.com/docs/guides/prompt-engineering/
six-strategies-for-getting-better-results.

Prompting

106

3.1.4 More Examples
In this subsection, we consider more examples of prompting LLMs to perform various NLP tasks.
The motivation here is not to give standard prompts for these tasks, but rather to use simple
examples to illustrate how LLMs can be prompted to deal with NLP problems.

3.1.4.1 Text Classification
Text classification is perhaps one of the most common problems in NLP. Many tasks can be
broadly categorized as assigning pre-defined labels to a given text. Here we consider the polarity
classification problem in sentiment analysis. We choose polarity classification for illustration because it is one of the most popular and well-defined text classification tasks. In a general setup of
polarity classification, we are required to categorize a given text into one of three categories: negative, positive, or neutral. Below is a simple prompt for doing this (for easy reading, we highlight
the task description in the prompt).
Analyze the polarity of the following text and classify it as positive, negative, or
neutral.
Text:
The service at the restaurant was slower than expected, which was a bit frustrating.
The polarity of the text can be classified as positive.
To make the example complete, we show the response generated by the LLM (underlined text).
Although the answer is correct, the LLM gives this answer not in labels but in text describing
the result. The problem is that LLMs are designed to generate text but not to assign labels to text
and treat classification problems as text generation problems. As a result, we need another system
to map the LLM’s output to the label space (call it label mapping), that is, we extract “positive”
from “The polarity of the text can be classified as positive”. This is trivial in most cases because
we can identify label words via simple heuristics. But occasionally, LLMs may not express the
classification results using these label words. In this case, the problem becomes more complicated,
as we need some way to map the generated text or words to predefined label words.
One method to induce output labels from LLMs is to reframe the problem as a cloze task. For
example, the following shows a cloze-like prompt for polarity classification.
Analyze the polarity of the following text and classify it as positive, negative, or
neutral.
Text:
The service at the restaurant was slower than expected, which was a bit frustrating.
The polarity of the text is positive

3.1 General Prompt Design

107

We can use LLMs to complete the text and fill the blank with the most appropriate word. Ideally, we wish the filled word would be positive, negative, or neutral. However, LLMs are not
guaranteed to generate these label words. One method to address this problem is to constrain the
prediction to the set of label words and select the one with the highest probability. Then, the output
label is given by
label = arg max Pr(y|x)

(3.1)

y∈Y

where y denotes the word filled in the blank, and Y denotes the set of label words
{positive, negative, neutral}.

Another method of using LLMs to generate labels is to constrain the output with prompts. For
example, we can prompt LLMs to predict within a controlled set of words. Here is an example.

Analyze the polarity of the following text and classify it as positive, negative, or
neutral.
Text:
The service at the restaurant was slower than expected, which was a bit frustrating.
What is the polarity of the text?
Just answer: positive, negative, or neutral.
Positive

Sentiment analysis is a common NLP problem that has probably been well understood by
LLMs through pre-training or fine-tuning. Thus we can prompt LLMs using simple instructions
to perform the task. However, for new classification problems, it may be necessary to provide
additional details about the task, such as the classification standards, so that the LLMs can perform
correctly. To do this, we can add a more detailed description of the task and/or demonstrate
classification examples in the prompts. To illustrate, consider the following example.

Prompting

108

Analyze the polarity of the following text and classify it as positive, negative, or
neutral. Here’s what each category represents:
Positive: This indicates that the text conveys a positive emotion or attitude. For
example, texts expressing happiness, satisfaction, excitement, or admiration are
considered positive.
Negative: This refers to a text that expresses a negative emotion or attitude. It
encompasses feelings of sadness, anger, frustration, or criticism.
Neutral: Neutral sentiment is used to describe texts that do not exhibit clear positive or negative emotions but instead convey informational, factual, or indifferent
tones.
Text:
The service at the restaurant was slower than expected, which was a bit frustrating.
What is the polarity of the text?
Positive

While it seems straightforward to use LLMs for classification problems, there are still issues
that have not been well addressed. For example, when dealing with a large number of categories,
it remains challenging to effectively prompt LLMs. Note that if we face a very difficult classification problem and have a certain amount of labeled data, fine-tuning LLMs or adopting “BERT +
classifier”-like architectures is also desirable.

3.1.4.2 Information Extraction
Many NLP problems can be regarded as information extraction problems, involving the identification or extraction of specific pieces of information from unstructured text. This information can
include named entities, relationships, events, and other relevant data points. The goal of information extraction is to transform raw data into a format that can be easily analyzed and used in
various downstream applications.
As information extraction covers a wide range of problems, we cannot discuss them all here.
Instead, we start with the task of named entity recognition — a task that has long been a concern
in NLP. Named entity recognition is a process that detects and classifies key information in text
into specific groups. These key pieces of information, known as named entities, typically include
proper names and are categorized into distinct classes such as people, locations, organizations,
dates, monetary values, and percentages. Consider first a simple example of extracting person
names from a text3 .

3

The text is from https://edition.cnn.com/travel

3.1 General Prompt Design

109

Identify all person names in the provided text.
Text:
Is the UK really doing that badly or have travel trends shifted, post-pandemic?
For Tom Jenkins, CEO of the European Tourism Organisation, it’s the latter. “I
think the UK is doing perfectly well but we’ll see more people going to Europe,”
he says of 2024, adding that there’s “allegedly a slight plateau of American demand for the UK.”
The person name identified in the provided text is: Tom Jenkins

Then, we can extract all named entities.

Identify and classify all named entities in the provided text into categories such
as person names, locations, dates, and organizations. List each entity with its type
on one line.
Text:
Is the UK really doing that badly or have travel trends shifted, post-pandemic?
For Tom Jenkins, CEO of the European Tourism Organisation, it’s the latter. “I
think the UK is doing perfectly well but we’ll see more people going to Europe,”
he says of 2024, adding that there’s “allegedly a slight plateau of American demand for the UK.”
1. Tom Jenkins - Person Name
2. European Tourism Organisation - Organization
3. UK - Location
4. Europe - Location
5. 2024 - Date

Given these identified named entities, we can further process the text using other information
extraction tasks. For example, we can identify the relationships between these named entities (call
it relation extraction). The corresponding prompt is shown as follows.

Prompting

110

Given a text and a list of named entities identified within it, analyze and describe
the relationships between these entities. Explain how each entity is contextually
related to others.
Text:
Is the UK really doing that badly or have travel trends shifted, post-pandemic?
For Tom Jenkins, CEO of the European Tourism Organisation, it’s the latter. “I
think the UK is doing perfectly well but we’ll see more people going to Europe,”
he says of 2024, adding that there’s “allegedly a slight plateau of American demand for the UK.”
Named Entity:
Tom Jenkins, European Tourism Organisation, UK, Europe, 2024
Tom Jenkins - European Tourism Organisation: Tom Jenkins is the CEO of the
European Tourism Organisation. This indicates a professional association where
Tom Jenkins likely has a significant role in directing the organization’s activities,
especially those related to tourism in Europe.
...
If LLMs have been fine-tuned with instruction following for information extraction tasks, it
is relatively easy to perform various information extraction tasks. For example, the following is a
prompt template for information extraction.
You will be provided with a text. Your task is to {∗task-description∗}
Text: {∗text∗}

Here task-description is the instruction for the task. Examples of the instruction include
extract the keywords from the text.
extract the key events from the text.
detect the coreference and anaphoric links between entities from the text.
and so on. Of course, we can provide a more detailed description of the task to prompt LLMs
more accurately.

3.1.4.3 Text Generation
In NLP, the term text generation is used broadly to cover a wide range of different problems, many
of which focus on generating text based on given text. These problems can be roughly categorized
into two classes: 1) continual writing based on the input text (call it text completion), and 2)
transformation of the input text into another text (call it text transformation). Here, we illustrate
the prompts for the former and consider the prompts for the latter in the following discussion.

3.1 General Prompt Design

111

A simple example of text completion is continuing to write a story starting with a few given
sentences.

You are a writer who is great at writing stories. You are provided with a few
sentences to start with. Please continue the story from the provided sentences.
Start of the story:
One sunny afternoon, Lily discovered a tiny, golden key hidden under a rock in
her backyard. Curious, she picked it up and wondered what it might unlock. Just
then, her cat, Whiskers, began to meow excitedly, leading her towards the old
oak tree at the edge of the garden.
Lily followed Whiskers, her heart pounding with excitement and a touch of
nervousness. The old oak tree had always seemed mysterious to her, its gnarled
branches casting long shadows even on the brightest days. As they reached the
tree, Whiskers darted ahead and started pawing at something buried beneath a
pile of autumn leaves ...

Likewise, we can create a conversation given the conversational context.

You are provided with a conversation between a tourist and a taxi driver. Please
complete the conversation.
Tourist: Could you take me to the downtown museum, please?
Taxi Driver: Of course! First time in the city?
Tourist: Yes, any must-see places around?
Taxi Driver: Definitely check out the botanical gardens after the museum.
They’re nearby and quite beautiful.
Tourist: Great, thanks for the tip!
Taxi Driver: No problem. Enjoy your visit!

Another application of text completion is generating text based on user requirements. The
requirements can include style, tone, length, and any specific content that the text should contain.
For example, users may need a formal report, a creative story, or a piece of programming code.
Here is a prompt for generating a five-character regulated Chinese poem.

Prompting

112

You are a poet. Please write a traditional Chinese poem. Here is the requirement.
Theme: Spring’s rejuvenation
Structure: Five-character regulated poem
Emotion: Convey a sense of happiness and renewal through vivid spring imagery.
春意盎然
翠柳风中舞，
红花雨后新。
山明水又绿，
天宽鸟自频。

If the LLM is trained to generate language and code, we can prompt it to perform code completion tasks. Here is an example.

Please write a Python function to calculate the average of a list of numbers.
def calculate_average(numbers):
if numbers: # Check if the list is not empty
return sum(numbers) / len(numbers)
else:
return 0 # Return 0 if the list is empty

3.1.4.4 Text Transformation
Typical text transformation tasks include machine translation, summarization, text style transfer,
and so on. The prompts used to perform these tasks are not significantly different from the ones
presented previously. Here, we present a few examples to showcase these prompts.

Translate the following text from English to Spanish.
Text:
The quick brown fox jumps over the lazy dog.
El rápido zorro marrón salta sobre el perro perezoso.

3.1 General Prompt Design

113

Summarize the following this article in no more than 50 words:
Article:
In recent years, urban areas have been facing significant challenges related to sustainability and environmental management. Cities are experiencing higher rates
of pollution, increased traffic congestion, and greater demands on infrastructure.
This growth has led to numerous environmental issues, including elevated levels of air and water pollution, increased waste production, and strained public
services ...
Urban areas are grappling with sustainability challenges, such as rising pollution,
traffic congestion, and infrastructure demands ...

Rewrite this text in a formal tone.
Text:
Hey, what’s up? Long time no see!
Hello, how have you been? It has been quite some time since we last met!

3.1.4.5 Question Answering
The question-answering format is inherently simple. For a given question, there is an answer that
corresponds to it. For example, in open-domain question answering, we expect the system to
return an answer in response to a user-submitted question. Prompt templates for general-purpose
question answering can be:

{∗question∗}

Question: {∗question∗}
Answer:

Question answering is important in NLP because many problems can be framed as questionanswering tasks. In particular, many recent reasoning tasks are defined in the form of question
answering. For example, in the MMLU benchmark [Hendrycks et al., 2021], each example consists of a multiple-choice question, and LLMs are required to select the correct answer. See the
following for an example prompt for answering a question in this dataset.

Prompting

114

(Z, ∗) is a group with a ∗ b = a + b + 1 for all a, b in Z. The inverse of a is
(A) 0

(B) −2

(C) a − 2

(D) (2 + a) ∗ −1

D

Another widely-used benchmark is the GSM8K dataset [Cobbe et al., 2021]. It consists of
thousands of grade school math word problems. For each problem, we use LLMs to return a
solution in natural language. For example, the following is a prompt used for solving a problem
in GSM8K.

DEMO

Q: Jerry’s two daughters play softball on different teams. They each have
8 games this season. Each team practices 4 hours for every game they
play. If each game lasts for 2 hours, how many hours will Jerry spend at
the field watching his daughters play and practice altogether?
A: Jerry will spend 8 games * 2 hours per game = ≪ 8 ∗ 2 = 16 ≫ 16
hours watching one daughter play her games. He will spend 16 ∗ 2 =≪
16 ∗ 2 = 32 ≫ 32 hours watching both daughters play their games. He
will spend 8 games * 4 hours of practice = ≪ 8 ∗ 4 = 32 ≫ 32 hours
watching one daughter practice. He will spend 32∗2 =≪ 32∗2 = 64 ≫
64 hours watching both daughters practice. He will spend a total of 32
hours watching games + 64 hours watching practice =≪ 32 + 64 =
96 ≫ 96 hours. #### 96
Q: Mary bought six apples from the store. From the apples she bought,
for each that Mary ate, she planted two trees from the remaining ones.
How many apples did Mary eat?
A: She planted eight trees. This means she used half of that amount of
apples, which is 8 trees /2 trees/apple =≪ 8/2 = 4 ≫ 4 apples. That
means that she planted four of the six apples she bought, leaving only 6
apples − 4 apples =≪ 6 − 4 = 2 ≫ 2 apples to be eaten. #### 2

USER

Q: Boris has 100 apples. Beck has 23 fewer apples than Boris. If Boris
gives Beck 10 apples, how many fewer apples does Beck have than Boris
now?
A: Boris starts with 100 apples and gives Beck 10 apples, leaving
him with 100 apples −10 apples =≪ 100 − 10 = 90 ≫ 90 apples.
Initially, Beck has 23 fewer apples than Boris, giving him a
starting count of 100 − 23 =≪ 100 − 23 = 77 ≫ 77 apples. After
receiving 10 apples from Boris, Beck’s total comes to 77 apples +10
apples =≪ 77 + 10 = 87 ≫ 87 apples. Consequently, Beck now has
90 − 87 =≪ 90 − 87 = 3 ≫ 3 fewer apples than Boris. #### 3

3.2 Advanced Prompting Methods

115

Here a few-shot prompt is adopted. The LLM learns from these demonstrations of problemsolution pairs not only the way of problem-solving but also the way of formatting the output. For
example, the final result of calculation follows the #### token, and ≪ ... ≫ annotates the detailed
calculation steps (called calculation annotation)4 .

3.2 Advanced Prompting Methods
So far in this chapter, we have introduced the basic concepts related to LLM prompting and presented a number of prompts for NLP tasks. We now consider several techniques for enhancing the
effectiveness of prompting.

3.2.1 Chain of Thought
We have encountered the concept of chain of thought (CoT) several times in this chapter and
previous ones [Wei et al., 2022c; Chowdhery et al., 2022]. CoT methods provide a simple way
to prompt LLMs to generate step-by-step reasoning for complex problems, thereby approaching
tasks in a more human-like manner. Rather than coming to a conclusion directly, the CoT methods
instruct LLMs to generate reasoning steps or to learn from demonstrations of detailed reasoning
processes provided in the prompts. To illustrate CoT, we consider the problem of algebraic calculation, as commonly described in the literature. Suppose we are given a algebraic problem
Calculate the average of the numbers 2, 4, and 6.
We can consider it as the question and prompt an LLM to answer it.
Q: Please calculate the average of the numbers 2, 4, and 9.
A: The answer is 6.
It seems difficult for the LLM to directly give a correct answer. A simple improvement is to
add demonstrations of similar problems in the prompt, and thus the LLM can learn from these
demonstrations.
Q: Please calculate the average of the numbers 1, 3, 5, and 7.
A: The answer is 4.
Q: Please calculate the average of the numbers 2, 4, and 9.
A: The answer is 7.
The problem here is that, although we have shown a similar question-answer pair, it remains
difficult for the LLM to reason out the correct answer. In CoT, not only can LLMs learn from the
4

During prediction, a calculator is used when we see ≪ ... ≫. More specifically, once the LLM encounters “=”
in a ≪ ... ≫, then the calculator calculates the expression on the left-hand side of “=”. This method helps reduce the
calculation errors made by LLMs.

Prompting

116

correspondence between questions and answers but they may gain more from detailed problemsolving steps that used to derive the answers. To do this, we can incorporate some reasoning steps
into the prompt to obtain a CoT prompt.
Q: Please calculate the mean square of the numbers 1, 3, 5, and 7.
A: Calculate the square of each number: 12 = 1, 32 = 9, 52 = 25, and 72 = 49.
Sum the squares, 1 + 9 + 25 + 49 = 84. There are 4 numbers in total. Divide
the sum by the number of items, 84/4 = 21. The answer is 21.
Q: Please calculate the average of the numbers 2, 4, and 9.
A: Calculate 2 + 4 + 9, which equals 15. There are three numbers. Divide the
total sum by the count, resulting in 15/3 = 5. The answer is 5.
Here we highlight the reasoning steps in green. By providing a detailed reasoning process for
a similar problem, the LLM learns to reason and thereby generates a problem-solving path that
leads to the correct answer.
There are several benefits of using CoT prompting. First, CoT allows LLMs to decompose complex problems into smaller, sequential reasoning steps. This somewhat mirrors human
problem-solving behaviors, making it particularly effective for tasks requiring detailed, multi-step
reasoning. Second, CoT makes the reasoning process more transparent and interpretable. Since all
reasoning steps are visible, we can understand and interpret how a conclusion was reached. Third,
if users can follow the logic behind the reasoning process, they will be more likely to trust the predictions of an LLM. This is particularly important when applying LLMs in fields like medicine,
education, and finance. Fourth, CoT is an in-context learning approach, and thus, it is applicable
to most well-trained, off-the-shelf LLMs. Moreover, CoT provides efficient ways to adapt LLMs
to different types of problems. It can even inspire more creative solutions by exploring various
alternative reasoning paths, which might not be obvious when arriving at a conclusion directly.
The method described above requires providing one or more examples of CoT reasoning, typically called the few-shot CoT method. By contrast, the zero-shot CoT method does not require
such examples. It instead prompts LLMs to reason step-by-step by incorporating specific instructions in prompts. For example, below is a zero-shot CoT prompt.
Q: Please calculate the average of the numbers 2, 4, and 9.
A: Let’s think step-by-step.
We have three numbers: 2, 4, and 9. Add these numbers together,
2 + 4 + 9 = 15. Determine how many numbers there are, which in this case
is three. The average is calculated by dividing the total sum by the number of
elements. Completing the division gives 15/3 = 5. So the answer is 5.
Following the instruction “Let’s think step by step”, the LLM is prompted to generate detailed
reasoning steps. As discussed in Kojima et al. [2022]’s work, prompting with such instructions
may result in LLMs generating only the reasoning steps without a clear conclusion. In this case,
a second round of prompting can be used to extract the answer from these reasoning steps. For
example, Kojima et al. [2022] create a second prompt which combines both the input and output

3.2 Advanced Prompting Methods

117

in the first round of prompting. Using this combined input, the LLM can continue its reasoning
process and then generate the correct answer. Furthermore, it is possible to prompt LLMs to
reason using instructions other than “Let’s think step by step”, such as “Let’s think logically” and
“Please show me your thinking steps first”.
While we have illustrated CoT methods using an algebraic reasoning problem, these methods
can be applied to a variety of different problems. Typical problem-solving scenarios for CoT
include mathematical reasoning, logical reasoning, commonsense reasoning, symbolic reasoning,
code generation, and so on. See Figure 3.1 for more examples of applying CoT in various tasks.
CoT today is one of the most active fields of prompt engineering. This has not only led to improved performance for LLM prompting but has opened the door to a wide range of methods for
studying and verifying reasoning capabilities of LLMs. Although we have focused on the basic
idea of CoT in this section, it can be improved in several ways. For example, we can consider the
reasoning process as a problem of searching through many possible paths, each of which may consist of multiple intermediate states (i.e., reasoning steps). In general, we wish the search space to
be well-defined and sufficiently large, so that we are more likely to find the optimal result. For this
reason, an area of current LLM research is aimed at designing better structures for representing
reasoning processes, allowing LLMs to tackle more complex reasoning challenges. These structures include tree-based structures [Yao et al., 2024], graph-based structures [Besta et al., 2024],
and so on. By using these compact representations of reasoning paths, LLMs can explore a wider
range of decision-making paths, analogous to System 2 thinking5 . Another line of research focuses on prompting LLMs with multi-round interactions. This involves decomposing complex
problems into sub-problems, verifying and refining model outputs, employing model ensembling,
and so on. Note that these methods and the issues involved are not limited to CoT. In fact, they
are often used as more general approaches to improving LLMs, while CoT can be seen as a way
to test the capabilities of LLMs. We will see discussions of some of these issues in the following
subsections.
Before leaving our discussion of CoT, we should consider its practical limitations. One of them
is the need for detailed, multi-step reasoning demonstrations in few-shot CoT scenarios, which
may be difficult to obtain, either automatically or manually. Also, there is no standard method for
breaking down complex problems into simpler problem-solving steps. This often heavily depends
on the user’s experience. In addition, errors in intermediate steps can also affect the accuracy of
the final conclusion. For further discussion on the pros and cons of CoT, the interested reader can
refer to recent surveys on this topic [Chu et al., 2023; Yu et al., 2023; Zhang et al., 2023a].

3.2.2 Problem Decomposition
We have seen that LLMs can benefit from solving a complex problem by breaking it down into
simpler problem-solving tasks. Such an approach can be seen as an example of a broader paradigm
known as problem decomposition, which has been extensively explored and discussed in psychology and computer science. From the psychological perspective, complex problem-solving
refers to a process of addressing a problem using knowledge that helps overcome the barriers of
5

System 1 and System 2 thinking, as described by Kahneman [2011], represent two different modes of cognitive
processing. System 1 is fast, automatic, intuitive, and emotional. This mode of thinking operates effortlessly and
quickly, and is often what guides our daily decisions, judgments, and impressions. System 2 is slow, deliberate, and
analytical. It is activated when we need to perform complex computations.

Prompting

118

Q: What home entertainment equipment requires cable?
Answer Choices: (a) radio shack (b) substation (c) television (d) cabinet
A: The answer must require cable. Of the
above choices, only television requires cable.
So the answer is (c).
Q: What type of person typically contracts
illness?
Answer Choices: (a) hospital (b) head (c)
sick person (d) elderly person (e) doctor’s office
A: The answer must be a person susceptible to illness. Of the choices, "elderly per
son" is correct because they typically have
weaker immune systems.
So the answer is (d).

Q: Do hamsters provide food for any animals?
A: Hamsters are prey animals. Prey are food
for predators. Thus, hamsters provide food
for some animals. So the answer is yes.
Q: Can you use Microsoft Office without internet?
A: Yes, you can use Microsoft Office applications , although features like cloud storage and online collaboration require internet. So the answer is yes.

(a) CSQA

(b) StrategyQA

Q: Complete the rest of the sequence, making sure that the parentheses are closed properly. Input: [ {
A: Let’s think step by step.
0: empty stack
1: [ ; stack: [
2: { ; stack: [ {
So the answer is } ].

Q: Take the last letters of the words in "Elon
Musk" and concatenate them.
A: The last letter of "Elon" is "n". The last
letter of "Musk" is "k". Concatenating them
is "nk". The answer is nk.

Q: Complete the rest of the sequence, making sure that the parentheses are closed properly. Input: < [ [
A: Let’s think step by step.
0: empty stack
1: < ; stack: <
2: [ ; stack: < [
3: [ ; stack: < [ [
So the answer is ] ] >.

(c) Dyck languages

Q: Take the last letters of each words in "Renata Mariela Mona Kristin" and concatenate
them.
A: The last letter of "Renata" is "a". The
last letter of "Mariela" is "a". The last letter of "Mona" is "a"". The last letter of
"Kristin" is "n". Concatenating them is
"aaan". The answer is aaan.

(d) Last Letter Concatenation

Fig. 3.1: CoT in four different reasoning tasks, including CSQA, StrategyQA, Dyck languages, and Last Letter Concatenation. The CoT parts are highlighted in green.

the problem6 . There are generally no standard or clear paths to a solution for a complex problem. However, it is often advantageous to employ strategies that decompose the problem, thereby
making it easier to tackle the corresponding sub-problems with less effort. For example, consider
writing a blog about the risks of AI. If we simply prompt an LLM with the instruction “Please
write a blog about the risks of AI”, the LLM may generate a blog with arbitrary structures and
6

A relatively formal definition can be found in Frensch and Funke [2014]’s book: complex problem-solving occurs
to overcome barriers between a given state and a desired goal state by means of behavioral and/or cognitive, multi-step
activities.

3.2 Advanced Prompting Methods

119

writing styles. A better method, instead, could be to outline the blog and provide more detailed
information about each section. Consider the following prompt
You are a blog writer. Please follow the provided outline below to write a blog
about the risks of AI.
• Introduction
Introduce AI, its relevance, and the importance of understanding its risks for youth.
• Privacy Concerns
Discuss how AI might compromise personal privacy through interactions online.
• Misinformation
Explore AI’s role in spreading misinformation and influencing young people’s decisions.
• Cyberbullying
Highlight how AI tools can be utilized in cyberbullying and the impact on mental
health.
• Tips for Safe AI Use
Offer guidelines for responsible AI usage and promote critical thinking.
• Conclusion
Recap main points and encourage proactive engagement with AI ethics.

Here we give the title and major points for each section. Then, the LLM can use this structure to
break down the writing task by filling in content for these sections. Note that the way to structure
the blog can be provided by humans or even generated automatically. For example, we can use
the LLM to first generate the outline, and then ask it to follow this outline to complete the writing.
In computer science, decomposing complex problems is a commonly used strategy in software
and hardware system design. A well-known example is the divide-and-conquer paradigm, which
is often used to design algorithms for computation problems that can be reduced to simpler, more
manageable problems. For example, consider a problem of determining whether a document
discusses the risks of AI. We can instruct the LLM with the following prompt.
You are provided with a text. Please determine whether it discusses the risks of
AI.
{∗document∗}

If the document is long, the computation will be expensive. Alternatively, we can divide
the document into relatively short segments and perform the same task on each segment. These
segments can be processed in parallel to further reduce the computational cost. Next, we determine

Prompting

120

the relevancy of each segment to the topic of AI risks. The final output is then generated using
another prompt.
Your task is to determine whether a text discusses the risks of AI. This text has
been divided into segments, and you have obtained the relevancy of each segment
to the topic of AI risks. Based on this, please provide your final result.
Segment 1: {∗relevancy-to-the-topic1∗}

Segment 2: {∗relevancy-to-the-topic2∗}

Segment 3: {∗relevancy-to-the-topic3∗}
...

Now let us return to a more general discussion of problem decomposition in prompting. While
problem decomposition can be applied to various NLP problems, it has been more extensively
discussed and tested in reasoning tasks recently. For complex reasoning tasks, we often need
a multi-step reasoning path to reach a correct conclusion. We can use LLMs to achieve this in
three different ways. First, LLMs can directly reach the conclusion. In other words, they can
predict without explicit reasoning processes, and there is a hidden and uninterpretable reasoning
mechanism. Second, LLMs are prompted to generate a multi-step reasoning path that leads to the
conclusion, like CoT. However, we run LLMs just once, and all intermediate steps in reasoning
are generated in a single prediction. Third, we break down the original problem into a number of
sub-problems, which are either addressed in separate runs of LLMs or tackled using other systems.
Here we focus our attention on the third approach, which is closely related to problem decomposition. Note, however, that a more comprehensive discussion could cover all these approaches,
while the first two have been discussed to some extent in this chapter.
A general framework for problem decomposition involves two elements.
• Sub-problem Generation. This involves decomposing the input problem into a number of
sub-problems.
• Sub-problem Solving. This involves solving each sub-problem and deriving intermediate
and final conclusions through reasoning.
These two issues can be modeled in different ways, leading to various problem decomposition
methods. One approach is to treat them as separate steps in a two-step process. For example,
consider the blog writing task described at the beginning of this subsection. In the first step, we
decompose the entire problem into sub-problems all at once (i.e., outline the blog). In the second
step, we solve the sub-problems either sequentially or in another order (i.e., fill in content for
each section as needed). The final output of this process combines the results from solving each
sub-problem. While this method is simple and straightforward, it assumes that the problem is
compositional, making it more suitable for tasks like writing and code generation.
However, many real-world problems require complex reasoning. One key characteristic of
these problems is that the reasoning steps may not be fixed. The reasoning path can vary for
different problems, and each step of reasoning may depend on the outcomes of prior steps. In

3.2 Advanced Prompting Methods

121

such cases, it is undesirable to use fixed sub-problem generation in advance. Instead, sub-problems
should be generated dynamically based on the input problem, and, if possible, generated on the
fly during the reasoning process. This makes problem decomposition more challenging compared
with designing divide-and-conquer algorithms. Ideally, we would like to jointly design both the
systems for sub-problem generation and sub-problem solving. But a more practical and widely
used approach is to adopt separate models for these tasks. A straightforward way to achieve this
is to adapt an LLM for these tasks by either prompting or tuning the model.
Here we consider a method based on the above idea, called least-to-most prompting [Zhou et al.,
2023b]. The motivation for this method arises from the challenges of solving difficult reasoning
problems — those that cannot be addressed by simply generalizing from a few examples. For
these problems, a more effective problem-solving strategy is to follow a progressive sequence of
sub-problems that systematically lead to the conclusion. More specifically, in the least-to-most
prompting method, sub-problem generation is performed by prompting an LLM with instructions
and/or demonstrations. For example, below is a 2-shot prompt for sub-problem generation in
least-to-most prompting.

TASK

Your task is to decompose a problem into several sub-problems. You will
be given a few examples to illustrate how to achieve this.

DEMO

Q: In a community, 5% of the population are infants, 15% are children,
40% are adults, and 40% are seniors. Which group makes up the largest
portion of the population?
A: To answer the question “Which group makes up the largest portion of the
population?”, we need to know: “How many percent are infants?”, “How
many percent are children?”, “How many percent are adults?”, “How many
percent are seniors?”.
Q: Alice, Bob, and Charlie brought beads for their group project in their
craft class. Alice has twice as many beads as Bob, and Bob has five times
as many beads as Charlie. If Charlie has 6 beads, how many beads can they
use for their craft project?
A: To answer the question “How many beads can they use for their craft
project?”, we need to know: “How many beads does Bob have?”, “How
many beads does Alice have?”.

USER

Q: The environmental study conducted from 2015 to 2020 revealed that the
average temperature in the region increased by 2.3 degrees Celsius. What
was the duration of the environmental study?
A: To answer the question “What was the duration of the environmental
study?”, we need to know: “When did the environmental study start?”,
“When did the environmental study end?”.

By learning from the examples, the LLM can generate two sub-problems for answering the
new problem “What was the duration of the environmental study?” (highlighted in blue and
orange). Given these sub-problems, we solve them sequentially. For each sub-problem, we take
all previously-generated QA pairs as context, and then produce the answer. For the example above,

Prompting

122
we need to answer the first sub-problem by prompting the LLM, like this

The environmental study conducted from 2015 to 2020 revealed that
the average temperature in the region increased by 2.3 degrees Celsius.
SUB-PROB1

Q: When did the environmental study start?
A: The environmental study started in 2015.

Once we have the answer to the first sub-problem, we proceed to the second one. This time,
we include both the first sub-problem and its corresponding answer in the input.
The environmental study conducted from 2015 to 2020 revealed that
the average temperature in the region increased by 2.3 degrees Celsius.
SUB-PROB1

Q: When did the environmental study start?
A: The environmental study started in 2015.

SUB-PROB2

Q: When did the environmental study end?
A: The environmental study started in 2020.

Finally, we use the LLM to solve the original problem given the answers to all the subproblems.
The environmental study conducted from 2015 to 2020 revealed that
the average temperature in the region increased by 2.3 degrees Celsius.
SUB-PROB1

Q: When did the environmental study start?
A: The environmental study started in 2015.

SUB-PROB2

Q: When did the environmental study end?
A: The environmental study started in 2020.

FINAL

Q: What was the duration of the environmental study?
A: The duration of the environmental study was 5 years.

The least-to-most method offers a basic approach to prompting LLMs to generate and solve
sub-problems separately. We can improve it in several ways. One simple improvement is to apply
various advanced prompting techniques, which do not require changes to the problem decomposition framework. For example, we can incorporate CoT into the prompting to enhance the
reasoning performance of sub-problem generation and solving.
Another improvement is to explore methods for better decomposing problems and organizing
problem-solving paths. To describe these approaches, we will use the symbol p0 to denote the

3.2 Advanced Prompting Methods

123

input problem, and use the symbols {p1 , ..., pn } to denote the sub-problems corresponding to p0 .
For least-to-most prompting, we decompose p0 into {p1 , ..., pn }, given by
{p1 , ..., pn } = G(p0 )

(3.2)

where G(·) denotes the function of sub-problem generation. Then, we solve the sub-problems
{p1 , ..., pn } sequentially, resulting in a sequence of answers {a1 , ..., an }. For answering the i-th
sub-problem pi , we include both the original problem p0 and all previously-seen problem-answer
pairs in the context for prediction. The answer ai is given by
ai = Si (pi , {p0 , p<i , a<i })

(3.3)

where p<i = {p1 , ..., pi−1 } and a<i = {a1 , ..., ai−1 }. Si (·) denotes the function that solves the
sub-problem pi given the context {p0 , p<i , a<i }. The last step is to generate the answer to the
original problem p0 , which can be expressed in a similar manner to Eq. (3.3).
a0 = S0 (p0 , {p≤n , a≤n })

(3.4)

One way to refine this model is to modify the G(·) function so that the model can dynamically
generate answers. Instead of generating all sub-problems at one time, we can generate each of
them during problem-solving [Dua et al., 2022]. To do this, we can replace Eq. (3.2) with
pi = Gi (p0 , {p<i , a<i })

(3.5)

Hence we obtain a sub-problem generation model that operates in a step-by-step manner. At each
step i, we first generate the sub-problem pi by prompting an LLM with the original problem p0
and the problem-solving history {p<i , a<i }. We then generate the answer ai for this sub-problem
using the same or a different LLM, based on the same contextual information (see Eq. (3.3)). This
method effectively expands the reasoning capacity of LLMs by allowing them to dynamically
generate and solve sub-problems in intermediate reasoning steps. As a result, the reasoning paths
are not fixed in advance, and the models can choose and adapt their reasoning strategies during
problem-solving.
Another way to improve the above model is to focus on developing better sub-problem solvers.
In our previous discussion, we restricted Si (·) to LLMs that are prompted to solve the sub-problem
pi . In fact, we can expand this function to any system that is capable of addressing the subproblem. For example, Si (·) could make calls to IR systems, thereby allowing us to access a
broader range of data for problem-solving. Another example is using Si (·) as a calculator to
accurately compute results in mathematical problem-solving. If the sub-problem pi is complex
and requires multiple intermediate problem-solving steps, it is also possible to further decompose
pi into smaller sub-problems. For example, Si (·) can be defined as a recursive program that
generates and solves sub-problems. This incorporates recursion into problem-solving and allows
us to address problems by iteratively decomposing them. As a result, we can define a hierarchical
structure for problem-solving [Khot et al., 2023].
If we generalize the above formulation a bit further, we can consider it as a reinforcement
learning problem. A typical method is to model a problem-solving process as a decision making
process. In each step of this process, an action is taken based on the current state. These actions

124

Prompting

can include all functions for sub-problem generation and solving (i.e., Gi (·) and Si (·)). Thus,
the action sequence corresponds to a problem-solving path. Since the discussion of reinforcement
learning problems is beyond the scope of this chapter, we skip the precise description of this
learning task. Nevertheless, developing an agent or controller to determine when and how to
generate and solve a sub-problem is also a natural choice.
In NLP, problem decomposition is related to a long line of research on multi-hop question
answering [Mavi et al., 2024]. This task requires the system to gather and combine information
from multiple pieces of text to provide an accurate answer to a complex question. For example,
to answer the question “What is the capital of the country where Albert Einstein was born?”, we
need to know “Where Albert Einstein was born?” and “What’s the capital of Germany?”. Earlier
work in this area and related ones has investigated the issue of problem decomposition, though the
methods might not be based on LLMs. For example, a popular method is to develop an additional
neural model to generate simpler questions that address different aspects of the original question
[Andreas et al., 2016; Talmor and Berant, 2018; Min et al., 2019]. This question generator can
create questions in a batch or sequential manner.
Broadly speaking, problem decomposition is also related to the compositionality issue in NLP
[Drozdov et al., 2022; Press et al., 2023]. For example, in semantic parsing, we map natural language sentences into structured meaning representations by breaking them down into constituent
parts and understanding the sentences based on the meanings of these parts and the rules used to
combine them. In early studies of this field, highly compositional sentences were considered easier
for testing systems, as it is relatively straightforward to decompose such sentences and compose
the meanings of their parts. However, the task becomes much more difficult when more generalization is required for modeling compositionality in new data. In this case, we want systems
to have improved abilities of compositional generalization. In more recent research on LLMs,
this issue has been frequently discussed in compositional reasoning tasks, such as SCAN7 , as it
is considered an important aspect of testing the language understanding and reasoning abilities
of LLMs. This also presents new tasks for developing and examining problem decomposition
methods.
In LLMs, one interesting application of problem decomposition is tool use. In some cases,
it is necessary to integrate external tools into LLMs to access accurate data not available during
training or fine-tuning. For example, LLMs can integrate with APIs to fetch real-time data such
as weather updates, stock market prices, or news feeds, enabling them to provide up-to-date responses to user queries. When using tools, LLM predictions might include markers that indicate
where and how to call external APIs. This requires decomposing the problem into sub-problems,
with some handled by the LLMs and others by external tools. More detailed discussions on this
topic will be presented in Section 3.2.5.

3.2.3 Self-refinement
In many cases, predictions of LLMs can be inaccurate or incorrect. Given that current LLMs can
perform tasks like refinement and correction, it makes sense to explore methods for these models
to self-refine their outputs. Self-refinement is a common phenomenon in human psychological
7

The SCAN tasks (Simplified versions of the CommAI Navigation tasks) are designed to evaluate the ability of
LLMs to perform compositional generalization [Lake and Baroni, 2018]. They involve translating natural language
commands into a sequence of actions. For example, a command “jump opposite left and walk thrice” can be translated
into the action sequence “LTURN LTURN JUMP WALK WALK WALK”.

3.2 Advanced Prompting Methods

125

activities and daily behavior. For example, when designing a product, a designer might first create
a basic prototype, then refine the design after evaluation and testing to enhance user experience
and functionality. The refinement can be iterated several times until the design is satisfactory. The
idea of predict-then-refine can also be found in NLP. One early example is Brill’s tagger [Brill,
1992], where an initial POS tagging result of a sentence can be iteratively refined using a rulebased system. In the era of deep learning, a good deal of work on sequence-to-sequence problems,
such as grammar correction and text rewriting, can also be seen as examples on this theme.
We can prompt LLMs to do self-refinement. Consider a simple example where we first use
an LLM to translate a text from Chinese to English and then use the same model to refine the
translation. Here is how we can prompt the model to translate:
Translate the following text from Chinese to English:
Chinese:
一系列考古发现奠定红山文化在中华文明起源研究中的重要地位。最新公
布的研究成果认为，大约从距今5800年开始，中华大地上各个区域相继出
现较为明显的社会分化，进入文明起源的加速阶段。
English:
A series of discoveries have cemented the significant role of the Hongshan
culture in studies on the origins of Chinese civilization. The latest research
findings suggests that, starting from around 5800 years ago, many regions
across China began to experience noticeable social differentiations, entering an
accelerated phase in the origin of civilization.
Then, we can prompt it to refine the translation:

NiuTrans

Prompting

126

You are provided with a Chinese text along with its English translation. Please
refine the translation to improve its accuracy and fluency if necessary.
Chinese:
一系列考古发现奠定红山文化在中华文明起源研究中的重要地位。最新公
布的研究成果认为，大约从距今5800年开始，中华大地上各个区域相继出
现较为明显的社会分化，进入文明起源的加速阶段。
English:
A series of discoveries have cemented the significant role of the Hongshan culture in studies on the origins of Chinese civilization. The latest research findings
suggests that, starting from around 5800 years ago, many regions across China
began to experience noticeable social differentiations, entering an accelerated
phase in the origin of civilization.
Refinement:
A series of archaeological discoveries have cemented the significant role
of the Hongshan culture in studies on the origins of Chinese civilization.
The latest research findings suggest that, starting from around 5800 years
ago, various regions across China began to experience noticeable social
differentiations, entering an accelerated phase in the origin of civilization.
Here modifications to the input translation are highlighted in green. This process can be iterative:
at each step, we use the translation from the previous step as input, and prompt the LLM to produce
a potentially better translation.
In the above example, the refinement of the initial model output was achieved by prompting the
LLM with a simple instruction like “Please refine it!”. However, the refinement is based solely on
the instruction-following ability of the LLM, and there is no guidance or supervision on how and
where to improve the model output. A more effective approach would be to consider feedback on
specific aspects that need refinement. For example, we can prompt the LLM with “Please correct
all the grammatical errors in the translation”, so that the model can focus more on grammatical
error correction during refinement.
A general framework of self-refinement with LLMs involves three steps [Madaan et al., 2024].
• Prediction. We use an LLM to produce the initial model output.
• Feedback Collection. We obtain feedback on the model output.
• Refinement. We use the LLM to refine the model output based on the feedback.
The last two steps can be repeated multiple times, which leads to an iterative self-refinement
process. In this framework, receiving accurate and detailed feedback is critical as it guides the subsequent refinement step. In general, feedback can be obtained in various ways [Fernandes et al.,
2023; Pan et al., 2024]. We can manually examine the model’s predictions to identify issues. Also,
we can use a feedback model to automatically generate feedback. For instance, a reward model
trained on some labeled data can help evaluate the model output and return a score indicating its
quality.

3.2 Advanced Prompting Methods

127

Since the focus of this chapter is on LLM prompting, we consider use LLMs to provide feedback. To illustrate, let us consider an example of generating good responses to user questions:

• First, we prompt an LLM to answer the input question, as usual.

Generate a response to the question: “What are some major environmental
concerns today?”
Response:
Major environmental concerns today include climate change, air and water
pollution, deforestation, loss of biodiversity, and ocean degradation.

• Then, using the following prompt, we prompt the same LLM again to get feedback on the
generated response.

You will be provided with a question and its corresponding response. Evaluate the accuracy and depth of this response, and identify any factual inaccuracies or areas where additional details could enhance understanding.
Question:
What are some major environmental concerns today?
Response:
Major environmental concerns today include climate change, air and water
pollution, deforestation, loss of biodiversity, and ocean degradation.
Feedback:
The response provided is accurate and highlights key environmental
concerns. However, it would benefit from more specific and detailed
descriptions of these concerns. Here are suggestions for improvement:
1. Include specific examples of environmental issues.
2. Add information about major pollutants.
3. Specify the regions most affected.

• Given the feedback the LLM produces, we prompt it to refine the response.

Prompting

128

You will be provided with a question, its corresponding response, and
some feedback on this response.
Question:
What are some major environmental concerns today?
Response:
Major environmental concerns today include climate change, air and water
pollution, deforestation, loss of biodiversity, and ocean degradation.
Feedback:
The response provided is accurate and highlights key environmental concerns. However, it would benefit from more specific and detailed descriptions of these concerns. Here are suggestions for improvement:
1. Include specific examples of environmental issues.
2. Add information about major pollutants.
3. Specify the regions most affected.
Now refine the response based on the provided feedback.
Refined Response:
Key environmental concerns include climate change, evidenced by rising
global temperatures and severe weather patterns; significant air and
water pollution in urban and industrial areas; rampant deforestation in
regions like the Amazon, causing biodiversity loss; and ocean degradation,
highlighted by coral reef bleaching and widespread overfishing.

Ideally, if a strong LLM is adopted, we would like to have it perform all three steps without
extra training. On the other hand, if we have enough labeled data for the task of interest, we can
enhance the performance of the LLM using supervised learning. For example, we can fine-tune
the LLM to better adapt it to refinement tasks, or alternatively, use task-specific models, which
may not necessarily be based on LLMs [Welleck et al., 2023; Schick et al., 2023]. In a broader
sense, improving LLMs for self-refinement tasks can be seen as an alignment issue. For example,
it has been found that some self-correction abilities can be activated through RLHF [Ganguli et al.,
2023]. However, discussing these issues is beyond the scope of this chapter. Further discussion
can be found in Chapter 4.
In LLMs, self-refinement is related to several concepts that reveal the psychological aspects
of these models, such as the ability to self-reflect. A view is that if LLMs are capable of selfreflection, their predictions can become more accurate and even possess self-correcting capabilities. This self-reflection can be activated in various ways, for example, by prompting these LLMs
to engage in more in-depth and careful thinking, or by providing examples from which the models
can learn and reflect. To illustrate, we consider here the deliberate-then-generate (DTG) method
presented in Li et al. [2023a]’s work, where LLMs are prompted to deliberate. In DTG, we are
given an initial model output which may contain errors. LLMs are then prompted to identify the
error types of this model output and provide an improved output. Below is a template of DTG
prompting for Chinse-to-English translation tasks.

3.2 Advanced Prompting Methods

129

Given the Chinese sentence: {∗source∗}
The English translation is: {∗target∗}

Please first detect the type of error, and then refine the translation.

Error Type:

We aim to first predict the error type (red), and then produce a refined translation (blue). This
process of deliberation is guided by the instruction “Please first detect the type of error, and then
refine the translation”. It encourages LLMs to initially engage in thoughtful analysis and then give
better results. Since error type prediction and refinement are performed in a single run of LLMs,
this method incorporates both steps of feedback and refinement into one process.
In the above prompts, we assume that the LLM we use is able to review the input translation
and correctly identify its error types. However, this raises new difficulties as the model may not
be good at finding errors in translations. This will in turn result in extra fine-tuning or prompting engineering efforts. So a simpler method is to reduce the burden of error identification and
use LLMs for deliberation only. To do this, we can replace the input translation with a random
translation and assign a default error type. An example of such a prompt is shown below.

Given the Chinese sentence:
一系列考古发现奠定红山文化在中华文明起源研究中的重要地位。
The English translation is:
A variety of innovative techniques have redefined the importance of modern art
in contemporary cultural studies.
Please first detect the type of error, and then refine the translation.
Error Type: Incorrect Translation

In this example, the input translation is not generated by LLMs but is instead randomly sampled from the dataset. So it is simply an incorrect translation for the source sentence, and we can
set the error type accordingly. The LLMs then generate a new translation by taking both the source
sentence and the incorrect translation as input. The design of this prompt can also be considered as
activating the learning capabilities of LLMs through “negative evidence” [Marcus, 1993], thereby
enabling them to reflect and produce better outcomes through contrastive analysis. Nevertheless,
this method does not rely on any feedback and can enhance the performance of a single LLM
prediction via simple prompting.
Note that while DTG is non-iterative, iterative learning and refinement are commonly used
in NLP. An advantage of these iterative approaches are that they mimics human learning and
problem-solving, where continuous feedback and adjustments lead to progressively improved outcomes. Iterative methods can be applied to a range of LLM prompting problems. For example, in
problem decomposition, one can incorporate new sub-problems and their solutions into the context at each step, and thus LLMs can progressively approach the solution of the original problem.

Prompting

130

On the other hand, iterative methods raise several issues that are absent in non-iterative methods, for example, errors in earlier steps may negatively impact subsequent problem-solving, and
determining when to stop iterating often requires additional engineering effort.

3.2.4 Ensembling
Model ensembling for text generation has been extensively discussed in the NLP literature. The
idea is to combine the predictions of two or more models to generate a better prediction. This
technique can be directly applicable to LLMs. For example, we can collect a set of LLMs and run
each of them on the same input. The final output is a combined prediction from these models.
For LLM prompting, it is also possible to improve performance by combining predictions
based on different prompts. Suppose we have an LLM and a collection of prompts that address
the same task. We can run this LLM with each of the prompts and then combine the predictions.
For example, below are three different prompt templates for text simplification.
Make this text simpler.
{∗text∗}

Condense and simplify this text.
{∗text∗}

Rewrite for easy reading.
{∗text∗}

Each of these prompts will lead to a different prediction, and we can consider all three predictions
to generate the final one.
Formally, let {x1 , ..., xK } be K prompts for performing the same task. Given an LLM Pr(·|·),
we can find the best prediction for each xi using ŷi = arg maxyi Pr(yi |xi ). These predictions
can be combined to form a “new” prediction:
ŷ = Combine(ŷ1 , ..., ŷK )

(3.6)

Here Combine(·) is the combination model, which can be designed in several different ways. For
example, we can select the best prediction by voting or by identifying the one that overlaps the
most with others. Another method for model combination is to perform model averaging during
token prediction. Let ŷj be the predicted token at the j-th step for model combination. The

3.2 Advanced Prompting Methods

131

probability of predicting ŷj is given by
ŷj = arg max
yj

K
X

k=1

log Pr(yj |xk , ŷ1 , ..., ŷj−1 )

(3.7)

In ensembling for LLM prompting, it is generally advantageous to use diverse prompts so that
the combination can capture a broader range of potential responses. This practice is common in
ensemble learning, as diversity helps average out biases and errors that may be specific to any
single model or configuration. From the Bayesian viewpoint, we can treat the prompt x as a latent
variable, given the problem of interest, p. This allows the predictive distribution of y given p to
be written as the distribution Pr(y|x) marginalized over all possible prompts
Z
Pr(y|p) =
Pr(y|x) Pr(x|p)dx
(3.8)
The integral computes the total probability of y by considering all possible values of x, weighted
by their likelihoods given p. Here Pr(y|x) is given by the LLM, and Pr(x|p) is the prior distribution of prompts for the problem. This is a good model because the integral effectively accounts
for the uncertainty in the choice of x, ensuring that the final predictive distribution Pr(y|p) is
robust and encompasses all potential variations and biases in the prompts. However, computing
this integral directly can be computationally infeasible due to the potentially infinite space of x.
One approach to addressing this issue is to employ methods like Monte Carlo sampling, which
approximate the integral using a manageable, finite number of prompts.
While the Bayesian treatment is mathematically well-defined, it is common practice in NLP
to assume a non-informative or uniform prior and focus instead on constructing a set of diverse
prompts. Consequently, the output can be computed using a straightforward combination model,
as described in Eq. (3.6). The issue of creating high-quality, diverse prompts has been studied in
CoT and other in-context learning areas. Most of the research focuses on incorporating a variety
of demonstration examples across different prompts. Here, we list some of these methods.
• Given a problem, we manually create a number of demonstrations and use different ones
for different prompts.
• Given a problem, we use LLMs to automatically generate demonstrations and prompts.
• Given a prompt, we create different prompts by changing the order of demonstrations in the
prompt.
• Given a prompt, we use LLMs to generate a number of similar prompts.
• Given a prompt, we transform it into other forms, e.g., translating it into other languages.
Of course, in practice, we can combine these methods to achieve greater diversity. An underlying assumption here is that diverse prompts can lead to diverse model outputs. This is particularly
the case when the problem we deal with is relatively new and difficult. For stronger and more robust LLMs, the variance in output for similar prompts might not be large. In this case, the benefit
of involving multiple prompts can be modest.

Prompting

132

In addition to providing diverse prompts for LLMs, another approach is to make use of the
inherent variance in the outputs of LLMs. One simple way to generate multiple outputs is to
sample outputs from the hypothesis space. This is straightforward for LLMs because their search
algorithms are naturally suited to producing multiple results. For example, if beam search is
used, we can easily collect all the complete hypotheses in the beam to generate multiple output
candidates. Also, we can modify the search algorithms to implement more powerful sampling
methods and explore a wider range of high-quality hypotheses.
The next question is how to combine diverse model predictions. This leads us back to the
model combination issue as described in Eq. (3.6). Here we consider the self-consistency method,
which outputs not the prediction with the highest probability, but rather the one that best aligns
with other predictions [Wang et al., 2022a; 2023b]. First, an LLM is prompted with CoT as
usual and generates multiple reasoning paths by sampling. Then, the frequency of each answer is
counted across these reasoning paths. The final output is the answer that has the highest count.
Below shows three different predictions by prompting an LLM with CoT.
Prediction 1 (Correct):

Q: Three friends play a game where they flip a fair coin. Each friend flips the
coin once. What is the probability that exactly one of them flips heads?
A: Let’s think step by step.
Each coin flip is an independent event with a 50% chance of landing heads
(H) and 50% tails (T). Consider the possible outcomes for three flips: HHH,
HHT, HTH, HTT, THH, THT, TTH, TTT. Out of these, only HHH, HHT, HTH,
and THH are relevant. We are only interested in scenarios with exactly one
head: HTT, THT, and TTH. There are 3 such outcomes out of 8 total, giving a
probability of 3/8.

Prediction 2 (Correct):

Q: Three friends play a game where they flip a fair coin. Each friend flips the
coin once. What is the probability that exactly one of them flips heads?
A: Let’s think step by step.
Using a binomial probability formula, where p (success or heads) is 0.5, and
each of the 3 trials (flips) is independent, we look for exactly one success
(head). The formula for exactly k successes in n trials is given by the


binomial coefficient formula nk pk (1 − p)n−k . Plugging in the values, we get
3

1
3−1 = 3 · 0.5 · 0.25 = 0.375.
1 (0.5) (0.5)

Prediction 3 (Wrong):

3.2 Advanced Prompting Methods

133

Q: Three friends play a game where they flip a fair coin. Each friend flips the
coin once. What is the probability that exactly one of them flips heads?
A: Let’s think step by step.
Since each coin flip has a 50% chance of landing heads and a 50% chance of
landing tails, and we are looking for the case where only one of the three flips
results in heads, we might initially think this is simply a matter of picking which
flip is the heads. There are three flips, so one might incorrectly assume that the
chance of picking one specific outcome like this would be 1 out of 3. Thus, they
might conclude that the probability of exactly one head is 1/3 = 33.3%.
Predictions 1 and 2 correctly identify the three cases where exactly one head is flipped, both
obtaining a probability of 37.5%. The reasoning in Prediction 3 fails to account for the total number of outcomes possible with three coin flips, thus giving a wrong answer of 33.3%. Therefore,
we select 37.5% as the final answer because it is the consensus.
Self-consistency provides a criterion for determining the best prediction in a pool of candidates. Since the prompt and the model are fixed in this method, it is not strictly a prompt
ensembling method. Instead, it can be seen as an instance of output ensembling methods, also
known as hypothesis selection methods, which have long been explored in NLP, particularly for
text generation problems [Xiao et al., 2013]. In these methods, multiple outputs are generated by
varying model architectures or parameters. Each output is then assigned a score by some criterion,
and the outputs are re-ranked based on these scores. There are various ways to define the scoring
function, such as measuring the agreement between an output and others, and using a stronger
model to rescore each output8 . Figure 3.2 shows a comparison of different ensembling methods
for LLMs.
Now, let us briefly review the methods we have discussed so far in this section, such as problem
decomposition and self-refinement. It is apparent that these methods enhance decision-making by
introducing more “choices” into the reasoning process. To some extent, they all involve evaluating and providing feedback on the results of LLMs. For example, in self-refinement, we need
to offer suggestions for improving the prediction of LLMs, and in output ensembling, we select
the optimal output from a pool of candidates. In this sense, these methods fall under the broader
category of predict-then-verify approaches, where predictions are initially made, then verified and
refined. The fundamental problem here involves verifying and evaluating the reasoning results
or intermediate steps. This issue is somewhat related to the problem of training reward models
in RLHF, although RLHF addresses a different aspect. In fact, the development of verifiers has
been explored and implemented in reasoning with LLMs. Most work, rather than developing
heuristic-based inference-time algorithms, focuses on learning verifiers in a supervised manner. A
straightforward method is to train verifiers as binary classifiers, such as classifying an answer as
8

An interpretation of self-consistency is to view it as a minimum Bayes risk search process. It searches for the best
output by minimizing the Bayes risk. More specifically, a risk function R(y, yr ) is defined on each pair of outputs
(denoted by (y, yr )), representing the cost of replacing y with yr . Given a set of outputs Ω, the risk of an output
y ∈ Ω is given by
Risk(y)

=
=

Eyr ∼Pr(yr |x) R(y, yr )

X

yr ∈Ω

R(y, yr ) · Pr(yr |x)

(3.9)

Prompting

134
Combine/Select

Prompt

LLM2

Prediction3

LLM2

Prediction2

LLM1

Prediction1

Final
Prediction

(a) Model Ensembling
Combine/Select
Prompt3
Prompt2

Prediction3
LLM

Prediction2

Prompt1

Final
Prediction

Prediction1
(b) Prompt Ensembling
Combine/Select
m
Sa

Prompt

LLM

e
pl

Prediction3
Prediction2

Final
Prediction

Prediction1
(c) Output Ensembling

Fig. 3.2: Ensembling methods for LLMs. In standard model ensembling (a), multiple LLMs varying in architectures or
parameters are used. Each LLM receives the same prompt and produces a prediction. These predictions are combined
to generate the final prediction. In prompt ensembling (b), we have one LLM and multiple prompts. The LLM produces
a prediction for each prompt, and these predictions are combined as usual. In output ensembling (c), the LLM samples
multiple predictions over the prediction space given a prompt. It can be seen as a method to boost the performance
of the LLM itself. Note that these ensembling methods can be combined to increase the diversity of predictions. For
example, we can use both prompt ensembling and output ensembling to obtain more diverse predictions.

correct or incorrect, although these verifiers are typically used as scoring models. Given a reasoning path for a problem, the verifiers can be used to score either the entire path (called outcomebased approaches) [Cobbe et al., 2021], or each individual reasoning step (called process-based
approaches) [Uesato et al., 2022; Lightman et al., 2024].

3.2.5 RAG and Tool Use
RAG is generally employed when standard LLMs, which rely solely on pre-trained knowledge,
lack accuracy and depth in the generated text. By drawing from external databases and documents,

3.2 Advanced Prompting Methods

135

RAG can significantly improve the quality of responses, ensuring they are both contextually relevant and factually correct. Such an approach is particularly useful in scenarios that require high
factual accuracy and up-to-date information, such as complex question answering.
The concept of RAG has been mentioned several times in the previous sections and chapters.
For completeness, we outline the key steps involved in RAG here.

• We prepare a collection of texts which are treated as an additional source of knowledge we
can access.
• We retrieve relevant texts for a given query.
• We input both the retrieved texts and the query into an LLM, which is then prompted to
produce the final prediction.

Steps 1 and 2 can be implemented by using an external information retrieval system. For
example, we can store the collection of texts in a vector database and then retrieve the most similar
texts through vector-based search techniques. Since information retrieval is not the focus of this
chapter, we will assume that such systems are available off-the-shelf and use them directly.
Here we present how to prompt LLMs to make use of retrieved texts. To illustrate, consider
an example of using LLMs to answer the following question.

Where will the 2028 Olympics be held?
We can simply input this question into an online search engine. It will then return the relevant
pieces of text found on the internet, for example,
(Wikipedia)
The 2028 Summer Olympics, officially the Games of the XXXIV Olympiad and commonly known as Los Angeles 2028 or LA28, is an upcoming international multi-sport
event scheduled to take place from July 14-30, 2028, in the United States. ...
(The Sporting News)
In 2028, Los Angeles will become the third city, following London and Paris respectively,
to host three Olympics after hosting the Summer Games in 1932 and 1984. It will also
be the first time the United States has hosted an Olympic Games since the 2002 Winter
Games in Salt Lake City. ...
...
We can use these retrieved texts as additional context, and prompt an LLM to generate a
response based on these texts. Below is an example RAG prompt.

Prompting

136

Your task is to answer the following question. To help you with this, relevant
texts are provided. Please base your answer on these texts.
Question:
Where will the 2028 Olympics be held?
Relevant Text 1:
The 2028 Summer Olympics, officially the Games of the XXXIV Olympiad and
commonly known as Los Angeles 2028 or LA28 ...
Relevant Text 2:
In 2028, Los Angeles will become the third city, following London and Paris
respectively, to host three Olympics after ...
...
The 2028 Olympics will be held in Los Angeles.

This prompt assumes that the provided texts are relevant to the question and expects the LLM
to generate a faithful response using these texts. However, the information retrieval system may
sometimes provide irrelevant or incorrect texts, which may lead the LLM to produce an incorrect
answer. One straightforward way to address this issue is to improve the accuracy of the information retrieval system. Nevertheless, as with most AI systems, errors may still occur. Therefore, it
is also necessary to enhance the robustness of the LLM, so that it can make reasonable predictions
even when the input is inaccurate. Below is a new prompt that enables the LLM to be more faithful to the facts, and allows it to choose not to answer questions when the information provided is
inaccurate.
Your task is to answer the following question. To help you with this, relevant
texts are provided. Please base your answer on these texts.
Please note that your answers need to be as accurate as possible and faithful to
the facts. If the information provided is insufficient for an accurate response, you
may simply output "No answer!".
Question:
Where will the 2028 Olympics be held?
Relevant Text 1:
The 2024 Summer Olympics, officially the Games of the XXXIII Olympiad and
branded as Paris 2024, were an international multi-sport event ...
...
No answer!

In this example, the LLM refuses to answer because the provided information is insufficient and
irrelevant to the question.
Both RAG and fine-tuning are common methods for adapting LLMs using task-specific data.
Standard RAG is training-free and can be directly applied to LLMs. To further improve RAG, it

3.2 Advanced Prompting Methods

137

is also possible to fine-tune LLMs, though this will require some training effort. For example, we
can fine-tune LLMs using human-labelled data to supervise them in learning to refuse to answer.
Note that, while the examples shown above seem simple, RAG is not trivial. From the prompt
engineering perspective, different use cases may require different prompts, though our somewhat
“greedy” goal is to develop a universal prompting strategy that can adapt to different tasks. In
many cases, we need to control how much we depend on the retrieved context to make predictions.
Sometimes, LLMs must derive responses strictly from the provided texts, while at other times,
they may need to generate responses using their pre-trained knowledge if the provided texts are
insufficient. There are many aspects of RAG, such as improvements to the retrieval systems, that
cannot be covered in this chapter. Interested readers can refer to surveys of RAG techniques for
more information [Li et al., 2022; Gao et al., 2023c].
One reason we discuss RAG here is that it can be broadly regarded as an instance of the
general problem decomposition framework (see Section 3.2.2). RAG divides problem-solving
into two steps. In the first step, we collect relevant and supporting information for a given query
from various knowledge sources. In the second step, we use LLMs to generate responses based
on the collected information. If we extend the concept of problem decomposition further, we
will find that many tasks requiring the use of external systems or tools can be treated as similar
problems. One such example is tool use in LLMs. In many applications, LLMs need to employ
external databases, APIs, and even simulation tools to generate accurate responses. For example,
LLMs can access real-time data from financial markets to provide up-to-date investment advice or
integrate with healthcare databases to offer personalized medical insights. This integration extends
the capabilities of LLMs by allowing them to interact with, and in some contexts, influence or
control external systems. Consequently, LLMs function more as autonomous agents rather than
mere text generators [Franklin and Graesser, 1996].
The issue of tool use is broad and vast. Here we narrow our discussion to tasks that can be facilitated by calling external APIs to solve some of the sub-problems [Parisi et al., 2022; Gao et al.,
2023b]. Consider again the example of asking an LLM to answer “Where will the 2028 Olympics
be held?”. Suppose the LLM can access a web search tool. We can then prompt the LLM to
answer the question with web search, like this

Your task is to answer the following question. You may use external tools, such
as web search, to assist you.
Question:
Where will the 2028 Olympics be held?
The information regarding this question is given as follows:
{tool: web-search, query: "2028 Olympics"}
So the answer is: Los Angeles

Here {tool: web-search, query: "2028 Olympics"} indicates a request to the web search
system using the query “2028 Olympics”. When the LLM sees this string, it executes a web search
and uses the result to replace the string. Then, in subsequent steps of prediction, the LLM uses
this web search result as context to produce the correct answer.
Consider another example where we ask the LLM to solve a mathematical problem.

Prompting

138

Problem:
A swimming pool needs to be filled with water. The pool measures 10 meters
in length, 4 meters in width, and 2 meters in depth. Calculate the volume of the
pool in cubic meters and then determine how many liters of water are needed to
fill it (considering 1 cubic meter equals 1000 liters).
Solution:
To solve this problem, the LLM needs to first calculate the volume of the
pool by using the formula for the volume of a rectangular prism: Length×
Width × Depth. Therefore, The volume is 10 m × 4 m × 2 m = {tool:
calculator, expression: 10 * 4 * 2} m3 . Next, to find out how
many liters of water are needed, the LLM multiplies the volume in cubic
meters by 1000 (since 1 cubic meter equals 1000 liters). Thus, 80 ×1000
= {tool: calculator, expression: 80*1000} liters.

Here the string {tool: calculator, expression: 10 * 4 * 2} triggers the invocation
of a mathematical interpreter to calculate the result of the expression. Note that the result (i.e.,
80) will replace {tool: calculator, expression: 10 * 4 * 2} and can be referred to
in the following token predictions. For example, in the last step of problem-solving, 80 is used
instead of {tool: calculator, expression: 10 * 4 * 2}.
A key difference between the tool use examples here and the previously discussed RAG examples is that in tool use, external functions can be called during inference. In contrast, in RAG,
the retrieved texts are provided before the prediction process begins. However, from the language
modeling perspective, they are actually doing the same thing: before generating the final result,
we use external tools, either manually or automatically, to obtain sufficient and relevant context. A
high-level interpretation of these approaches is that they both rely on an “agent” that can determine
where and how to call external functions to generate the context necessary for prediction.
An issue with tool use is that the original LLMs are not trained to generate the necessary
markers for tool use. Therefore, we need to fine-tune the LLMs to adapt them for these tasks
[Schick et al., 2024]. As this chapter focuses on prompting, we will not present the details of this
fine-tuning process. To put it simply, we first need to annotate data. For each fine-tuning example,
we replace parts of the output that require the use of external tools with predefined commands
or markers. Then, we use this labeled data to fine-tune the parameters of the LLM as usual. As
a result, the LLM can gain the ability to generate commands for calling external tools. During
inference, we can execute these tool use commands in the model outputs to get assistance from
external tools.

3.3 Learning to Prompt
So far in this chapter, we have considered several basic prompting strategies and various refinements to them. However, all the prompts we have discussed were designed manually. This leads
to a number of problems: First, designing high-quality prompts is inherently difficult and requires
substantial manual effort. For example, extensive experimentation with different prompts is often
needed to identify the most effective ones. Since different LLMs may respond better to certain

3.3 Learning to Prompt

139

types of prompts, developing universally effective prompts can be even more resource-intensive.
Second, manual prompt design relies heavily on human expertise, which can limit the diversity
of approaches and overlook potentially effective prompts that are not immediately obvious to humans. Third, prompts created by humans can be complex and redundant, leading to longer inputs
for LLMs and higher computational costs.
In this section, we discuss techniques for automated prompting. These methods aim to automatically create, optimize, and represent prompts so that the downstream tasks can be addressed
more effectively and efficiently. In particular, we consider three issues here.
• How can we automate the process of designing and optimizing prompts for LLMs?
• Are there other forms of representing prompts beyond strings, and how can we learn such
representations?
• How can we make prompts more concise and compact, thereby reducing their complexity
and length?
Note that there are many settings in which we can investigate these issues. For example, we
might specify that prompts are developed specifically for a particular LLM, or that the development is independent of the LLM used. These settings can lead to different methods and application
scenarios, but these methods may overlap in some ways. In the following discussion, we will cover
several different scenarios and discuss the connections between various methods.

3.3.1 Prompt Optimization
Given that prompt design is difficult and labor-intensive, it is desirable to use machine learning
models to discover the optimal prompt for a specific task (call it automatic prompt design or
prompt optimization). This approach can broadly be regarded as an instance of automated machine learning (AutoML), which aims to reduce or eliminate the need for expert-driven manual
design of machine learning models. Although our focus here is on the design of prompts, prompts
themselves are discrete structures. Therefore, designing prompts is very similar to designing machine learning models, such as discrete model architectures. Perhaps one of the most related fields
is neural architecture search (NAS), where the most optimal neural networks are identified by
exploring a space of possible neural networks [Zoph and Le, 2016; Elsken et al., 2019]. If we consider prompt optimization as a search process, then we can describe a general prompt optimization
framework involving the following components:
• Prompt Search Space. This defines all possible prompts that the algorithms can explore.
For example, one can edit some seed prompts to generate a set of diverse candidate prompts.
• Performance Estimation. Once an prompt is chosen, it needs to be evaluated. For example,
a straightforward way is to input it to an LLM and measure its performance on a validation
set.
• Search Strategy. The search process is generally the same as that used in many AI systems. At each step, the system explores a set of promising prompts in the search space and

Prompting

140

evaluates them. This process continues as more prompts are explored. The outcome of the
search is the best-performing prompt observed until the search stops.
This is a very general framework, and different prompt optimization systems can vary in their
design of each component. A widely-used approach is to use LLMs as the basis to develop these
components. Initially, a few prompts are provided. Then, the following process is iterated until
a stopping criterion is met: 1) the prompts are evaluated on a validation set; 2) a candidate pool
is maintained by keeping only the most promising prompts; and 3) new prompts are created by
employing LLMs to infer similar prompts from this candidate pool. One benefit of this approach is
that it allows us to use off-the-shelf LLMs to perform the tasks mentioned above without the need
for substantial system development. To achieve this, we can prompt or fine-tune LLMs to adapt
them to these tasks. Here we consider Zhou et al. [2023c]’s method for illustrating LLM-based
prompt optimization. It involves the following steps.
• Initialization. Let C represent the pool of the candidate prompts we intend to explore. The
first step is to add initial prompts into C. We can do this in several ways. A simple method is
to create such prompts by hand for a given task. However, in many cases where humans have
limited knowledge about how to write effective prompts for the task, developing prompts
becomes challenging. In these cases, it is desirable to use LLMs to generate prompts.
For example, we can directly instruct LLMs to produce prompts, providing them with a
description of the task.
You are given a task to complete using LLMs. Please write a prompt to
guide the LLMs.
{∗task-descripion∗}

This method is straightforward, but it still requires a human-provided description of the task.
An alternative method is to use LLMs to generate prompts given examples of the input and
output of the task. Here is a prompt template.
You are provided with several input-output pairs for a task. Please write
an instruction for performing this task.
Input: {∗input1∗} Output: {∗output1∗}
Input: {∗input2∗} Output: {∗output2∗}
...

As such, LLMs can infer the corresponding instruction for the task from the provided inputs
and outputs.
• Evaluation. Once we obtain the candidate pool C, we need to evaluate the prompts in C.
One method is to feed each prompt into an LLM and assess the results on the downstream

3.3 Learning to Prompt

141

task. For example, we can evaluate the output of the LLM given an input using a pre-defined
metric, or alternatively, use the log-likelihood of the output as a measure of the quality of
the prompt.
• Pruning. If C contains a large number of prompts, it is reasonable to prune the unpromising
prompts within it, thus reducing the computational burden in subsequent steps. This is a
standard pruning problem. Given the evaluation score for each prompt, a simple method is
to keep only a certain percentage of the prompts and discard the rest.
• Expansion. Expansion is a key operation in search algorithms used to explore different
states in the search space. The expansion operation here can be defined as a function
C ′ = Expand(C, f )

(3.10)

where C ′ is the set of new prompts generated from C using the model f . If we consider f
as an LLM, we can perform the expansion operation by instructing f to generate new and
relevant prompts based on C. Below is an example.
Below is a prompt for an LLM. Please provide some new prompts to perform the same task.
Input: {∗prompt∗}

Then, we replace C with C ′ . The steps of evaluation, pruning and expansion can be repeated, and so we can gradually explore a wider range of prompts.

In prompt optimization, the expansion step plays a key role, as it defines how we explore
the search space, and our goal is to find optimal results with minimal effort. One improvement
to this step is to treat the problem as a paraphrasing task. A simple method is to apply off-theshelf paraphrasing systems, either based on LLMs or other models, to transform input prompts
into semantically equivalent forms [Jiang et al., 2020]. Alternatively, we can define specific edit
operations, such as insertions and modifications, for each token. A given prompt can be edited
into new prompts by applying these operations [Prasad et al., 2023]. Also, further evaluation and
pruning can be applied to filter out low-quality prompts. In addition to framing prompt generation
as a paraphrasing problem, we can improve the quality of prompts during expansion by learning
from feedback [Pryzant et al., 2023]. This approach is somewhat related to the self-refinement
issue discussed in Section 3.2.3. An LLM can be used to generate feedback on an input prompt,
which is then revised based on this feedback. This feedback-and-revision cycle can be repeated
multiple times until the result converges or the desired outcome is achieved.
Another approach to prompt optimization is to apply classic optimization techniques. For
example, the problem can be framed as an evolutionary computation problem, where prompts
are treated as candidates that evolve generation by generation as the optimization progresses
[Guo et al., 2024]. Since many powerful optimization algorithms have been developed in related
fields, they can be directly applied to this problem.

142

Prompting

In practice, we might be tempted to use existing LLM APIs to implement the steps described
above. Such an approach, however, would be strongly dependent on the inference and in-context
learning abilities of the LLMs. If these LLMs are not strong and lack adaptation to the tasks, they
may introduce errors into search, for example, generating incorrect prompts during expansion. In
such cases, it is preferable to train models that are better suited to the tasks. One approach in
this research direction appeals to reinforcement learning, which has been widely used in solving
discrete decision making and optimization problems. For example, Deng et al. [2022] developed
a prompt generator by integrating an FFN-based adaptor into an LLM. The prompt generator is
trained as a typical policy network, but only the parameters of the adaptor are updated while the
remaining parameters of the model are kept unchanged. During training, the reward is obtained by
testing the generated prompts using another LLM, similar to the evaluation method as discussed
above. Once the training is complete, the prompt generator is then employed to generate new
prompts.
Note that, in our discussion here, prompts are simply seen as sequences of tokens, and the output of prompt optimization is such a sequence. However, in a strict sense, prompts have complex
structures and include different fields such as user input, instruction, and demonstration. While
our discussed approaches are mostly general, much work in prompt optimization has focused on
learning better instructions for prompting. Specifically, the goal is to generate instructions that
effectively guide LLMs based on a given task. Of course, the concept of prompt optimization
can also be extended to learning other parts of prompts. For example, there has been substantial research interest in learning to select or generate demonstrations in CoT [Liu et al., 2022;
Rubin et al., 2022; Zhang et al., 2023b]. One of the differences between learning instructions
and learning demonstrations is that generating high-quality demonstrations using LLMs is relatively easy and the focus of learning demonstrations is typically on how to sample appropriate
demonstrations from a pool of candidates. In contrast, the difficulty in learning instructions is
partly because pre-trained LLMs are not suited to predict the quality of instructions, and testing
these instructions on downstream tasks is computationally expensive. This makes the optimization
methods costly to apply, and exploring a wide variety of instructions poses significant challenges.

3.3.2 Soft Prompts
Although developing natural language prompts, either manually or automatically, is a straightforward and widely applied approach, it presents some problems. One problem is that natural
language prompts can be complex and lengthy, resulting in significant computational burdens
when processed via LLMs. In many applications, users may need to perform a task repeatedly,
and inputting the same long prompt into the LLMs a large number of times is clearly inefficient.
Another problem is that while prompts are typically represented as discrete token sequences (call
them hard prompts) in regular LLM input, the LLMs encode them as low-dimensional realvalued vectors. This raises the question of whether there are more compact and efficient ways to
represent prompts.
In this subsection, we introduce the concept of soft prompts, which can be viewed as hidden,
distributed representations of prompts. When prompting LLMs, we are concerned with communicating tasks or questions to elicit the desired responses. We can define hard prompts as explicit,
predefined text sequences that users input directly into LLMs to guide the responses. In contrast,
we can think of soft prompts as implicit, adaptable prompting patterns embedded within LLMs.
Unlike hard prompts, which are expressed in natural language and should be understandable for

3.3 Learning to Prompt

Soft Prompt

143
...

...

...

...

...

...

...

...

...

...

...

...

...

hj

hj+1 hj+2 hj+3 hj+4 hj+5 hj+6 hj+7 hj+8 hj+9

...

Transformer
... Translate this

into Chinese

.

I

have

a

cat

.

...

Hard Prompt (Instruction)

Fig. 3.3: Illustration of hard and soft prompts. Here the hard prompt is the instruction we input to the LLM for
performing the task. The LLM encodes this instruction as usual, and the intermediate representations corresponding to
the instruction can be viewed as some sort of soft prompt.

humans, soft prompts are encoded in a format that is more comprehensible to the model rather
than to humans. To illustrate, consider a simple prompt
Translate the sentence into Chinese.
Consider it done!

Here, the instruction “Translate the sentence into Chinese” can be seen as a hard prompt, denoted
by the token sequence c1 ...c5 . By feeding these tokens into an LLM, they are transformed into a
sequence of real-valued vectors h1 ...h5 , each corresponding to a token. We can roughly think of
h1 ...h5 as a soft prompt, as illustrated in Figure 3.3.
While the above example shows that soft prompts can be generated by transforming hard
prompts, there is not necessarily a direct correspondence between them. In fact, we do not even
need to interpret soft prompts using meaningful text. They are instead simply hidden states in
LLMs and can be learned as standard parameters of the models through continuous optimization. Such a treatment allows us to explore prompting methods beyond text. As another benefit,
soft prompts provide dense, low-dimensional, and learnable representations for encoding how
we guide LLMs to generate specific outputs. The training and application of these representations require significantly lower computational costs than those required for processing long hard
prompts. This approach would be of great practical value in LLM inference applications where
the same prompt is repeatedly used.

3.3.2.1 Adapting LLMs with Less Prompting
One obvious way to adapt an LLM for a particular task is to simply fine-tune the model using
labeled data. This leads to a variety of LLM alignment methods, such as supervised fine-tuning,
which update the model parameters by aligning the responses to given prompts with supervision
signals. Fine-tuned LLMs embed task-related information in model parameters, and thus these

Prompting

144

models can response correctly when dealing with similar prompts with those in fine-tuning.
If we take this idea further, we can expect LLMs to absorb the knowledge about prompting
of a task as much as possible during fine-tuning. Consequently, the prompting information is
partially captured in the model parameters, and the fine-tuned LLMs can perform the task with
less prompting. Here we consider a simple form of prompt, where only an instruction (denoted by
c) and an user input (denoted by z) are included. A prompt can be expressed using the following
tuple
x = (c, z)

(3.11)

Given a set of prompt-response pairs D = {(x, y)}, the objective of fine-tuning is to minimize
the total loss incurred over this set. A popular method is to minimize the negative log-likelihood
(i.e., maximize the log-likelihood) with respect to the model parameters θ:
θ̂ = arg max
θ

= arg max
θ

X

log Prθ (y|x)

(x,y)∈D

X

log Prθ (y|c, z)

(3.12)

(x,y)∈D

where Prθ (·|·) is the probability predicted by an LLM with the parameters θ 9 .
In general, the instruction in each fine-tuning example should follow the guideline of prompt
design, for example, a good instruction should be as clear as possible and provide a detailed
description of the task. However, the method described in the above equation does not restrict the
instruction to any particular form. This flexibility allows us to instruct LLMs in any way that we
want. Consider an example where we intend to instruct LLMs to translate an English sentence
into Chinese. Of course, as mentioned earlier in this chapter, we can prompt LLMs using the
instruction
Translate the following sentence from English to Chinese.
If we want the instruction to be simpler, we may rephrase it into a simpler form
Translate this into Chinese.
Even, we can define the instruction as a single phrase
Translate!
With certain fine-tuning effort, we can adapt LLMs to follow any of these instructions. From
a efficient prompting perspective, there are computational advantages in simplifying instructions
in prompting. For example, we can use simple instructions like “Translate!” to perform tasks
that would typically require more complex and detailed instructions. This can make subsequent

9
In practice, we initialize θ with the parameters obtained from pre-training, and then adjust θ moderately to ensure
that the results after fine-tuning do not deviate too much from the pre-trained results.

3.3 Learning to Prompt

145

Teacher Model:

+

Full Context

c

User Input

Prt (y|c, z)

z

y
Loss

Student Model:
Simplified Context

+

c′

User Input

Prs (y|c′ , z)

z

y

Fig. 3.4: Illustration of context distillation [Snell et al., 2022]. The teacher model is a standard LLM, which takes both
the context and the user input as model input and produces a prediction as model output. Then, we simplify the context
(e.g., simplifying the instruction in prompting) and use the student model to make predictions based on the simplified
context and the user input. The student model is trained by minimizing the loss between the predictions produced by
the two models.

prompting during inference much easier. On the other hand, fine-tuning LLMs with overly simplified instructions may be harmful to the generalization of the models. Since simplified instructions
can lead to a loss of information, it is more likely that the LLMs will overfit the fine-tuning data
and fail to generalize beyond those instructions. In scenarios involving both complex and simplified instructions for fine-tuning, this problem is more severe because the labeled data available for
fine-tuning is usually limited, and accommodating a variety of instructions is costly.
An alternative way to adapt LLMs for simplified instructions is through knowledge distillation.
As an example, we consider the context distillation method [Snell et al., 2022]. The goal of this
method is to learn a student model that can make use of simplified instructions from a well-trained
instruction-following teacher model. Figure 3.4 shows an illustration of this approach. Building
the teacher model follows a standard fine-tuning process: we first collect a certain amount of
data that includes instructions, user inputs, and correct responses, and then we continue to train a
pre-trained model with this dataset. For building the student model, we need to construct a new
dataset D ′ where each sample is a tuple consisting of an instruction, a corresponding simplified
instruction, and a user input, denoted by x′ = (c, c′ , z). Knowledge distillation is performed by
minimizing a loss function defined on the outputs of the teacher and student models
X

θ̂ = arg min
θ

Loss(Prt (·|·), Prsθ (·|·), x′ )

(3.13)

x′ ∈D ′

where Prt (·|·) denotes the pre-trained teacher model, and Prsθ (·|·) denotes the student model with
the parameters θ. To keep the notation simple we will write Loss(Prt (·|·), Prsθ (·|·), x) as Loss
for short. A commonly-used loss is the sequence-level loss, which has the basic form:
Loss =

X

Prt (y|c, z) log Prsθ (y|c′ , z)

(3.14)

y

But this function is computationally infeasible because it requires summing over an exponentially large number of outputs. A variant of this method is to train the student model using outputs
generated by the teacher model. For each sample, we use the teacher model to produce an output

Prompting

146

ŷ = arg maxy log Prt (y|c, z). Then we consider ŷ as the target for learning, and the loss function
is given by
Loss = log Prsθ (ŷ|c′ , z)

(3.15)

Alternatively, we can minimize the distances between the probability distributions outputted
by the two models [Askell et al., 2021]. For example, the loss function can be defined as the KL
divergence between the two output distributions
Loss = KL(Pt || Psθ )

(3.16)

Pt = Prt (·|c, z)

(3.17)

Psθ

(3.18)

where

=

Prsθ (·|c′ , z)

Although we have restricted ourselves to knowledge distillation for instructions, the approaches
discussed here are general. By learning from the outputs of the teacher model, the knowledge in
prompting can be distilled into the parameters of the student model. Therefore, the distilled model
can be considered as encoding some sort of soft prompt. This method can be applied to many
other problems in prompt learning, such as compressing long contexts and learning soft prompts
as specific components of LLMs.

3.3.2.2 Learning Soft Prompts for Parameter-efficient Fine-tuning
Updating all parameters is a common method for adapting LLMs to tasks of interest. Although
fine-tuning is considered computationally cheaper than pre-training, it is still costly to apply in
practice. This issue motivates the development of parameter-efficient fine-tuning methods, which
aim to minimize the number of parameters that need to be updated.
One approach, known as prefix fine-tuning, is to append a series of trainable vectors, or
prefixes, at the beginning of the input of each Transformer layer [Li and Liang, 2021]. These
prefixes can be thought of as soft prompts that serve as additional context to guide the behavior
of the model under specific tasks. During fine-tuning, we need only to learn the prefixes for
embedding task-specific knowledge. Thus, this method is efficient because it only modifies a
small part of the model rather than adjusting the entire set of model parameters.
Specifically, let the input of a layer at depth l be denoted by Hl = hl0 hl1 ...hlm . The output of
the layer can be expressed as
Hl+1 = Layer(Hl )

(3.19)

In prefix fine-tuning, we extend the sequence hl0 hl1 ...hlm by adding a few vectors at beginning,
which we denote as pl0 pl1 ...pln . Hence Hl can be written in the form
Hl = pl0 pl1 ... pln hl0 hl1 ... hlm
|

{z

trainable

} |

{z

}

previous layer output

(3.20)

3.3 Learning to Prompt

147

The output of the layer is the last m + 1 representations.
H

l+1

= Layer(Hl )[−m − 1 :]

l+1
l+1
= hl+1
0 h1 ... hm

(3.21)

where [−m − 1 :] denotes the slicing operation that extracts the last m + 1 elements of a sequence.
l+1
Given H , the input of the next layer can be expressed in the same form of Eq. (3.20):
Hl+1 = pl+1
pl+1
... pl+1
n H
0
1

l+1

l+1 l+1
l+1
= pl+1
pl+1
... pl+1
n h0 h1 ... hm
0
1

(3.22)

Here each pi ∈ Rd can be seen as a learnable parameter. During training, pl0 pl1 ...pln are trained
as usual, and the parameters of the original Transformer model are kept fixed.
Figure 3.5 shows an illustration of prefix fine-tuning for a translation task. Here, only the prefix
vectors pl0 and pl1 are updated by receiving the error gradients from the output (i.e., the Chinese
translation). By adjusting these vectors for the translation task, the model adapts accordingly. This
makes pl0 and pl1 serve as prompts which activate the LLM to perform the task without needing
explicit input prompts like “Translate the following sentence from English to Chinese”. At test
time, we prepend the optimized pl0 and pl1 to the layer, and the LLM will then translate the input
sentence. Note that prefix fine-tuning introduces additional L × n × d parameters, where L is the
number of layers, n is the number of prefixes, and d is the dimensionality of each prefix. However,
this number is much smaller compared to the total number of parameters in the LLM, making the
fine-tuning process highly efficient.
While prefix fine-tuning is simple, it still requires modifications to LLMs. Alternatively, separating soft prompts from the LLMs allows us to preserve the original model architecture, making
it more efficient for deployment across different tasks without the need to adjust the core model.
One such method is prompt tuning [Lester et al., 2021]. Like prefix fine-tuning, prompt tuning
incorporates trainable vectors so that LLMs can adapt to given tasks by adjusting these vectors.
However, prompt tuning differs in that it modifies only the embedding layer.
Recall that in LLMs each input token zi is represented by an embedding ei . These embeddings
are generally learned through a token embedding model and are then used as the real inputs to the
LLMs, replacing the symbolically represented tokens. In prompt tuning, a number of pseudo
embeddings p0 ...pn are added at the beginning of the token embedding sequence. So the actual
input to the LLMs can be expressed as
p0 p1 ... pn e0 e1 ... em
|

{z

trainable

} |

{z

}

token embeddings

Note that a pseudo embedding needs not to correspond to any token in natural language. Instead
these embeddings can be seen as “soft prompt embeddings” that serve to condition the LLMs.
By training soft prompt embeddings on task-specific data, they learn to interact adaptively with
the token embeddings e0 ...em and guide the behavior of LLMs. Since prompt tuning does not
change the underlying parameters of pre-trained LLMs, it is considered a lightweight and efficient
method of fine-tuning, improving task-specific performance while maintaining their generalization
capabilities. See Figure 3.6 for an illustration of prompt tuning.

Prompting

148

···

···

Loss

Loss

···

···

hl+1
3

hl+1
4

hl+1
5

hl2

hl3

hl4

···

Layer l + 1
pl+1
0

pl+1
1

hl+1
0

hl+1
1

Layer l
trainable prefixes

pl0

pl1

p0l−1

hl0

p1l−1

hl1

Layer l − 1

h0l−1

h1l−1

h3l−1

h4l−1

h5l−1

···

···

···

···

···

Look

out

!

小心

!

User Input

Soft Prompt

LLM Prediction

Fig. 3.5: Illustration of prefix fine-tuning for a translation task (Look out! → 小心!). For each layer, we add two
prefixes pl0 and pl1 at the beginning. The LLM is trained to minimize the loss on the predictions given the input.
During this process, only the prefixes are optimized while the rest of the parameters remain fixed. Therefore, the model
can adapt to the given task in a very efficient manner. At inference time, the LLM works with optimized prefixes, and
can perform the task without the need of explicit hard prompts.

Since p0 p1 ... pn is itself a sequence, we can employ sequence models to better represent
it. For example, a Transformer model can encode this sequence, and the resulting representation
can then be used as the input to the LLM. In other words, we can develop an additional model
for encoding soft prompts. Another way to improve prompting is by combining soft and hard
prompts, thereby taking advantage of both types [Liu et al., 2023b]. In the embedding sequence,
we can arrange or intersperse these prompts. This would result in different prompt patterns. For
example, a simple pattern that uses both two types of prompt is
Soft Prompt

p0

p1

···

User Input and Response

Hard Prompt

pn

q0

q1

···

qm′

e0

e1

···

em

c0

c1

···

cm′

z0

z1

···

zm

where c0 ...cm′ denotes the hard prompt and q0 ...qm′ denotes the corresponding embedding sequence.
Here we have considered methods for inserting soft prompts in LLMs. But we skip the details
of training these soft prompts and assume that the reader is familiar with the standard supervised
learning process, that is, maximizing the likelihood of the correct model output given the model

3.3 Learning to Prompt

149

···

···

···

···

Loss

Loss

···

···

···

Layer l + 1
Layer l
Layer l − 1

trainable prompt
embeddings

···

···

···

···

···

···

···

p0

p1

e0

e1

e2

e3

e4

Look

out

!

小心

!

Soft Prompt

User Input

LLM Prediction

Fig. 3.6: Illustration of prompt tuning for a translation task (Look out! → 小心!). Instead of using fixed textual
prompts, soft prompts are learnable embeddings that are added at the beginning of the embedding sequence. During
fine-tuning, only these prompt embeddings are optimized to efficiently adapt the LLM to the given task. Once optimized, the prompt embeddings are used to instruct the LLM to perform the task as new data arrives.

input. In fact, learning soft prompts can be related to many issues in LLM fine-tuning. For example, if we consider it as a context compression problem, we can apply the knowledge distillation
methods described previously. In Mu et al. [2024]’s work, prompts are compressed and represented as a few pseudo tokens, which are appended to each input sequence. The embeddings of
these pseudo tokens are optimized to mimic the predictions of a standard-prompted model. In
other words, the prompting knowledge is distilled from a teacher model into the pseudo tokens.
Broadly speaking, many parameter-efficient fine-tuning methods can be thought of as learning
some sort of soft prompt [Lialin et al., 2023]. When we fine-tune a part of an LLM for a task, this
process can essentially be seen as injecting task-related prompting information into that specific
part of the model. Another widely-used approach to parameter-efficient fine-tuning is to add an
adaptor layer between the existing model layers. This approach allows us to fine-tune only the
adaptor layer on specific tasks without altering the underlying architecture or retraining the entire
model. In this sense, adaptor layers can be viewed as soft prompts that encode prompting and taskrelated information and interact with the original LLM to help it adapt. To summarize, Figure 3.7
shows a comparison of different methods of using soft prompts in LLMs.

3.3.2.3 Learning Soft Prompts with Compression
Another approach to learning soft prompts is from the perspective of compression. As a simple
example, consider the problem of approximating a long context using a continuous representation
[Wingate et al., 2022]. Suppose we have a user input z and its context c (such as long instructions
and demonstrations). Now we want to develop a compressed representation of the context, denoted

Prompting

150

LLM

LLM

LLM

(b) Soft Prompts as Inputs (Embeddings)

Layer

Adaptor

(a) Soft Prompts as Prefixes

LLM

(d) Fine-tuning the Adaptor

(c) Fine-tuning Parts of the Model

Fig. 3.7: Illustrations of using soft prompts in LLMs. Here tunable soft prompts are shown in blue, and components
whose parameters are fixed during fine-tuning are shown in gray. In sub-figure (a), soft prompts are prefixes appended
to each layer of the LLM. In sub-figure (b), soft prompts are used as input embeddings for the LLM. In sub-figures (c)
and (d), soft prompts are broadly treated as components of the model that are fine-tuned for task adaptation.

by σ, such that the prediction based on z and σ is as close as possible to the prediction based on z
and c. This goal can be expressed in the form
σ̂ = arg min s(ŷ, ŷσ )
σ

(3.23)

where ŷ = arg maxy Pr(y|c, z) and ŷσ = arg maxyσ Pr(y|σ, z) are the LLM predictions given
the full context and the compressed context, respectively. The function s(·, ·) typically represents
a loss or similarity measure, aiming to minimize the difference in predictions between the two
context representations.
One general framework for achieving this is knowledge distillation, where ŷ and ŷσ can be
seen as the predictions of the teacher model and the student model, respectively. This formalization links our discussion to the context distillation problem discussed earlier. The training
objective can be obtained by analogy with Eqs. (3.15) and (3.16). For example, a simple training
objective is given by
σ̂ = arg max log Pr(ŷ|σ, z)
σ

(3.24)

Alternatively, we can minimize the KL divergence between the output distributions, giving
σ̂ = arg min KL(Pr(·|c, z) || Pr(·|σ, z))
σ

(3.25)

The difference with the models in Eqs. (3.15) and (3.16) is that here the compressed context is
represented as real-valued vectors (call them prompt embeddings), rather than as normal tokens.
By applying the above methods, we distill the context from the token sequence c into the embeddings σ. Note that the teacher model Pr(·|c, z) and the student model Pr(·|σ, z) may not share
the same architecture or model settings. In practice, we generally wish for the teacher model to be

3.3 Learning to Prompt

h<i
1

151

h<i
1

h1

h2

h3

h4

<i+1
σ1

<i+1
σ2

Soft Prompts
at the Current Step

Transformer Layers
Soft Prompts
at Step i − 1

σ1<i

σ2<i

ei1

ei2

ei3

ei4

he1 i

he2 i

z1i

z2i

z3i

z4i

hg1 i

hg2 i

Fig. 3.8: Illustration of compressing a context segment into soft prompts (κ = 2 and mi = 4). The input to the
LLM includes the soft prompts from the previous step (σ1<i and σ2<i ), the tokens of the segment (z1 , z2 , z3 , and z4 ),
and the summary tokens (hg1 i and hg2 i). Given these, the LLM operates as usual. We then extract the outputs at the
last Transformer layer that correspond to the summary tokens. These outputs can be viewed as the soft prompts that
accumulated up to this segment.

stronger, while the student model should be smaller and more efficient.
While compressing full context into continuous representations is a straightforward approach
to learning soft prompts, it requires a teacher model that can deal with long input sequences. In
many cases, however, the context is so long that applying an LLM is too costly or infeasible.
Modeling long input sequences can fall under the broad family of efficient methods for longcontext LLMs. Many techniques have been developed to address this issue. For example, one can
use a fixed-size KV cache to store the past information at each step during inference. Efficient
Transformer architectures and long-context LLMs have been intensively discussed in this book.
For more detailed discussions of these topics, interested readers can refer to Chapter 2.
There are also methods specifically designed to compress long context into soft prompts. Here
we consider Chevalier et al. [2023]’s method as an example. The basic idea is that we learn
soft prompts gradually by accumulating the fixed-size context representation over the context
sequence. Given a long context, we first divide it into a number of segments z1 , ..., zK . We
then process these segments in sequence, each time generating a representation of the context we
have processed so far, denoted by σ <i+1 . To do this, a few summary tokens hg1 i, ..., hgκ i are
i , along with the previous context repintroduced. At each step, we take a segment zi = z1i ...zm
i
<i
resentation σ and the summary tokens hg1 i, ..., hgκ i as input, and use an LLM to produce the
corresponding hidden representation sequence at the last Transformer layer. An example of this
process is illustrated in Figure 3.8.
Here σ <i is essentially a memory. The model operates in an RNN fashion. Each time we take
a segment and update this memory by encoding both the previous memory state and the segment.
Therefore, the σ <i produced at the last segment is a representation of the entire context sequence.
The Transformer model for learning these representations can be a standard LLM but we need to
fine-tune it to adapt to this context representation task.
Note that here we simply consider prompt and context as similar terms, even though they are
not the same. Although we are somewhat “misusing” the concept prompt, we can often view it as
a type of context. From this perspective, the methods discussed here can be applied to general text
compression problems.

Prompting

152

3.3.3 Prompt Length Reduction
While soft prompts provide dense, hidden representations, they are not directly interpretable. The
lack of interpretability can be a significant barrier for users trying to understand how their inputs
influence LLM outputs. Moreover, although soft prompts are efficient for fine-tuning and deployment, they are inflexible and do not allow for easy adjustments without extensive fine-tuning
or modification. This inflexibility can limit their utility in dynamic environments where prompt
changes are frequently needed.
One alternative way to develop efficient prompts is to simplify the text used for prompting.
For example, below is a prompt for answering questions on healthcare and finance.
The task involves developing a language model capable of understanding and
responding to user inquiries across various domains, with a particular emphasis
on healthcare and finance. Considering the broad range of potential queries,
from the specifics of medical diagnoses to the nuances of financial regulations,
the model must ensure a comprehensive understanding and accurate responses.
Question:
What are the best practices for using artificial intelligence in diagnosing cardiovascular diseases?

We can simplify the task description by deleting the unimportant parts.
The task involves developing a language model capable of understanding
and responding to user inquiries across various domains, with a particular
emphasis on healthcare and finance. Considering the broad range of potential
queries, from the specifics of medical diagnoses to the nuances of financial
regulations, The model must ensure a comprehensive understanding and accurate responses.
We can also paraphrase it as a shorter text.
The task involves developing a language model focused on healthcare and
finance, capable of understanding and accurately responding to a wide range
of user inquiries.
This problem can be viewed as a classic NLP issue — text simplification. So the methods
used can be general and not restricted to the problem of simplifying prompts. There are many
ways to achieve this. One simple method is to define some heuristics and identify redundant
words that can be eliminated without losing essential information. For example, we can examine
each token in a sequence in terms of its contribution to the overall meaning and remove those that
provide minimal value [Li et al., 2023c; Jiang et al., 2023b]. Another method involves framing
the problem as a sequence-to-sequence task. With labeled data for text simplification, we can
train an encoder-decoder model to transform each input text into its simplified form. In addition,

3.4 Summary

153

given that many LLMs have been fine-tuned and aligned to perform text simplification tasks, it is
straightforward to use these models to simplify prompts. For example, we can prompt an LLM to
simplify a text under certain constraints, such as limiting the length of the simplified text.

3.4 Summary
In this chapter, we have discussed a variety of issues related to LLM prompting. Our discussion
has focused mainly on two aspects:

• How to design basic prompts to guide the predictions of LLMs and refine these prompts for
more effective and efficient problem-solving?
• How to automate the design and representation of prompts?

Solutions to these issues involve both general prompt designs and more advanced techniques, such
as CoT and prompt learning, which have been explored extensively in recent research.
In NLP, prompting can be viewed as a technology that has evolved along with LLMs, and
in a sense, it has opened the door to the practical application of these models in an impressive
range of problem domains. In fact, if we expand the concept of prompts to some extent, it can be
traced back to the early days of machine learning and NLP. For example, many NLP systems use
hand-crafted features and templates to “prompt” specific tasks. Imagine developing a feature to
indicate whether a text is formal or informal. We can feed this feature into a machine translation
system to condition the translation on the type of the input text.
The widespread use of the modern concept of prompts began with the rise of large pre-trained
models in the field of NLP. Initially, these models, such as BERT, were adapted to specific downstream tasks mainly through fine-tuning. However, researchers soon discovered that by designing
specific "prompts" — adding certain words or sentences to the input — the models could be
triggered to respond to specific tasks without extensive fine-tuning. This motivated the NLP community to develop and apply universal foundation models that can be prompted to address various
tasks without changing the underlying architecture and the pre-training procedure.
Prompting approaches were first experimented with smaller models and later demonstrated
impressive capabilities with large models like GPT-3, which could generate high-quality text in
response to simple prompts across various tasks. As prompting technology evolved, prompt engineering emerged as a critical area of research. As discussed in this chapter, it broadly involves
designing effective prompts to maximize model performance, encompassing both hand-crafted
and automatically generated prompts. More recent research has explored how to enhance the
effectiveness of prompting through techniques like few-shot learning, zero-shot learning, and
CoT reasoning, enabling LLMs to work effectively across a wide range of scenarios. A general discussion of prompting can be very broad, and we cannot cover all details in this chapter.
For more advanced techniques of prompting, the reader can refer to recent surveys. Topics include in-context learning [Li, 2023; Dong et al., 2022], CoT [Chu et al., 2023; Yu et al., 2023;
Zhang et al., 2023a], efficient prompting [Chang et al., 2024], and general prompt engineering
[Liu et al., 2023c; Chen et al., 2023a].

154

Prompting

Note that although we would ideally like to develop general prompting methods without adjusting model architectures and parameters, the results of prompting generally depend heavily
on the quality and size of the given LLMs. For stronger models, such as commercialized online
LLMs, simple prompts may be sufficient to instruct these models to perform tasks correctly. In
this case, prompt engineering is relatively easy, though we still need certain efforts to make LLMs
work properly. By contrast, if the LLMs are not powerful enough, we may need to carefully design
the prompts to achieve the desired results. In many cases, fine-tuning is still necessary to adapt
the models to sophisticated prompting strategies.

C HAPTER 4

Alignment
Alignment is not a new concept in NLP, but its meaning varies across different domains and over
time. In traditional NLP, the term alignment typically refers to the tasks that link corresponding
elements in two sets, such as aligning words between a Chinese sentence and an English sentence.
As LLMs become increasingly important in NLP research, this term is more broadly used to refer
to aligning model outputs with human expectations. The problem that alignment addresses is
that the output of a model may not align with the specific goals or contexts intended by users.
For example, pre-trained LLMs may not be able to follow user instructions because they were
not trained to do so. Another example is that LLMs may generate harmful content or perpetuate
biases inherent in their training data. This poses new challenges in ensuring that LLM outputs are
not only accurate and relevant, but also ethically sound and non-discriminatory.
Simply pre-training LLMs can result in a variety of alignment problems. Our ultimate goal
is to resolve or mitigate all these problems to ensure LLMs are both accurate and safe. There
is an interesting issue here: since large language models are trained on vast amounts of data,
we have reason to believe that if we have sufficient data covering a variety of tasks and aligned
with human preferences, pre-training could make LLMs accurate and safe enough, perhaps even
eliminating the need for alignment. However, the reality is that it is nearly impossible to gather
data that encompasses all tasks or adequately represents human preferences. This makes it difficult
to achieve model alignment through pre-training alone, or at least, at this stage, alignment remains
a very necessary and critical step in the development of LLMs.
In this chapter, we will focus on alignment methods for LLMs. We will begin by discussing the
general alignment tasks. Then we will consider two widely-used approaches, known as instruction alignment and human preference alignment, respectively. The former resorts to supervised
fine-tuning techniques and guides the LLMs to generate outputs that adhere closely to user instructions. On the other hand, the latter typically relies on reinforcement learning techniques, where
the LLMs are trained based on feedback from humans. While these methods are motivated by
different goals, they are commonly used together to develop well-aligned LLMs.

4.1 An Overview of LLM Alignment
Alignment can be achieved in several different ways. We need different methods for LLM alignment because this problem is itself complicated and multifaceted, requiring a blend of technical
considerations. Here we consider three widely-used approaches to aligning LLMs.
The first approach is to fine-tune LLMs with labeled data. This approach is straightforward
as it simply extends the pre-existing training of a pre-trained LLM to adapt it to specific tasks.
An example of this is supervised fine-tuning (SFT), in which the LLM is further trained on a
dataset comprising task-specific instructions paired with their expected outputs. The SFT dataset
is generally much smaller compared to the original training set, but this data is highly specialized.
The result of SFT is that the LLM can learn to execute tasks based on user instructions. For
example, by fine-tuning the LLM with a set of question-answer pairs, the model can respond to
specific questions, even if not directly covered in the SFT dataset. This method proves particularly
155

Alignment

156

Pre-training
Stage

Pre-training

Alignment
Stage

Instruction
Alignment
(e.g., SFT)

Training & Fine-tuning

Human Preference
Alignment
(e.g., RLHF)

Prompting

Inference

Fig. 4.1: Schematic illustration of the pre-train-then-align method for developing LLMs. In the pre-training stage, we
train an LLM on vast amounts of data using next token prediction. Then, in the alignment stage, we align the LLM
to user instructions, intents, and preferences. This includes instruction alignment, human preference alignment, and
prompting.

useful when it is relatively easy to describe the input-output relationships and straightforward to
annotate the data.
The second approach is to fine-tune LLMs using reward models. One difficulty in alignment
is that human values and expectations are complex and hard to describe. In many cases, even for
humans themselves, articulating what is ethically correct or culturally appropriate can be challenging. As a result, collecting or annotating fine-tuning data is not as straightforward as it is with SFT.
Moreover, aligning LLMs is not just a task of fitting data, or in other words, the limited samples
annotated by humans are often insufficient to comprehensively describe these behaviors. What we
really need here is to teach the model how to determine which outputs are more in line with human
preferences, for example, we not only want the outputs to be technically accurate but also to align
with human expectations and values. One idea is to develop a reward model analogous to a human
expert. This reward model would work by rewarding the LLM whenever it generates responses
that align more closely with human preferences, much like how a teacher provides feedback to a
student. To obtain such a reward model, we can train a scoring function from human preference
data. The trained reward model is then used as a guide to adjust and refine the LLM. This frames
the LLM alignment task as a reinforcement learning task. The resulting methods, such as reinforcement learning from human feedback (RLHF), have been demonstrated to be particularly
successful in adapting LLMs to follow the subtleties of human behavior and social norms.
The third approach is to perform alignment during inference rather than during training or
fine-tuning. From this perspective, prompting in LLMs can also be seen as a form of alignment,
but it does not involve training or fine-funing. So we can dynamically adapt an LLM to various
tasks at minimal cost. Another method to do alignment at inference time is to rescore the outputs
of an LLM. For example, we could develop a scoring system to simulate human feedback on the
outputs of the LLM (like a reward model) and prioritize those that receive more positive feedback.
The three methods mentioned above are typically used in sequence once the pre-training is
complete: we first perform SFT, then RLHF, and then prompt the LLM in some way during
inference. This roughly divides the development of LLMs into two stages — the pre-training stage
and the alignment stage. Figure 4.1 shows an illustration of this. Since prompting techniques have
been intensively discussed in the previous chapter, we will focus on fine-tuning-based alignment
methods in the rest of this chapter.

4.2 Instruction Alignment

157

4.2 Instruction Alignment
One feature of LLMs is that they can follow the prompts provided by users to perform various
tasks. In many applications, a prompt consists of a simple instruction and user input, and we want
the LLM to follow this instruction to perform the task correctly. This ability of LLMs is also
called the instruction-following ability. For example, below is a prompt where we want the LLM
to extract key points and provide a concise summary for a lengthy article.
Instruction
Input

Summarize this text in three sentences.
Daylight Savings Time (DST) - the process of moving clocks forward
by one hour in the summer - was started in Germany in 1916 ...

Output

This task requires the LLM to understand the instruction “Summarize this text in three sentences”
and perform the summarization accordingly. However, LLMs are typically trained for next-token
prediction rather than for generating outputs that follow instructions. Applying a pre-trained LLM
to the above example would likely result in the model continuing to write the input article instead
of summarizing the main points. The goal of instruction alignment (or instruction fine-tuning) is
to tune the LLM to accurately respond to user instructions and intentions. The rest of this section
will discuss some issues related to instruction alignment, including fine-tuning LLMs to follow
instructions, generating or collecting instruction data, and generalizing instruction alignment.

4.2.1 Supervised Fine-tuning
One straightforward approach to adapting LLMs to follow instructions is to fine-tune these models
using annotated input-output pairs [Ouyang et al., 2022; Wei et al., 2022a]. Unlike standard language model training, here we do not wish to maximize the probability of generating a complete
sequence, but rather maximizing the probability of generating the rest of the sequence given its
prefix. This approach makes instruction fine-tuning a bit different from pre-training. The SFT
data is a collection of such input-output pairs (denoted by S), where each output is the correct
response for the corresponding input instruction. For example, below is an SFT dataset
x (instruction + user input)
Summarize the following article.
Article: In recent years, solar energy has seen
unprecedented growth, becoming the fastest-growing ...
Extract the main financial figures from the following
earnings report.
Report: The company reported a revenue of $10 million
in the first quarter with a profit margin of 15% ...
Classify the following email as spam or not spam.
Text: Congratulations! You’ve won a $500 gift card.
Click here to claim now.
Provide a solution to the following technical issue.
Issue: my computer is running slow and often freezes.

y (output)
{∗summary∗}
Revenue: $10 million,
Profit Margin: 15%

Spam

First, check for ...

Alignment

158

where the instructions are highlighted. This dataset contains instructions and the corresponding
outputs for several different NLP problems, and so we can fine-tune an LLM to handle multiple
tasks simultaneously.
Let x = x0 ...xm be an input sequence (e.g., instruction + user input) and y = y1 ...yn be the
corresponding output sequence. In SFT, we aim to maximize the probability of the output y given
the input x. Consider an LLM with pre-trained parameters θ̂. The fine-tuning objective can then
be formulated as:
X

θ̃ = arg max
θ̂ +

(4.1)

log Prθ̂+ (y|x)

(x,y)∈D

where θ̃ denotes the parameters optimized via fine-tuning, and θ̂ + represents an adjustment to θ̂.
Here we will omit the superscript + and use θ to represent θ̂ + to keep the notation uncluttered. But
the reader should keep in mind that the fine-tuning starts from the pre-trained parameters rather
than randomly initialized parameters.
The objective function log Prθ (yi |x, y<i ) is computed by summing the log-probabilities of
the tokens in y, conditional on the input x and all the previous tokens y<i :
log Prθ (y|x) =

n
X
i=1

(4.2)

log Prθ (yi |x, y<i )

This formulation is equivalent to minimizing the cross-entropy loss.
Note that minimizing the conditional log-probability log Prθ (y|x) is not a standard language
model training problem. If we concatenate x and y as a single sequence, a more general form of
language modeling is based on the joint log-probability log Prθ (x, y), that is, we minimize the
loss over all tokens of the sequence seq x,y = [x, y]. We can write the probability of this sequence
using the chain rule
log Prθ (seq x,y ) = log Prθ (x, y)
= log Prθ (x) + log Prθ (y|x)
|

{z

set to 0

}

|

{z

}

(4.3)

loss computation

There are two terms on the right-hand side of the equation. We can simply set the first term
log Prθ (x) to 0, focusing solely on the second term log Prθ (y|x) for loss computation. As a
result, the training can be implemented using standard LLMs. For the sequence seqx,y , we first
run the forward pass as usual. Then, during the backward pass, we force the loss corresponding
to x to be zero. Figure 4.2 shows an illustration of this process.
By taking log Prθ (seq x,y ) as the objective function, we can describe SFT using a regular form
of language model training:
θ̃ = arg max
θ

X

log Prθ (seq x,y )

(4.4)

(x,y)∈D

The problem we considered above is fundamentally a single-round prediction problem, where
the LLM generates a response based on a single input without any further interaction or feedback
from the user. The input is processed, and the output is generated in one go. This is typical in

4.2 Instruction Alignment

159
Output

Input

Loss = 0

Loss 6= 0

x1

x2

x3

y1

y2

x1

x2

x3

y1

y2

x0

x1

x2

x3

y1

x0

x1

x2

x3

y1

(b) Backward Pass

(a) Forward Pass

Fig. 4.2: Illustration of supervised fine-tuning for LLMs. We concatenate the input and the output into a single sequence. During the forward pass, we run the LLM as usual. During the backward pass, we compute the loss only for
the output part and simply set the loss for the input part to 0.

scenarios where a single question is asked, and a single answer is provided, with no follow-up
questions or clarifications. However, in practice, we sometimes have to handle multi-round prediction problems, for example, an LLM engages in a dialogue over multiple turns. In this setting,
the LLM not only generates responses based on the initial input but also incorporates subsequent
inputs that might refine or expand on earlier interactions. For example, we can use the LLM to act
as a healthcare assistance chatbot and have a conversation with the user, like this

User
Chatbot

I’ve been feeling very tired lately.
I’m sorry to hear that. Besides feeling tired, have you noticed any
other symptoms?

User

Yes, I’m also experiencing headaches frequently.

Chatbot

How long have these symptoms been going on?

User
Chatbot
User
Chatbot

About a week now.
It might be good to check in with a healthcare professional. Would
you like help setting up an appointment?
Yes, please. Can it be after work hours?
Sure, I can arrange that. There are slots available next Wednesday
and Thursday after 5 PM. Which day works better for you?
...

In this task, there are several rounds of conversation, each involving the generation of a response based on the user’s request or question and the conversational history. Suppose we have
K rounds of conversation, denoted by {x1 , y1 , x2 , y2 , ..., xK , yK }. Here xk and yk denote the
user request and the response, respectively, for each round k. The log-probability of generating
the response can be written as log Prθ (yk |x1 , y1 , ..., xk ). Our goal is then to maximize the sum
of these log-probabilities
θ̃ = arg max
θ

K
X

k=1

log Prθ (yk |x1 , y1 , ..., xk )

(4.5)

Alignment

160

A straightforward implementation of this involves calculating the conditional probability for
each k. However, it requires running the LLM K times, each time with an increased conversational history to make predictions. A more efficient method is to perform loss computation of all
responses in a single run of the LLM. To do this, we represent the conversation as a sequence
seqx1 ,y1 ,...,xK ,yK = [x1 , y1 , ..., xK , yK ] (or seq for short). The log-probability of this sequence
is given by
log Prθ (seq) = log Prθ (x1 , y1 , ..., xK , yK )
= log Prθ (x1 ) + log Prθ (y1 |x1 ) + · · · +
|

{z

set to 0

|

}

{z

loss computation

}

log Prθ (xK |x1 , y1 , ..., yK−1 ) +
|

{z

set to 0
K 1 1

log Prθ (y |x , y , ..., xK )
|

{z

loss computation

}

}

(4.6)

The trick here is that we ignore the loss for generating user inputs, as illustrated in Figure 4.3.
Hence we only compute the probabilities of generating the responses given their conversational
histories, in other words, the value on the right-hand side of Eq. (4.6) is actually equal to the value
on the right-hand side of Eq. (4.5). As with Eq. (4.4), the training of this multi-round prediction
model can be achieved by maximizing the log likelihood over a training dataset D:
θ̃ = arg max
θ

X

log Prθ (seq)

(4.7)

seq∈D

While implementing the SFT methods introduced above seems trivial as they are fundamentally the same as regular language model training, there are still issues that need to be considered
in practice. For example,
• SFT requires labeled data. This makes SFT quite different from pre-training, where raw text
is used as training data and is readily available. As in other supervised machine learning
problems, data annotation and selection in SFT are not simple tasks. In general, we wish
to develop SFT data that is both substantial in quantity and high in quality, and this data
should be highly relevant to the tasks the LLM will perform. On the other hand, there is
a need to fine-tune LLMs with less data to minimize computational and data construction
costs. Often, the quality of LLMs is highly dependent on the data used in SFT. Thus, such
data must be carefully developed and examined. As we will see in later subsections, SFT
can be more efficient and effective through more advanced techniques for data construction.
• SFT is still computationally expensive for LLMs due to their big size. As a result, maintaining and updating such models is resource-intensive. For example, applying gradient updates to billions of parameters within an LLM requires significant computational power and
memory. This often requires high-performance computing environments, which are costly
to operate. To address these challenges, various optimization strategies, such as pruning,
quantization, and the use of more efficient training algorithms, have been explored. In particular, there has been significant interest in parameter-efficient fine-tuning methods which
are designed to maintain state-of-the-art performance without the need for extensive computational resources. We have seen in Chapter 3 that applying techniques like soft prompts can

4.2 Instruction Alignment

161

Loss = 0

Loss 6= 0

Loss = 0

Loss 6= 0

Prθ (x1 )

Prθ (y1 |x1 )

Prθ (x2 |x1 , y1 )

Prθ (y2 |x1 , y1 , x2 )

x1

y1

x2

y2

···

User: I’ve been feeling very tired lately.

Chatbot: I’m sorry to hear that. Besides feeling tired,
have you noticed any other symptoms?

User: Yes, I’m also experiencing headaches frequently.

Chatbot: How long have these symptoms been going on?

..
.
Fig. 4.3: Illustration of supervised fine-tuning for conversational models. Here the LLM acts as a chatbot to respond to
each request based on the conversational history. The conversation progresses by alternating between the user and the
chatbot. In SFT, we treat the entire conversation as a sequence, just like in standard LLMs, but compute the loss only
for the responses of the LLM.

make the fine-tuning process more efficient. For further discussion on parameter-efficient
methods, the reader can refer to related papers on this issue [Houlsby et al., 2019; Hu et al.,
2022; Han et al., 2024].
• SFT can be regarded as a post-training step following pre-training. It is a separate training
phase designed to preserve the advantages of the initial pre-training while incorporating new
adjustments. This may seem paradoxical because updating a pre-trained LLM with further
data potentially causes the model to forget some of its prior knowledge. Imagine a scenario
where we have a large amount of SFT data and extensively fine-tune the LLM. In this
case, the LLM could overfit the data, which in turn may reduce generalization performance
or cause catastrophic forgetting. A common strategy to mitigate this issue is to employ
regularization and early stopping techniques. Another practical approach is to use a smaller
learning rate to gently adjust the weights of the LLM. In addition, fine-tuning with data from
diverse sources and problem domains can also be beneficial. Nevertheless, in practice, the
SFT step is often carefully examined and requires substantial engineering and experimental
efforts to optimize.

4.2.2 Fine-tuning Data Acquisition
Fine-tuning data is so important that much recent work in LLM has focused on developing various
datasets for instruction fine-tuning. As with most work in machine learning, there are generally
two approaches to data acquisition — manual data generation and automatic data generation.

Alignment

162
4.2.2.1 Manually Generated Data

One straightforward method is to recruit human annotators to create input-output pairs for the
tasks of interest. Unlike data annotation in conventional NLP, such as text classification, where
annotators simply assign labels to collected texts according to guidelines, creating fine-tuning data
for LLMs requires more steps and effort, making it thus more challenging. Suppose we want to
obtain fine-tuning data for the English-to-Chinese machine translation task. The first step is to
write a prompt template to describe the task and format the problem clearly. For example,
Instruction

Translate the text from English to Chinese.

User Input

{∗text∗}

Output

{∗translation∗}

Then, we collect pairs of source and target texts (i.e., Chinese texts and the corresponding
translations), and replace the variables {∗text∗} and {∗translation∗} to generate the fine-tuning
samples. For example, given a pair of English and Chinese sentences
How’s the weather today?
{∗text∗}

→

今天天气怎么样？
{∗translation∗}

we can generate a fine-tuning sample using the prompt template, like this
Instruction

Translate the text from English to Chinese.

User Input

How’s the weather today?

Output

今天天气怎么样？

That is,
x = Translate the text from English to Chinese.\n How’s the weather today?
y =

今天天气怎么样？

We can use this (x, y) pair to fine-tune the LLM, as described in the previous subsection.
One difficulty here is that there are many, many different ways to write prompt templates
for the same task, and different people may produce prompt templates with varying qualities
and complexities. Sometimes, we may write prompt templates with overly complex or verbose
instructions. Sometimes, we may not even know exactly what the target task is and how to describe it. A widely-adopted strategy is to create prompt templates for existing NLP tasks, given
that there have been so many well-established NLP problems and benchmarks [Bach et al., 2022;
Wang et al., 2022b; Mishra et al., 2022]. In this case, annotators can be given the original task
description and many examples. Then, they can use their own ways to express how to prompt the
LLM to perform the tasks. Note that, while such a method can ease the process of creating and
writing prompts, we still need annotation frameworks and crowdsourcing systems to manage the
work and conduct quality control. For example, we generally need to design annotation guidelines

4.2 Instruction Alignment

163

and a unified format for writing prompt templates, especially when many annotators are contributing to the same task. One advantage of inducing prompts from existing NLP tasks is that, once the
prompt templates have been developed, it is easy to generate prompts using the annotated samples
in the original tasks. For example, given a bilingual dataset for English-to-Chinese translation, we
can easily create a number of fine-tuning examples by filling the slots in the above template with
the sentence pairs in this dataset.
Another approach is to directly use the naturally existing data available on the internet. A
common example is by collecting question-and-answer pairs from QA websites to fine-tune LLMs
for open-domain QA tasks [Joshi et al., 2017]. Many benchmarks in QA are built in this way
because there are so many types of questions that it is impossible to think of them all by a small
group of people. Instead, using data from those websites can ensure that the LLM fine-tuning data
is at a good or acceptable level in terms of quantity and quality.
In addition to employing existing resources, another straightforward way to develop a finetuning dataset is to crowdsource the data. A simple approach is to allow users to input any question, after which responses are either manually given or automatically generated by an LLM and
then manually annotated and corrected. It is thus possible to capture real user behavior and consequently gather inputs and outputs for a large number of “new” problems that traditional NLP tasks
do not cover.
An issue related to the construction of the fine-tuning datasets is that we usually want the
data to be as diverse as possible. Many studies have found that increasing the diversity of finetuning data can improve the robustness and generalization ability of LLMs. For this reason, there
has been considerable interest in involving more diverse prompts and tasks in LLM fine-tunining
datasets. We will give further discussions on the generalization of fine-tuning in Section 4.2.4.
4.2.2.2 Automatically Generated Data
One limitation of manual data generation is that the quality and diversity largely depend on human
experience and creativity. Therefore, if we want LLMs to handle a broad range of tasks, that
is, to effectively execute any instruction, relying on human-annotated data for LLM fine-tuning
is often inefficient. Moreover, the coverage of such data can be limited, and the data may even
contain biases introduced by the annotators themselves. An alternative approach is to generate data
automatically. For example, we can collect a number of questions through crowdsourcing, and
employ a well-tuned LLM to generate answers to the questions. These question-answer pairs are
then used as fine-tuning samples as usual. This method, though very simple, has been extensively
applied to generate large-scale fine-tuning data for LLMs.
The above way of producing synthetic fine-tuning data is similar to those used in data augmentation for NLP. If we have an LLM, we can produce a prediction in response to any input.
Repeating this process for different inputs allows us to create a sufficient number of fine-tuning
samples. Such a method is particularly useful for fine-tuning new LLMs using a well-tuned LLM.
However, one disadvantage of this approach is that it relies on human-crafted or collected inputs
for data generation, which may turn out to be inappropriate for generalizing LLMs. In many LLM
applications, a significant challenge arises from the broad range of users’ questions and requests,
many of which are not covered in existing NLP tasks and datasets. In these cases, it becomes
necessary to generate not only the predictions but also the inputs themselves.
Here we consider self-instruct as an example to illustrate how to generate LLM fine-tuning

Alignment

164

Initialization

Initialize the task pool with a number of instructions
and corresponding input-output samples.
Sample 1: (Instruction, User-input, Output)
Sample 2: (Instruction, User-input, Output)

Task Pool

···

Draw a few instructions from the pool

Sampling

Instructiona
Instructionb

sampling

Task Pool

Instructionc

Instruction
Generation

Sample
Generation

Filtering

Prompt the LLM to generate a new instruction based on
the drawn instructions.
You are provided several different instructions for performing
some tasks. Please generate an instruction based on these.
Task 1: Instructiona
Task 2: Instructionb
Task 3: Instructionc
New Task: Instructionnew

Given the newly-generated instruction and a few
input-output samples, generate a new sample.
You are provided with a set of input-output samples tasks,
each composed of an instruction, a user input, and an output.
Please generate a new sample based on these.
Sample 1: Samplea
Sample 2: Sampleb
New Sample: Instructionnew User-inputnew Outputnew

Filter out invalid and low-quality samples.
Add the remaining samples into the pool.

Fig. 4.4: Illustration of self-instruct [Wang et al., 2023b]. This method maintains a pool of instructions and corresponding input-output samples. Initially, the pool contains a number of hand-crafted instructions and samples. Each time,
we draw a few instructions from the pool. An LLM is then prompted to generate new instructions and samples based
on those drawn. Finally, the newly-generated instructions and samples are filtered and added to the pool.

samples [Wang et al., 2023d; Honovich et al., 2023]. The idea is that we can prompt an LLM to
create a new instruction by learning from other instructions. Given this instruction, the LLM can
then fill in other fields (such as the user input) and produce the predictions. Figure 4.4 shows a
schematic illustration of self-instruct. Here we give a brief outline of the key steps involved.

• The self-instruct algorithm maintains a pool of tasks. Initially it contains a number of seed
hand-crafted tasks, each with an instruction and input-output sample. As the algorithm
proceeds, LLM-generated instructions and samples will be added to this pool.

4.2 Instruction Alignment

165

• At each step, a small number of instructions are drawn from the instruction pool. For example, we can randomly select a few human-written instructions and a few LLM-generated
instructions to ensure diversity.
• The selected instructions are then used as demonstration examples. Thus, the LLM can
in-context learn from these examples and produce a new instruction. Below is an example
template for prompting the LLM.
You are provided several different instructions for performing some tasks.
Please generate an instruction based on these.
Task 1: {instruction1}

Task 2: {instruction2}
Task 3: {instruction3}
Task 4: {instruction4}

New Task:

• Given the generated instruction, the LLM is then prompted to complete the sample by filling
in the remaining input fields and generating the corresponding output. Below is a prompt
template.
You are provided with a set of input-output samples, each composed of
an instruction, a user input, and an output. Please generate a new sample
based on these.
Sample 1: {instruction1}
Input: {user-input1}
Output: {output1}

Sample 2: {instruction2}
Input: {user-input2}
Output: {output2}

New Sample: {new-instruction}

• This newly-generated sample is examined by some heuristic rules (such as filtering out
samples or instructions that are similar to those already in the pool). If it passes, the sample
and instruction are added to the pool.
This generation process can be repeated many times to obtain a sufficient number of finetuning samples. Note that, above, we just show simple prompt templates for generating instruction
and fine-tuning samples. Of course, we can develop better templates to generate more diverse and
accurate instruction and fine-tuning samples. For example, for certain tasks like text classification,

166

Alignment

the LLM may tend to produce biased predictions, for example, most generated samples belong to
a single class. In such cases, we can adjust the order of generation of different fields. More
specifically, we can specify the output (i.e., the class) with some prior, and prompt the LLM
to generate user input given both the instruction and the output. This method resembles input
inversion, where the LLM generates the input based on the specified output [Longpre et al., 2023].
Using LLM-generated instructions and fine-tuning samples has been a common method for
developing LLMs, especially given that manually developing such data is so expensive that most
research groups cannot afford it. In several well-tuned LLMs, their fine-tuning datasets include
a certain amount of synthetic data, which has proved useful [Ouyang et al., 2022; Taori et al.,
2023; Chiang et al., 2023]. There have been further studies on synthetic data generation for LLM
fine-tuning. For example, one can generate more diverse instructions by introducing evolutionary
algorithms [Xu et al., 2024], or use synthetic data as supervision signals in a more advanced finetuning process [Chen et al., 2024b]. More recently, there has also been considerable interest in
using synthetic data in the pre-training stage [Gunasekar et al., 2023; Allal et al., 2024].
In many applications, a real-world scenario is that, given a task, we can collect or annotate a
relatively small amount of fine-tuning data, for example, we can recruit experts to create questions
for QA tasks in a specific domain. But the quantity and diversity of this data are in general not
sufficient. In this case, we can use self-instruct techniques to generate more diverse questionanswer pairs, and thus augment the fine-tuning data. This provides a way of bootstrapping the
LLM starting from a seed set of fine-tuning samples. Note that using self-generated data is a common practice and has long been applied in NLP. For example, this approach has been successfully
used in parsing and machine translation [Charniak, 1997; Sennrich et al., 2016].

4.2.3 Fine-tuning with Less Data
With the increasing prominence of instruction fine-tuning, there has been a surge in demand for
large-scale, high-quality fine-tuning data. For example, the FLAN fine-tuning dataset, which
is compiled from 1,836 tasks, contains 15 million samples [Longpre et al., 2023]. Fine-tuning
LLMs with such large datasets is typically a computationally expensive task, especially given
that updating the large number of parameters in LLMs is resource-intensive. One approach for
mitigating this issue is to explore efficient model training methods, for example, one can use
parameter-efficient methods to update only a small portion of the model. However, many finetuning datasets contain a large amount of synthetic data, where errors and biases are still inevitable.
Another approach to efficient fine-tuning is to consider only the most relevant and impactful
examples for fine-tuning. We can thus reduce the amount of data that needs to be processed while
still maintaining the quality of the model updates. There are several methods to achieve this. For
example, Zhou et al. [2023a] built an instruction-following dataset containing only 1,000 samples by carefully crafting the prompts and collecting samples from a variety of NLP tasks. They
showed that the LLaMa 65B model fine-tuned with this dataset could be competitive with or even
better than models with much more fine-tuning effort. This suggests that LLMs can be adapted
to respond to diverse tasks without necessarily needing fine-tuning on all types of instructionfollowing data. Chen et al. [2024a] developed a system based on the GPT-3.5 model to assess
the quality of each instruction-following sample. Therefore, they could select high-quality samples from existing datasets, showing better fine-tuning performance with fewer fine-tuning samples. Researchers have also developed methods to either select or filter out data using heuristics

4.2 Instruction Alignment

167

[Zhao et al., 2024; Ge et al., 2024], or to prioritize data that more significantly influences the finetuning process [Xia et al., 2024]. In fact, most of these methods can be seen as instances of larger
families of data selection and filtering methods. And it is often the case that using higher quality
(but maybe less) data is beneficial for training NLP models.
The discoveries in instruction fine-tuning somewhat differ from traditional views in NLP: the
ability of models to handle complex problems can be activated with a small amount of annotated
data, rather than requiring massive amounts of supervised data for extensive training. One possible explanation is that the ability of generating correct responses given instructions has been
learned during pre-training, but such instruction-response mappings are not with high probabilities during inference. Fine-tuning can slightly adjust the models to get them to follow instructions, requiring significantly less training effort than pre-training. This is closely related to what
is known as the superficial alignment hypothesis, which suggests that learning primarily occurs
during pre-training, and the subsequent fine-tuning or alignment phase does not significantly contribute to the underlying knowledge base of an LLM [Zhou et al., 2023a]. Since the core abilities
and knowledge of the model are already established from pre-training, effective fine-tuning for
alignment with user needs can be achieved with relatively small training fine-tuning effort. This
implies the possibility of fine-tuning LLMs with very little data. In another direction, it may not
be necessary to restrict fine-tuning to paired instruction-response data. For example, Hewitt et al.
[2024] found that instruction-following can be implicitly achieved by fine-tuning LLMs only on
responses, without corresponding instructions.
A concept related to the discussion here is sample efficiency. A machine learning method is
called sample efficient if it can learn effectively from a small number of training examples. In this
sense, instruction fine-tuning is sample efficient compared with pre-training. From the perspective
of machine learning, sample-efficient methods can be seen as efficient ways to sample the space
of data, and are advantageous as they make optimal use of scarce data. Therefore, sampling-based
learning techniques, such as many reinforcement learning algorithms, can benefit from these sample efficient approaches. For example, in human preference alignment, we can either efficiently
sample preference data via reward models [Liu et al., 2024b] or improve sampling efficiency in
policy learning [Wang et al., 2024].

4.2.4 Instruction Generalization
In many machine learning and NLP problems, training a model to generalize is a fundamental
goal. For example, in text classification, we expect our model to correctly classify new texts that
were not seen during training. However, generalization poses additional challenges in instruction
fine-tuning. We expect instruction-fine-tuned LLMs to not only generate appropriate responses for
different inputs within a task but also to accurately perform various tasks as described by different
instructions. To illustrate this issue, consider an LLM Pr(y|c, z), where c is an instruction, z
is a user input, and y is the corresponding model output (i.e., the response). Suppose that the
performance of this model is evaluated in terms of a metric, written as Performance(Pr(y|c, z))
or P(c, z, y) for short. Informally, when we say this model can generalize within a given task
(indicated by the instruction c∗ ), we mean that there may be a value ǫ such that the average
performance on new inputs is above this value:
1 X
P(c∗ , z′ , y′ ) > ǫ
|Z| z′ ∈Z

(4.8)

Alignment

168

where Z is the set of new inputs, and z′ and y′ are an input in this set and the corresponding
output, respectively.
Likewise, we can say that this model can generalize across tasks if the average performance
over all instruction-input pairs is above some ǫ:
X
1
P(c′ , z′ , y′ ) > ǫ
|D| (c′ ,z′ )∈D

(4.9)

where D is the set of new instruction-input pairs.

Here, we need to deal with variations in two dimensions: instruction and user input. This
makes the generalization problem very complex, because, intuitively, a model needs to learn from
a vast number of tasks and different input-output pairs associated with each task to achieve good
generalization. As we have discussed several times in this book, achieving such generalization
incurs much lower cost than pre-training. In general, fine-tuning LLMs with instruction-response
data to some extent can lead to models yielding instruction following on new tasks. Nevertheless,
it is typically believed that certain efforts are still needed to adapt LLMs to make them understand
and execute instructions broadly.
One way to generalize instruction fine-tuning is to increase the diversity of the fine-tuning
data. In earlier studies on instruction fine-tuning, researchers developed many datasets, covering a
wide variety of NLP tasks and different instructions for each task [Wang et al., 2022b; Sanh et al.,
2022; Longpre et al., 2023]. By transforming these tasks into a unified format, one can finetune an LLM with a sufficiently large number of samples, for example, there have been several
instruction fine-tuning datasets that involve over 100 NLP tasks and 1M samples. However, these
early datasets mostly focus on existing academic problems, but not those that users want to deal
with in real-world applications. Much recent work has shifted focus to addressing new and more
practical problems. For example, there has been considerable interest in constructing datasets
that contain large and complicated demonstrations and responses from SOTA models to real user
queries [Wang et al., 2023c; Teknium, 2023].
Perhaps the use of large and diverse fine-tuning datasets has its origins in attempts to scale
LLMs in different dimensions. Indeed, scaling laws have been used broadly to motivate the development of a wide range of different instruction-fine-tuned LLMs. And it is reasonable to scale
instruction fine-tuning to make an LLM follow broad instructions. From the perspective of LLM
alignment, however, scaling instruction fine-tuning might not be efficient to achieve generalization.
One problem is that instruction fine-tuning relies on supervised learning that learns to generalize and perform tasks based on instruction-response mappings. However, such an approach does
not capture subtle or complex human preferences (e.g., tone, style, or subjective quality) because
these are hard to encode as explicit instruction-response data. Moreover, the generalization performance is bounded by the diversity and quality of the instruction-response dataset. Given these
limitations, we would instead like to employ preference models as an additional fine-tuning step
following instruction fine-tuning, so the LLMs can generalize further (see Section 4.3).
Another view is that some instruction-response mappings may already be learned during pretraining, and so the pre-trained LLMs have encoded such mappings. However, since we often do
not know exactly what data is used in the pre-training, it is hard to judge whether we need to learn
such mappings in the fine-tuning. A related question is whether out-of-distribution generalization

4.2 Instruction Alignment

169

is primarily achieved during pre-training or fine-tuning. While directly answering this question is
beyond the scope of this chapter, it has been shown that pre-training on large and diverse datasets
is effective in improving out-of-distribution performance [Hendrycks et al., 2020; Radford et al.,
2021; Gunasekar et al., 2023]. This raises an interesting problem: if an LLM has been well pretrained at scale, fine-tuning may not be as essential for out-of-distribution generalization, since the
model may have already encountered sufficient distributional variation. This prompts researchers
to fine-tune LLMs with modest effort or to explore new methods to achieve instruction-following.
As discussed in the previous sub-section, for example, instruction following can be yielded by
fine-tuning on a small number of carefully selected instruction-response pairs [Zhou et al., 2023a],
or even by using methods that are not explicitly designed to do so [Kung and Peng, 2023].
The above discussion provides two different strategies: one requires scaling up fine-tuning
datasets for larger diversity, the other requires small but necessary fine-tuning datasets for efficient
LLM adaptation. However, in practice, involving diverse instructions often helps. In many cases,
we need to adapt our LLM for specific purposes. But the LLM, which has possibly encoded broad
instruction-following mappings during pre-training, might tend to behave as a general-purpose
instruction executor even with modest fine-tuning. An interesting phenomenon is that when finetuning on math data, the resulting LLM might not specialize in math outputs. Instead, this model
might respond normally to general instructions, for example, it could generate poetry if instructed
to do so [Hewitt, 2024]. This is not a bad thing, but it shows that LLMs may not easily change their
nature of following general instructions. In this case, additional adaptations with more diverse
data may help adjust the way the LLM follows instructions, particularly for those tasks we aim to
address.

4.2.5 Using Weak Models to Improve Strong Models
So far we have explored a variety of instruction fine-tuning methods based on labeled data. One
of the limitations of many such methods is that they require the data to be annotated by humans or
generated by strong LLMs, which can provide accurate supervision signals in fine-tuning. However, in many cases, the LLM we have in hand is already strong (or at least is advantageous in
specific aspects of problem solving), and thus it is not easy to find a superior model for supervision. Even for human experts, when the problem becomes complex, providing correct and detailed
answers might be difficult, or sometimes infeasible. For example, when faced with an extremely
long document, the experts would find it challenging to identify any inconsistencies, subtle biases,
or missing key points without conducting an exhaustive and time-consuming review.
One may ask at this point: can we use weak LLMs to supervise strong LLMs? This seems
to be a significant challenge, but it may reflect a future scenario where we need to supervise AI
systems that are smarter than humans or any other AI systems [Burns et al., 2023b]. The problem
of using smaller, less complex models to improve the training of larger, more complex models
is also called the weak-to-strong generalization problem. While there have not been mature
approaches to weak-to-strong generalization, using smaller models to assist stronger models has
indeed proven useful in several areas of LLMs.
For instruction fine-tuning, one of the simplest ways of applying weak LLMs is to use these
models to generate synthetic fine-tuning data. Suppose we have a collection of inputs X, where
each input includes an instruction and a user input if necessary. For each x ∈ X, we use a weak
LLM Prw (·) to generate a prediction ŷ = arg maxy Prw (y|x). Then, the strong LLM Prsθ (·) can

Alignment

170
be trained on these generated predictions (see Eq. (4.1)):
θ̃ = arg max
θ

X

log Prsθ (ŷ|x)

(4.10)

x∈X

where θ is the model parameters.
The above form transforms the fine-tuning problem into a knowledge distillation problem, in
other words, we distill knowledge from the weak model to the strong model. Consequently, we
can employ various knowledge distillation methods to achieve this goal. However, explaining
weak-to-strong fine-tuning from the perspective of knowledge distillation is not straightforward.
A major concern is that the strong model may merely imitate or overfit the errors of the weak
model and fail to generalize. For example, the fine-tuned strong model still cannot solve difficult
problems that the weak model cannot accurately predict. Fortunately, preliminary experiments in
this line of research have shown positive and promising results. For example, Burns et al. [2023a]
found that fine-tuning the strong pre-trained GPT-4 model with GPT-2-level supervision could
improve generalization across several NLP tasks. To measure how the weak model improves the
generalization of the strong model, we define the following terms:
• Weak Performance (Pweak ). This is the test-set performance of the weak model, which
can be regarded as the baseline performance.
• Weak-to-strong Performance(Pweak→strong ). This is the test-set performance of the strong
model that is fine-tuned with the weak model.
• Strong Ceiling Performance (Pceiling ). This is the test-set performance of the strong model
that is fine-tuned with ground truth data. For example, we fine-tune the strong model with
human-annotated predictions and take the resulting model as a ceiling.
Then, the performance gap recovered (PGR) can be defined as
n

PGR = max 0,

Pweak→strong − Pweak o
Pceiling − Pweak

(4.11)

This metric measures how much of the performance gap between the ceiling model and the
weak model can be recovered by the weak-to-strong model. A PGR of 1 indicates that the weakto-strong fine-tuning can completely closes the performance gap, whereas a PGR of 0 indicates
no improvement. In Burns et al. [2023a]’s work, it is shown that PGR can be around 0.8 on 22
NLP classification tasks. It should be noted that, while the potential of weak-to-strong fine-tuning
is promising, achieving substantial weak-to-strong generalization remains a challenging goal that
needs further investigation [Aschenbrenner, 2024].
Fine-tuning LLMs with weak supervision is just one choice for using small models to improve
large models. Although this section primarily focuses on fine-tuning LLMs, we also mention
other methods here to give a more complete discussion (see Figure 4.5 for illustrations of these
methods).
• Instead of using small models to generate synthetic data, it is also straightforward to incorporate knowledge distillation loss based on these models. For example, a simple loss

4.2 Instruction Alignment

171

Compute Loss & Train

Compute Loss & Train

ŷ

y

Small Model
Input

Predict

Dataset
x ŷ

LM Loss

KD Loss

Small Model

Large Model

Large Model

Dataset

=⇒

=⇒

x y
x

x

(a) Fine-tuning on data generated by a small model
(weak-to-strong generalization)

(b) Fine-tuning with KD Loss from a small model
(weak-to-strong generalization)

Dataset
y
Compute Loss & Train

Small Model

Combination Model

y

Data
Selection

Large Model

Dataset
x y

=⇒

Small Model 1

Small Model 2

Small Model 3

x

x

x

x

(d) Ensemble of multiple small models

(c) Data selection with a small model

If Step 1 is not satisfactory, go to Step 2

Step 1

Step 2

(cheap)

(expensive)

y2

y1

Large Model

Small Model
x

x

(e) Cascading (at inference time)

Fig. 4.5: Illustrations of using small models to improve large models in LLMs. One approach involves using smaller
models for the fine-tuning or pre-training of larger models. This includes generating synthetic data (a), incorporating
auxiliary loss (b), and selecting appropriate data (c). Another approach involves combining small models and large
models. This includes learning a strong model by aggregating multiple small models (d), and cascading small models
with large models (e).

function that measures the difference between the small and large models can be defined as:
Losskd = KL(Prw (·|x) || Prsθ (·|x))

(4.12)

Then, we can add this loss to the original loss of language modeling, and yield the following
training objective
θ̃ = arg max
θ

X

(x,y)∈D

log Prsθ (y|x) − λ · Losskd

(4.13)

Alignment

172

where D is the set of input and output pairs, and λ is the coefficient of the interpolation. This
method can be employed in either the pre-training or fine-tuning phase. We can adjust λ to
control how much the small model influences the training. For example, we can gradually
decrease λ to make the training rely more on the original language modeling loss as the
large model becomes more capable.
• Another approach to involving small models in LLM pre-training and fine-tuning is to use
them to do data selection or filtering. Given a sequence, we can compute the likelihood
or cross-entropy using a small model. These quantities can then be used as criteria for
selecting or filtering data. For example, sequences with low likelihood or high cross-entropy
might be excluded from the training set, as they are less aligned with the small model’s
learned distribution. Conversely, sequences with high likelihood or low cross-entropy can
be prioritized, ensuring that the training focuses on more relevant or high-quality data.
• Ensemble learning is a simple and effective way to build a strong model by combining multiple weak models. Applying this technique to LLMs is straightforward. We can aggregate
distributions predicted by multiple small models or specialized submodels, and derive the
final prediction from the aggregated results. This aggregation can be done using methods
such as majority voting, weighted averaging, or stacking.
• Small models can also be employed at inference time to improve overall efficiency. Suppose
we have a large model that is slow but more accurate, and a small model that is fast but
less accurate. In model cascading, the small model first processes the input data, quickly
generating preliminary results. If these results meet certain pre-defined criteria, they can be
directly used. However, if the initial results are not sufficiently good, the input is then passed
to the larger, more accurate model to produce a better result. This approach significantly
reduces computational costs and latency, as the small model can effectively handle many
inputs without access to the large model.

4.3 Human Preference Alignment: RLHF
So far in this chapter, we have focused on fine-tuning LLMs using input-output paired labeled data.
This approach allows us to adapt LLMs for instruction-following via supervised learning. In many
applications, however, LLMs are required not only to follow instructions but also to act in ways
that are more aligned with human values and preferences. Consider a scenario where a user asks an
LLM how to hack into a computer system. If the LLM is not appropriately aligned, it may respond
by providing details on how to perform this illegal activity. Instead, a more desirable response
might be to advise the user against engaging in illegal activities and offer a general overview of
the consequences of such actions. The difficulty in achieving this is that the ethical nuances and
contextual considerations required for an LLM to respond appropriately in such scenarios are not
always straightforward to encode into a fine-tuning dataset. What’s even more challenging is that,
often, humans themselves cannot precisely express their own preferences.
In this section, we discuss an alternative LLM fine-tuning method, called reinforcement learning from human feedback or RLHF for short [Christiano et al., 2017; Stiennon et al., 2020]. The
basic idea behind RLHF is that LLMs can learn from comparisons of model outputs using reward

4.3 Human Preference Alignment: RLHF

173

y

Objective (MLE):

Gold-standard Predictions

max Pr(y|x)

(n one-hot distributions)

LLM

Predicted Token Distributions

where
x: input
y: gold-standard output

(n token distributions)
x

(a) Supervised fine-tuning (maximizing the prediction probability given the input)

Generate multiple
outputs via sampling

Human preference data

train

Reward Model

yy21

Objective (RL Loss Minimization):

LLM

Prediction y1
Prediction y2

x

min L(x, {y1 , y2 }, r)
where
L(·): loss function
r(·): reward model

(b) Reinforcement Learning from Human Feedback

Fig. 4.6: Supervised fine-tuning vs. reinforcement learning from human feedback. In supervised fine-tuning, we
optimize the LLM by maximizing the probability of the prediction given the input. In reinforcement learning from
human feedback, we first train a reward model on human preference data (on each pair of predictions, evaluators are
asked to choose which one they prefer). Then, we use this reward model to supervise the LLM during fine-tuning.

models (see Figure 4.6). To do this, we can recruit human experts who indicate their preferences
between pairs of outputs generated by the LLM. This preference data is used to train a reward
model that can predict the perceived quality of LLM outputs. Once trained, the reward model
provides feedback by assigning scores to new outputs that the LLM generates in response to the
inputs. The LLM uses these scores to update its parameters through reinforcement learning algorithms. In the rest of this section, we will first introduce the basic knowledge of reinforcement
learning to facilitate the discussion, and then discuss methods for training reward models and
aligning LLMs with these models.

4.3.1 Basics of Reinforcement Learning
We begin by looking at some basic concepts of reinforcement learning. Note that the notation used
here slightly differs from that used in the previous sections and chapters because we want to make
our description more consistent with those in the reinforcement learning literature. Nevertheless,
we will show how this notation corresponds to the language modeling notation. The reader who
is already familiar with reinforcement learning techniques may skip or skim this subsection.
A general reinforcement learning framework describes how an agent interacts with a dynamic
environment. This interaction is modeled as a sequence of actions taken by the agent in response
to the state of the environment. At each time step, the agent observes the current state, chooses an
action based on its policy, performs the action, and then receives feedback from the environment
in the form of a reward and a new state. This sequence of observe-act-receive feedback is repeated

Alignment

174
until the agent achieves its goal.
A reinforcement learning system involves several components

• Agent. This is the learner or decision-maker in reinforcement learning. In the context of
LLMs, it can be seen as the LLM itself.
• Environment. This includes everything external to the agent with which the agent interacts.
But the environment in LLMs is less about a physical or virtual space and more about the
framework within which the agent (e.g., an LLM) receives feedback and learns.
• State (s). A state represents the current situation of the environment. Given a sequence of
tokens for language modeling, a state at a time step can be viewed as the tokens we observed
so far, that is, the context tokens we take to predict the next token. For example, we can
define (x, y<t ) as the state when predicting the next token at the time step t.
• Action (a). Actions represent possible decisions the agent can make. We can see them as
possible predicted tokens in the vocabulary.
• Reward (R). The reward is the feedback from the environment that evaluates the success
of an action. For example, r(s, a, s′ ) denotes the reward the agent receives for taking the
action a at the state s and moving to the next state s′ . If the state-action sequence is given,
we can denote the reward at the time step t as rt = r(st , at , st+1 ). Also note that if the
decision-making process is deterministic, we can omit st+1 because it is can be determined
by st and at . In such cases, we can use r(st , at ) as shorthand for r(st , at , st+1 ).
• Policy (π). For an LLM, a policy is defined as the probability distribution over the tokens
that the LLM predicts, given the preceding context tokens. Formally, this can be expressed
as
π(a|s) = Pr(yt |x, y<t )

(4.14)

where a corresponds to the token y, and s corresponds to the context (x, y<t ). Figure 4.7
illustrates how an LLM can be treated as a policy in the reinforcement learning framework.
• Value Function (V and Q). A state-value function (or value function, for short) assesses
the expected discounted return (i.e., accumulated rewards) for an agent starting from a particular state s and following a specific policy π. It is defined as:
h

V (s) = E r(s0 , a0 , s1 ) + γr(s1 , a1 , s2 ) + γ 2 r(s2 , a2 , s3 ) + · · · s0 = s, π
h

= E r0 + γr1 + γ 2 r2 + · · · s0 = s, π
= E

∞
hX
t=0

γ t rt s0 = s, π

i

i

i

(4.15)

where γ ∈ [0, 1] is the discount factor that adjusts the importance of future rewards, s0 = s
indicates that the agent starts with the state s, and the expectation E is performed over all
possible trajectories (i.e., state-action sequences). Similarly, an action-value function (or

4.3 Human Preference Alignment: RLHF

175
Feedback

Value Functions

Action at
y1

y2

...

yt

V ( st ) and Q( st , at )

yt−1

Reward Model

Policy (LLM)
x0

x1

...

xm

y1

...

R( st , at )

State st (x and y<t )

Fig. 4.7: LLM as policy in reinforcement learning. At each step t, the LLM predicts a token yt given the model
input x and the previously-generated tokens y<t . This process can be framed as a reinforcement learning problem,
where yt serves as the action, (x, y<t ) as the state, and the predicted distribution Pr(yt |x, y<t ) as the policy. Once
yt is predicted, the LLM inputs both (x, y<t ) and yt to the reward model, which evaluates how effectively the chosen
token contributes to achieving the desired textual outcome. This evaluation generates reward scores which are used to
compute the value functions V (st ) and Q(st , at ). These functions then provide feedback to the LLM and guide the
policy training.

Q-value function) measures the expected return starting from a state s taking an action a
and thereafter following a policy π, given by
Q(s, a) = E

∞
hX

γ t rt s0 = s, a0 = a, π

t=0

i

(4.16)

where a0 = a indicates that the action taken at the initial state is a.
The goal of reinforcement learning is to learn a policy that maximizes the cumulative reward (or return) the agent receives over the long run. Given a state-action sequence τ =
{(s1 , a1 ), ..., (sT , aT )}1 , the cumulative reward over this sequence can be written as
T
X

R(τ ) =

(4.17)

rt

t=1

The expectation of this cumulative reward over a space of state-action sequences is given in
the form
J(θ) =
=

E
τ ∼D

X

h

R(τ ) πθ

i

Prθ (τ )R(τ )

τ ∈D

=

X

τ ∈D
1

Prθ (τ )

T
X

rt

(4.18)

t=1

We assume the state-action sequence begins with s1 and a1 , rather than s0 and a0 , to align with the notation
commonly used in this chapter, where the prediction y typically starts from y1 . Of course, it is also common to denote
a state-action sequence as {(s0 , a0 ), ..., (sT , aT )} or {(s0 , a0 ), ..., (sT −1 , aT −1 )} in the literature. But this variation
in notation does not affect the discussion of the models presented here.

Alignment

176

where τ ∼ D indicates that τ is drawn from the state-action sequence space D, and the subscript
θ indicates the parameters of the policy. J(θ) is also called the performance function.
Then the training objective is to maximize J(θ):

θ̃ = arg max J(θ)

(4.19)

θ

Now, we have a simple reinforcement learning approach: 1) we sample a number of stateaction sequences; then, 2) we evaluate each sequence using the performance function; then, 3) we
update the model to maximize this performance function. If we take Eq. (4.18) and use gradient
descent to optimize the policy, this approach would constitutes a form of policy gradient methods
[Williams, 1992].
Note that in many NLP problems, such as machine translation, rewards are typically sparse.
For instance, a reward is only received at the end of a complete sentence. This means that rt = 0
for all t < T , and rt is non-zero only when t = T . Ideally, one might prefer feedback to
be immediate and frequent (dense), and thus the training of the policy can be easier and more
efficient. While several methods have been proposed to address sparse rewards, such as reward
shaping, we will continue in our discussion to assume a sparse reward setup, where the reward is
available only upon completing the prediction.
The model described in Eqs. (4.17-4.19) establishes a basic form of reinforcement learning,
and many variants and improvements of this model have been developed. Before showing those
more sophisticated models, let us take a moment to interpret the objective function J(θ) from the
perspective of policy gradient. In gradient descent, we need to compute the gradient of J(θ) with
respect to θ:
∂J(θ)
∂θ

∂

=

P

τ ∈D Prθ (τ )R(τ )

∂θ

X ∂Prθ (τ )

=

τ ∈D

X

=

∂θ

Prθ (τ )

∂Prθ (τ )/∂θ
R(τ )
Prθ (τ )

Prθ (τ )

∂ log Prθ (τ )
R(τ )
∂θ

τ ∈D

X

=

τ ∈D

R(τ )

(4.20)

In some cases, we will assume that every sequence in D is equally probable (i.e., Prθ (τ ) =
Prθ (τ )
and
1/|D|). In this case we can simplify Eq. (4.20) and need only consider the terms ∂ log∂θ
R(τ ):
∂J(θ)
∂θ

=

1 X ∂ log Prθ (τ )
R(τ )
m τ ∈D
∂θ

(4.21)

One advantage of this result is that R(τ ) does not need to be differentiable, which means that we
can use any type of reward function in reinforcement learning.
By treating the generation of the sequence τ as a Markov decision process, we can further

4.3 Human Preference Alignment: RLHF

177

Prθ (τ )
, and obtain:
derive ∂ log∂θ

∂ log Prθ (τ )
∂θ

T
Y
∂
log
πθ (at |st ) Pr(st+1 |st , at )
∂θ
t=1

=

T
T
∂ X
∂ X
log πθ (at |st ) +
log Pr(st+1 |st , at )
|
{z
} ∂θ
{z
}
∂θ t=1 |
t=1

=

policy

(4.22)

dynamics

where the gradient is decomposed into two parts: the policy gradient and the dynamics gradient.
The policy component, log πθ (at |st ), determines the log-probability of taking action at given
state st , and it is parameterized by θ. The dynamics component, log Pr(st+1 |st , at ), represents
the log-probability of transitioning to state st+1 from state st after taking action at . In typical
reinforcement learning settings, the dynamics are not directly influenced by the policy parameters
θ, and thus, their derivatives are often zero. In this case, therefore, Eq. (4.22) can be simplified to:
∂ log Prθ (τ )
∂θ

=

T
∂ X
log πθ (at |st )
∂θ t=1

(4.23)

In other words, we only concentrate on optimizing the policy without concerning ourselves with
the underlying dynamics.
Substituting Eq. (4.23) into Eq. (4.21), and expanding R(τ ), we then obtain
∂J(θ)
∂θ

=

T
T

X
1 X ∂ X
log πθ (at |st )
rt
|D| τ ∈D ∂θ t=1
t=1

(4.24)

While this policy gradient approach is straightforward, it suffers from the problem that the
variance of the estimated gradients can be very high, making the learning process noisy and inefficient. One reason for this high variance problem is that rewards can vary greatly across different
steps or scenarios. Imagine that in a sequence of action decisions, the reward model tends to assign
small rewards to good actions (e.g., Rt = 2) and large penalties to poor actions (e.g., Rt = −50).
Such varying reward scales for good and poor actions can result in a very low total reward for the
entire sequence, even if it includes good actions.
One simple method for reducing the variance of the gradient is to set a baseline b and subtract
P
P
it from Tt=1 rt , resulting in Tt=1 rt − b.2 Here, the baseline can be interpreted as a reference
point. By centering the rewards around this baseline, we remove systematic biases in the reward
signal, making the updates more stable and less sensitive to extreme fluctuations in individual
rewards.

2

PT

In fact, the use of a baseline b does not change the variance of the total rewards t=1 rt . However, it is important
to note that while introducing a baseline does not alter the overall variance of the rewards, it helps reduce the variance
of the gradient estimates. This is because subtracting the baseline from the total rewards effectively
PT reduces fluctuations
around their mean, which makes the gradient estimates more stable. In general, the operation t=1 rt − b centers the
PT
rewards around zero (e.g., b is defined as the expected value of t=1 rt ), which can lead to reduced variance in the
PT
PT
product t=1 log πθ (at |st )( t=1 rt − b).

Alignment

178
This policy gradient model with a baseline can be given by
∂J(θ)
∂θ

T
T

 X
1 X ∂ X
rt − b
log πθ (at |st )
|D| τ ∈D ∂θ t=1
t=1

=

T
T
i
X
1 X ∂ hX
rk − b
log πθ (at |st )
|D| τ ∈D ∂θ t=1
k=1

=

T
t−1
T
i
X
X
1 X ∂ hX
rk − b
rk +
log πθ (at |st )
|D| τ ∈D ∂θ t=1
k=t
k=1

=

P

P

P

(4.25)

t−1
rk and Tk=t rk to distinguish between the
Here we write Tk=1 rk as the sum of two terms k=1
rewards accrued before and after the action at time step t. Note that in Markov decision processes,
the future is independent of the past given the present. Therefore, the action taken at time step t
cannot influence the rewards received before t, or in other words, the rewards prior to t are already
P
“fixed” by the time the action at t is chosen. The term t−1
k=1 rk does not contribute to the gradient
and can be omitted, leading to a simplified version of Eq. (4.25)

∂J(θ)
∂θ
Also note that removing

=

T
T
i
X
1 X ∂ hX
rk − b
log πθ (at |st )
|D| τ ∈D ∂θ t=1
k=t

(4.26)

PT

k=t rk can further reduce the variance of the gradient.

There are many ways to define the baseline b. Here we consider the value function of the state
st , that is, the estimated value of being in state st : V (st ) = E(rt + rt+1 + · · · + rT ). Hence we
have
A(st , at ) =
=

T
X
k=t
T
X
k=t

rk − b
rk − V (st )

(4.27)

P

where Tk=t rk represents the actual return received, and V (st ) represents the expected return.
A(st , at ) (or At for short) is called the advantage at time step t, which quantifies the relative
benefit of the action at compared to the expected value of following the policy from the state st
onward.
By using the advantage function A(st , at ), the gradient of J(θ) can be written in the form
∂J(θ)
∂θ

=

T

1 X ∂ X
log πθ (at |st )A(st , at )
|D| τ ∈D ∂θ t=1

(4.28)

This optimization objective corresponds to the advantage actor-critic (A2C) method in reinforcement learning [Mnih et al., 2016]. In this method, the actor aims at learning a policy. It
updates the policy parameters using Eq. (4.28) to help focus more on actions that are likely to
improve performance. The critic, on the other hand, updates its estimation of the value function,
which is used to calculate the advantage function A(st , at ), thus serving as the evaluator of the

4.3 Human Preference Alignment: RLHF

179

policy being learned by the actor.
In the A2C method, A(st , at ) is typically expressed as the difference of the action-value function Q(st , at ) and the state-value function V (st )
A(st , at ) = Q(st , at ) − V (st )

(4.29)

At first glance, this model may seem challenging to develop because it requires two separate submodels to calculate Q(st , at ) and V (st ) respectively. Fortunately, considering that Q(st , at ) can
be defined as the return rt + V (st+1 ), we can rewrite Eq. (4.29) as
A(st , at ) = rt + V (st+1 ) − V (st )

(4.30)

or alternatively, introduce the discount factor γ to obtain a more general form
A(st , at ) = rt + γV (st+1 ) − V (st )

(4.31)

A(st , at ) = rt + γV (st+1 ) − V (st ) is also called the temporal difference (TD) error. What
we need is to train a critic network for the value function V (st ), and then use it to compute the
advantage function3 .
Up to this point, we have spent considerable space discussing the basics of reinforcement
learning, especially on how to derive the optimization objective for the A2C method. However,
reinforcement learning is a vast field, and many technical details cannot be covered here. The interested reader can refer to reinforcement learning books for more details [Sutton and Barto, 2018;
Szepesvári, 2010]. Nevertheless, we now have the necessary knowledge to further discuss RLHF.
In the subsequent subsections, we will return to the discussion on LLM alignment, demonstrating
how to use the A2C method for aligning with human preferences.

4.3.2 Training Reward Models
We have shown that reward models play a very important role in the general reinforcement learning framework and form the basis for computing value functions. We now consider the problem
of training these reward models.
In RLHF, a reward model is a neural network that maps a pair of input and output token
sequences to a scalar. Given an input x and an output y, the reward can be expressed as
r = Reward(x, y)

(4.33)

where Reward(·) is the reward model. r can be interpreted as a measure of how well the output y
aligns with the desired behavior given the input x. As discussed in the previous subsection, both x
3

The training loss for the value network (or critic network) in A2C is generally formulated as the mean squared
error between the computed return rt + γV (st+1 ) and the predicted state value V (st ). Suppose that the value network
is parameterized by ω. The loss function is given by
Lv (ω)

=


2
1 X
rt + γVω (st+1 ) − Vω (st )
M

where M is the number of training samples, for example, for a sequence of T tokens, we can set M = T .

(4.32)

Alignment

180
Reward (Scalar)
Wr
Representation

hx0

hx1

hx2

···

hxm

hy1

hy2

···

hlast

···

yn

Linear Map

at Each Position

Transformer Decoder (LLM)
x0

x1

x2

···

xm

y1

y2

(Last Token hEOSi)

Fig. 4.8: Architecture of the reward model based on Transformer. The main component of this model is still an LLM.
We use the Transformer decoder as the sequence representation model. We extract the representation of the last position
of the decoder as the representation of the entire sequence [x, y]. We then map this representation to a scalar through a
linear transformation, which serves as the reward score for [x, y].

and y are assumed to complete texts. This means that the reward model evaluates the relationship
between inputs and outputs that provide full semantic content. For example, when applying the
reward model, it assigns a value of 0 (or another predetermined value) at each position t in the
output sequence y = y1 ...yn . Only at the final position, when t = n, does the reward model
generate the actual reward score. To keep the notation uncluttered, we will use r(x, y) to denote
the reward model Reward(x, y) from here on.
There are many ways to implement the reward model. One simple approach is to build the
reward model based on a pre-trained LLM. More specifically, we can concatenate x and y to form
a single token sequence seq x,y = [x, y]. We run a pre-trained LLM on this sequence, as usual,
and at each position, we obtain a representation from the top-most Transformer layer. Then, we
take the representation at the last position (denoted by hlast ) and map it to a scalar via linear
transformation:
r(x, y) = hlast Wr

(4.34)

where hlast is a d-dimensional vector, and Wr is a d × 1 linear mapping matrix. This architecture
of the reward model is illustrated in Figure 4.8.
To train the reward model, the first step is to collect human feedback on a set of generated
outputs. Given an input x, we use the LLM to produce multiple candidate outputs {y1 , ..., yN }.
Human feedback can be obtained in several ways:

• Pairwise Comparison (Pairwise Ranking). Given two different outputs, human experts
select which one is better.
• Rating. Human experts provide a score or rating to each output. This score is often a
continuous or discrete numerical value, such as a score on a scale (e.g., 1-5 stars, or 1-10
points). In some cases, the rating might be binary, indicating a “yes/no” or “positive/negative” preference.

4.3 Human Preference Alignment: RLHF

181

• Listwise Ranking. Human experts are asked to rank or order the given set of possible
outputs.
Here we consider pairwise comparison feedback as it is one of the simplest and most common forms of human feedback used in RLHF. In this setting, each time, two outputs (ya , yb ) are
randonly drawn from the candidate pool {y1 , ..., yN }. Human experts are then presented with
these pairs and asked to decide which output they prefer based on specific criteria, such as clarity,
relevance, and accuracy. The human feedback can be encoded as a binary label, ya ≻ yb for a
preference for ya , and yb ≻ ya for a preference for yb .

One simple and widely used model for describing such pairwise comparisons is the BradleyTerry model [Bradley and Terry, 1952]. It is a probabilistic model that estimates the probability
that one item is preferred over another. Adapting this model to the notation used here, we can
write the probability that ya is preferred over yb in the form
er(x,ya )
er(x,ya ) + er(x,yb )
er(x,ya )−r(x,yb )
= r(x,y )−r(x,y )
a
b + 1
e
= Sigmoid(r(x, ya ) − r(x, yb ))

Pr(ya ≻ yb |x) =

(4.35)

When training the reward model, we want to maximize this preference probability. A loss
function based on the Bradley-Terry model is given by




Lr (φ) = −E(x,ya ,yb )∼Dr log Prφ (ya ≻ yb |x)

(4.36)

where (x, ya , yb ) is drawn from a human-annotated dataset Dr consisting of preference pairs of
outputs and their corresponding inputs. φ represents the parameters of the reward model, which
includes both the parameters of the Transformer decoder and the linear mapping matrix Wr . In
practice, assuming (x, ya , yb ) is uniformly sampled from Dr , we can replace the expectation with
a summation
Lr (φ) = −

X
1
log Prφ (ya ≻ yb |x)
|Dr | (x,y ,y )∈D
a

b

(4.37)

r

The goal of training the reward model is to find the optimal parameters φ̂ that minimize this
loss function, given by
φ̂ = arg min Lr (φ)

(4.38)

φ

Since the reward model itself is also an LLM, we can directly reuse the Transformer training
procedure to optimize the reward model. The difference from training a standard LLM is that we
only need to replace the cross-entropy loss with the pairwise comparison loss as described in Eq.
(4.37). After the training of the reward model, we can apply the trained reward model rφ̂ (·) to
supervise the target LLM for alignment.
It is worth noting that although we train the reward model to perform pairwise ranking, we
apply it to score each input-output pair independently during the alignment process. The pairwise

Alignment

182

ranking objective ensures that the reward model is sensitive to subtle differences between outputs,
but we rely on the continuous scores produced by the reward model to guide the optimization of
the LLM. An advantage of this approach is that we can choose from or combine various ranking
loss functions, and still apply the resulting reward models in the same way as we have done in this
subsection. This consistency ensures a unified framework for aligning the LLM, regardless of the
specific ranking loss used during reward model training.

4.3.3 Training LLMs
Having obtained the reward model, we then train the policy (i.e., the LLM) via the A2C method.
Recall from Section 4.3.1 that a state-action sequence or trajectory τ can be evaluated by the utility
function
U (τ ; θ) =

T
X
t=1

(4.39)

log πθ (at |st )A(st , at )

where A(st , at ) is the advantage of taking the action at given the state st . An estimate of A(st , at )
is defined as the TD error rt + γV (st+1 ) − V (st ), where the value function V (st ) is trained with
the reward model.
Given this utility function, the A2C-based loss function can be written in the form




L(θ) = −Eτ ∼D U (τ ; θ)
= −Eτ ∼D

T
X
t=1



log πθ (at |st )A(st , at )

(4.40)

where D is a space of state-action sequences. As usual, the goal of training the policy is to
minimize this loss function
(4.41)

θ̃ = arg min L(θ)
θ

If we map the problem back to the language modeling problem and adopt the notation from
LLMs, the loss function can be written as:




L(θ) = −E(x,y)∼D U (x, y; θ)

(4.42)

where
U (x, y; θ) =

T
X
t=1

log πθ (yt |x, y<t )A(x, y<t , yt )

Here πθ (yt |x, y<t ) = Prθ (yt |x, y<t ) is the LLM parameterized by θ.

(4.43)

In general, we do not have a human annotated input-output dataset D in RLHF, but a dataset
containing inputs only. The outputs, in this case, are typically the predictions made by the LLM.

4.3 Human Preference Alignment: RLHF

183

The loss function is then defined as




(4.44)

L(θ) = −Ex∼D Ey∼πθ (·|x) U (x, y; θ)

where D denotes the input-only dataset, and y ∼ πθ (·|x) denotes that the output y is sampled by
the policy πθ (·|x).
The above formulation provides a basic form of the A2C method for LLMs. Improved versions
of this model are more commonly used in RLHF. In the following discussion, we will still use
the reinforcement learning notation to simplify the presentation and will get back the language
modeling notation later.
One common improvement of policy gradient methods is to use importance sampling to
refine the estimation of U (τ ; θ). This can be written as
T
X
πθ (at |st )

U (τ ; θ) =

t=1

πθref (at |st )

(4.45)

A(st , at )

Here we replace the log-probability log πθ (at |st ) with the ratio ππθ θ (a(at |st |st )t ) . θref denotes the paref
rameters of the previous policy (such as an initial model from which we start the training). So
πθ (at |st )
πθref (at |st ) , also called the ratio function, can be interpreted as the log-probability ratio between
the current policy πθ and the previous policy πθref (call it the reference policy). By using the
ratio function we reweight the observed rewards based on the likelihood of the actions under the
current policy versus the reference policy. When ππθ θ (a(at |st |st )t ) > 1, the action at is more favored by
ref

the current policy compared to the reference policy. By contrast, when ππθ θ (a(at |st |st )t ) < 1, the action
ref

at is less favored by the current policy4 .
4

Consider a more general case where we wish to evaluate the policy using its expected reward (also see Eq. (4.18))
J(θ)

=

h

Eτ ∼πθ R(τ )

i

(4.46)

Here τ ∼ πθ means that the sequence τ is generated by the policy πθ . Alternatively, we can write J(θ) in another form
J(θ)

=

Eτ ∼πθref

h Pr (τ )
θ

Prθref (τ )

R(τ )

i

(4.47)

It is not
that the right-hand sides of these equations are essentially
h
i the same since
h difficult toi find
P
P
Prθ (τ )
Prθ (τ )
Eτ ∼πθref Prθ (τ ) R(τ ) = τ Prθref (τ ) Prθ (τ ) R(τ ) = τ Prθ (τ )R(τ ) = Eτ ∼πθ R(τ )
ref

ref

Note that this equivalence holds only when the expectation is performed over the entire sequence space. In practice,
however, we often only sample a relatively small number of sequences using a policy in policy learning. As a result,
the sampling method itself matters. Eq. (4.47) offers an interesting manner to separate the sampling and reward
computation processes: we first use a baseline policy (with θref ) to sample a number of sequences, and then use the
target policy (with θ) to compute the expected reward. In this way, we separate the policy used for collecting the data,
and the policy used for computing the gradient. This approach avoids the need to directly sample from the policy we are
evaluating, which can be beneficial in
sequences from the target policy is expensive or difficult.
h cases where generating
i
In reinforcement learning, Eτ ∼πθref

Prθ (τ )
R(τ )
Prθ
(τ )

is often called a surrogate objective.

ref

Eq. (4.47) can also be interpreted from a policy gradient perspective. For Eτ ∼πθref

θ = θref is given by

h

Prθ (τ )
∂
Eτ ∼πθref
R(τ )
∂θ
Prθref (τ )

i

=
θ=θref

Eτ ∼πθref

h

h ∂Pr (τ )|
θ

Prθ (τ )
R(τ )
Prθ
(τ )
ref

θ=θref

∂θ

R(τ )

i

i

, the gradient at

(4.48)

The right-hand side is a standard form used in policy gradient methods, meaning that we compute the direction of

Alignment

184

A problem with the model presented in Eq. (4.47) (as well as in Eq. (4.39)) is that the
variance in the gradient estimates is often high, making the learning process unstable. To mitigate
this issue, techniques such as clipping are often employed to bound the importance weights and
prevent large updates. A clipped version of the utility function (also called the clipped surrogate
objective function) is given by
Uclip (τ ; θ) =
Clip

 π (a |s ) 
θ

t

t

πθref (at |st )

T
X

 π (a |s ) 

t=1

= min

t

θ

Clip

t

πθref (at |st )

 π (a |s )
θ

t

t

πθref (at |st )

A(st , at )

, bound



πθ (at |st )
, 1 − ǫ, 1 + ǫ
πθref (at |st )

(4.49)
(4.50)

Here the function bound( ππθ θ (a(at |st |st )t ) , 1 − ǫ, 1 + ǫ) constrains the ratio function to the range [1 −
ǫ, 1 + ǫ].

ref

A further improvement to the above model is to consider trust regions in optimization [Schulman et al.,
2015]. In reinforcement learning, a large update to the policy can lead to instability, where the
agent may start performing worse after an update. A reasonable idea is to optimize the model in
the trust region, which refers to a region around the current parameter estimate where the model
is well-behaved. One approach to incorporating trust regions is to impose a constraint on the size
of the policy update, ensuring that the current policy does not deviate too significantly from the
reference policy. This can be achieved by adding a penalty based on some form of divergence
between the current and reference policies to the objective function. A simple form of such a
penalty is given by the difference in the log-probability of the sequence τ under the current policy
versus the reference policy:
Panalty = log πθ (τ ) − log πθref (τ )

(4.51)

In practice, this penalty can be approximated by considering only the policy probabilities and
ignoring the dynamics. This gives
Penalty =

T
X
t=1

log πθ (at |st ) −

T
X
t=1

log πθref (at |st )

(4.52)

By including this penalty in the optimization objective, we encourage the current policy to remain
close to the reference policy, limiting very large updates that could destabilize learning.
We can incorporate this penalty into the clipped surrogate objective function, and obtain
Uppo-clip (τ ; θ) = Uclip (τ ; θ) − βPenalty

(4.53)

where β is the weight of the penalty. This training method is called proximal policy optimization
(PPO), which is one of the most popular reinforcement learning methods used in LLMs and many
other fields [Schulman et al., 2017].
Now we can write the objective of training LLMs in the form of PPO.

the parameter update at the point θ = θref on the optimization surface.

4.3 Human Preference Alignment: RLHF

185

U (x, y; θ) = Uppo-clip (x, y; θ) − βPenalty

(4.54)

where
Uppo-clip (x, y; θ) =

T
X

Clip

 π (y |x, y



T
X

log Prθ (yt |x, y<t ) −

T
X

t=1

θ

t

<t )

πθref (yt |x, y<t )

A(x, y<t , yt )

(4.55)

Penalty = log Prθ (y|x) − log Prθref (y|x)
=

t=1

t=1

log Prθref (yt |x, y<t )

(4.56)

Although the notation here appears a bit tedious, the idea of PPO is simple: we develop an
objective by combining the clipped likelihood ratio of the target and reference policies with an
advantage function, and then impose a penalty that ensures policy updates are not too large. The
PPO-based RLHF is illustrated in Figure 4.9.
To summarize, implementing RLHF requires building four models, all based on the Transformer decoder architecture.
• Reward Model (rφ (·) where φ denotes the parameters). The reward model learns from
human preference data to predict the reward for each pair of input and output token sequences. It is a Transformer decoder followed by a linear layer that maps a sequence (the
concatenation of the input and output) to a real-valued reward score.
• Value Model or Value Function (Vω (·) where ω denotes the parameters). The value function receives reward scores from the reward model and is trained to predict the expected
sum of rewards that can be obtained starting from a state. It is generally based on the same
architecture as the reward model.
• Reference Model (πθref (·) = Prθref (·) where θref denotes the parameters). The reference
model is the baseline LLM that serves as a starting point for policy training. In RLHF, it
represents the previous version of the model or a model trained without human feedback. It
is used to perform sampling over the space of outputs and contribute to the loss computation
for policy training.
• Target Model or Policy (πθ (·) = Prθ (·) where θ denotes the parameters). This policy
governs how the LLM decides the most appropriate next token given its context. It is trained
under the supervision of both the reward model and the value model.
In practice, these models need to be trained in a certain order. First, we need to initialize them
using some other models. For example, the reward model and the value model can be initialized
with a pre-trained LLM, while the reference model and the target model can be initialized with a
model that has been instruction fine-tuned. Note that, at this point, the reference model is ready for
use and will not be further updated. Second, we need to collect human preference data and train the
reward model on this data. Third, both the value model and the policy are trained simultaneously
using the reward model. At each position in an output token sequence, we update the value model

Alignment

186
Human preference data Dr = {(x, ya , yb )}
To Learn

Reward Model
Training

Minimizing the loss based on
the Bradley-Terry model
P
1

Reward Model
rφ (x, y)

min − |D |

log σ(rφ (x, ya ) − rφ (x, yb ))

r

φ

(x,ya ,yb )∈Dr

To Learn

Fixed

To Learn

Ref Model

LLM Policy

Value Function

Prθold (yt |x, y<t )

Prθ (yt |x, y<t )

Vω (x, y<t )

Evaluate the state-action pair using the advantage
function or the TD error (based on the reward
model and the value function)
x1

x2

y1

···

y2

yt

···

Action yt
(sampled with Prθ

old

)

LLM Policy
x0

Policy Training

x1

xm

···

y1

yt−1

···

State (x, y<t )

LLM Policy

Value Function

Minimizing the clipped PPO loss
with the penalty

Minimizing the MSE between the
computed return and the predicted
state value

min −

h

θ

Clip

P

x∈D,y∼Prθ

old



Prθ (yt |x,y<t )

Prθ

old

(yt |x,y<t )

PT

(·|x)

t=1

At −

β · log Prθ (yt |x, y<t )−


i

1
min M



ω

P

x∈D

PT

t=1

2

rt + γVω (x, y<t+1 ) − Vω (x, y<t )

log Prθold (yt |x, y<t )

∗∗ rt = r(x, y<t+1 ) denotes the reward received as step t.
∗∗ At denotes the advantage at step t, and can be defined as rt + γVω (x, y<t+1 ) − Vω (x, y<t )

Input-only data D = {x}
Fig. 4.9: Illustration of RLHF. The first step is to collect human preference data and train the reward model using this
data. Once the reward model is optimized, along with the reference model, we proceed to train both the policy and
the value function. At each prediction step, we compute the sum of the PPO-based loss and update the parameters of
the policy. This requires access to the reward model, the reference model, and the value function at hand. At the same
time, we update the parameters of the value function by minimizing the MSE loss.

by minimizing the MSE error of value predition, and the policy is updated by minimizing the PPO
loss.

4.4 Improved Human Preference Alignment

187

4.4 Improved Human Preference Alignment
In the previous section, we reviewed the basic concepts of reinforcement learning and the general
framework of RLHF. In this section, we will discuss some refinements of RLHF and alternative
methods to achieve human preference alignment.

4.4.1 Better Reward Modeling
In Section 4.3.2, we highlighted the task of learning from human preferences as well as the use
of pairwise ranking loss for training reward models. Here we consider more methods for reward
modeling. Our discussion will be relatively general, and since the reward model is widely used in
many reinforcement learning problems, it will be easy for us to apply the methods discussed here
to RLHF and related applications.

4.4.1.1 Supervision Signals
The training of reward models can broadly be seen as a ranking problem, where the model learns
to assign scores to outputs so that their order reflects the preferences indicated by humans. There
are several methods to train a reward model from the perspective of ranking.
One approach is to extend pairwise ranking to listwise ranking. For each sample in a dataset,
we can use the LLM to generate multiple outputs, and ask human experts to order these outputs.
For example, given a set of four outputs {y1 , y2 , y3 , y4 }, one possible order of them can be
y2 ≻ y3 ≻ y1 ≻ y4 . A very simple method to model the ordering of the list is to accumulate the
pairwise comparison loss. For example, we can define the listwise loss by accumulating the loss
over all pairs of outputs:
Llist = −E(x,Y )∼Dr

h

i
X
1
log Pr(ya ≻ yb |x)
N (N − 1) y ∈Y,y ∈Y
a

(4.57)

b
ya 6=yb

where Y is a list of outputs, and N is the number of outputs in the list. Pr(ya ≻ yb |x) can be
defined using the Bradley-Terry model, that is, Pr(ya ≻ yb |x) = Sigmoid(r(x, ya ) − r(x, yb )).
Here we omit the φ superscript on the Pr(·) to keep the notation uncluttered.
An extension to the Bradley-Terry model for listwise ranking could involve a ranking mechanism that takes into account the entire list of outputs rather than just pairwise comparisons. One
such model is the Plackett-Luce model, which generalizes the Bradley-Terry model to handle
multiple items in a ranking [Plackett, 1975]. In the Plackett-Luce model, for each item in a list,
we define a “worth” for this item that reflects its relative strength of being chosen over other items.
For the reward modeling problem here, the worth of y in the list Y can be defined as
α(y) = exp(r(x, y))

(4.58)

Alignment

188
Then the probability of selecting y from Y is given by
Pr(y is selected|x, Y ) =
=

P
P

α(y)
′
y′ ∈Y α(y )
exp(r(x, y))
′
y′ ∈Y exp(r(x, y ))

(4.59)

Suppose Y̊ is an ordered list yj1 ≻ yj2 ≻ · · · ≻ yjN . The overall log-probability of this
ordered list can be defined as the sum of the conditional log-probabilities at each stage of selection,
given by
log Pr(Y̊ |x) = log Pr(yj1 ≻ yj2 ≻ · · · ≻ yjN |x)

= log Pr(yj1 |x, {yj1 , yj2 , ..., yjN }) +
log Pr(yj2 |x, {yj2 , ..., yjN }) +
··· +

log Pr(yjN |x, {yjN })
=

N
X

k=1

(4.60)

log Pr(yjk |x, Y̊≥k )

where Y̊≥k represents the subset of the list of outputs that remain unselected at the k-th stage, i.e.,
Y̊≥k = {yjk , ..., yjN }. Given the log-probability log Pr(Y̊ |x), we can define the loss function
based on the Plackett-Luce model by




Lpl = −E(x,Y̊ )∼Dr log Pr(Y̊ |x)

(4.61)

There are also many other pairwise and listwise methods for modeling rankings, such as
RankNet [Burges et al., 2005] and ListNet [Cao et al., 2007]. All these methods can be categorized into a large family of learning-to-rank approaches, and most of them are applicable to the
problem of modeling human preferences. However, discussing these methods is beyond the scope
of this chapter. Interested readers can refer to books on this topic for more details [Liu, 2009; Li,
2011].
In addition to pairwise and listwise ranking, using pointwise methods to train reward models
offers an alternative way to capture human preferences. Unlike methods that focus on the relative
rankings between different outputs, pointwise methods treat each output independently. For example, human experts might assign a score to an individual output, such as a rating on a five-point
scale. The objective is to adjust the reward model so that its outputs align with these scores. A
simple way to achieve pointwise training is through regression techniques where the reward of
each output is treated as a target variable. Let ϕ(x, y) be the score assigned to y given x by
humans. Pointwise reward models can be trained by minimizing a loss function, often based on
mean squared error or other regression losses, between the predicted reward r(x, y) and the actual
human feedback ϕ(x, y). For example, the loss function could be


2

Lpoint = −E ϕ(x, y) − r(x, y)

(4.62)

While pointwise methods are conceptually simpler and can directly guide the reward model to

4.4 Improved Human Preference Alignment

189

predict scores, they might not always be the best choice in RLHF. A problem is that these methods
may struggle with high variance in human feedback, especially when different experts provide
inconsistent scores for similar outputs. Because they focus on fitting to absolute scores rather than
relative differences, inconsistencies in scoring can lead to poor model performance. Moreover,
fitting to specific scored outputs might discourage generalization, particularly given that training
data is often very limited in RLHF. In contrast, methods that consider relative preferences can
promote the learning of more generalized patterns of success and failure. Nevertheless, there are
scenarios where pointwise methods might still be suitable. For example, in tasks where training
data is abundant and the costs of obtaining accurate, consistent annotations are low, pointwise
methods can prove effective.
In fact, to make the supervision signal for training the reward model more robust, we can also
introduce additional regularization terms into training. For example, if we consider the first term
Uppo-clip (x, y; θ) in Eq. (4.54) as a type of generalized reward, then the second term (i.e., the
penalty term) can be viewed as a form of regularization for the reward model, except that here the
goal is to train the policy rather than the reward model. Another example is that Eisenstein et al.
[2023] develop a regularization term based on the squared sum of rewards, and add it to the
pairwise comparison loss in RLHF:


2

Lreg = Lpair + (−E(x,ya ,yb )∼Dr r(x, ya ) + r(x, yb ) )




= −E(x,ya ,yb )∼Dr log Prφ (ya ≻ yb |x)


2

−E(x,ya ,yb )∼Dr r(x, ya ) + r(x, yb )

(4.63)

Optimizing with this regularization term can help mitigate the underdetermination of reward models5 .

4.4.1.2 Sparse Rewards vs. Dense Rewards
As discussed in Section 4.3, the rewards in RLHF are very sparse: they are observed only at the
end of sequences, rather than continuously throughout the generation process. Dealing with sparse
rewards has long been a concern in reinforcement learning, and has been one of the challenges in
many practical applications. For example, in robotics, it often needs to shape the reward function
to ease optimization rather than relying solely on end-of-sequence rewards. Various methods
have been developed to address this issue. One common approach is reward shaping, where the
original function is modified to include intermediate rewards, thereby providing more immediate
feedback. Also, one can adopt curriculum learning to sequentially structure tasks in a way that the
complexity gradually increases. This can help models to master simpler tasks first, which prepares
them for more complex challenges as their skills develop. There are many such methods that can
mitigate the impact of sparse rewards, such as Monte Carlo methods and intrinsic motivation. Most
of these methods are general and the discussion of them can be found in the broader literature on
reinforcement learning, such as Sutton and Barto [2018]’s book.
Although we do not discuss methods for mitigating sparse rewards in detail here, an interesting
question arises: why are sparse rewards so successful in RLHF? Recall from Section 4.3.1 that
the supervision signal received at each time step t is not the reward for the current action, but
5
A model is called underdetermined if there are multiple alternative sets of parameters that can achieve the same
objective.

Alignment

190

rather some form of the accumulated rewards from t until the last time step. Such supervision
signals are dense over the sequence, because the reward obtained at the end of the sequence can
be transferred back to that time step, regardless of which time step it is. In other words, the sparse
rewards are transformed into the dense supervision signals. Furthermore, from the perspective of
reward shaping, Ng et al. [1999] show that the reward at t can be defined as
r ′ (st , at , st+1 ) = r(st , at , st+1 ) + f (st , at , st+1 )

(4.64)

where r ′ (·) is the transformed reward function, r(·) is the original reward function, and f (·) is
the shaping reward function. To ensure the optimality of the policy under the transformed reward
function, the shaping reward function can be given in the form
f (st , at , st+1 ) = γΦ(st+1 ) − Φ(st )

(4.65)

where Φ(s) is called the potential value of the state s. If we define Φ(s) as the common value
function as in Eq. (4.15) and substitute Eq. (4.65) into Eq. (4.64), we obtain
r ′ (st , at , st+1 ) = r(st , at , st+1 ) + γV (st+1 ) − V (st )

(4.66)

It is interesting to see that this function is exactly the same as the advantage function used in PPO.
This relates advantage-based methods to reward shaping: the advantage is essentially a shaped
reward.
On the other hand, one of the reasons for adopting end-of-sequence rewards lies in the nature
of the RLHF tasks. Unlike traditional reinforcement learning environments where the agent interacts with a dynamic environment, RLHF tasks often involve complex decision-making based
on linguistic or other high-level cognitive processes. These processes do not lend themselves easily to frequent and meaningful intermediate rewards because the quality and appropriateness of
the actions can only be fully evaluated after observing their impact in the larger context of the
entire sequence or task. In this case, the reward signals based on human feedback, though very
sparse, are typically very informative and accurate. Consequently, this sparsity, together with the
high informativeness and accuracy of the human feedback, can make the learning both robust and
efficient.

4.4.1.3 Fine-grained Rewards
For many applications, our objective will be more complex than merely evaluating an entire text.
For example, in sentiment analysis, we often do not just determine the sentiment of a text, but need
to analyze the sentiment in more detail by associating it with specific aspects of a topic discussed
in the text. Consider the sentence "The camera of the phone is excellent, but the battery life is
disappointing." In this example, we would need to separately analyze the sentiments expressed
about the camera and the battery. Such analysis, known as aspect-based sentiment analysis, helps
provide a finer-grained understanding of the customer review compared to general sentiment analysis.
For the problem of reward modeling, we often need to model different parts of a sequence as
well. A simple and straightforward way to do this is to divide a sequence into different segments
and then compute the reward for each segment [Wu et al., 2023]. Suppose that an output token

4.4 Improved Human Preference Alignment

191

sequence y can be divided into ns segments {ȳ1 , ..., ȳns } by some criterion. We can use the
reward model to evaluate each of these segments. By taking x, y and ȳk as input to the reward
model, the reward score for the k-th segment is given by
r k = r(x, y, ȳk )

(4.67)

Then the reward score for the entire output sequence is given by
ns
X

r(x, y) =

(4.68)

r(x, y, ȳk )

k=1

Here r(x, y) can be used to train the policy as usual.
A problem with this model is that training reward models at the segment level is not as straightforward as learning from human preferences on entire texts, as it is difficult to obtain segment-level
human preference data. For rating-like problems (e.g., we rate a segment according to its level
of misinformation), one simple approach is to assign a rating score to each segment and train the
reward model using pointwise methods. For example, we can use a strong LLM to rate the sequences ȳ1 ...ȳk−1 and ȳ1 ...ȳk , and obtain the scores s(ȳ1 ...ȳk−1 ) and s(ȳ1 ...ȳk ). We can then
define the score of the segment ȳk as the difference between s(ȳ1 ...ȳk ) and s(ȳ1 ...ȳk−1 )
s(ȳk ) = s(ȳ1 ...ȳk ) − s(ȳ1 ...ȳk−1 )

(4.69)

Using these segment-level scores, we can train the reward model with a regression loss function


2

Lrating = −Eȳk s(ȳk ) − r(x, y, ȳk )

(4.70)

Sometimes, alignment can be treated as a classification problem, for example, we assess
whether a segment has ethical issues. In this case, the segment can be labeled as ethical or
unethical, either by humans or using additional classifiers. Given the label of the segment, we
can train the reward model using some classification loss function. For example, suppose that
r(x, y, ȳk ) = 1 if the segment is classified as unethical, and r(x, y, ȳk ) = −1 otherwise6 . The
hinge loss of training binary classification models is given by
Lhinge = max(0, 1 − r(x, y, ȳk ) · r̂)

(4.71)

where r̂ ∈ {1, −1} denotes the ground truth label.

The remaining issue here is how to split y into segments. One approach is to define a fixedlength segmentation, where y is divided into equal-length chunks. However, this may not always
be ideal, as the content of the sequence may not align well with fixed boundaries. An alternative
approach is to segment y based on specific linguistic or semantic cues, such as sentence boundaries, topic shifts, or other meaningful structures in the text. Such a segmentation can be achieved
by using linguistic segmentation systems or prompting LLMs to identify natural breaks in the sequence. Another approach is to use dynamic segmentation methods based on the complexity of
6
To allow the reward model to output categories, we can replace the linear layer described in Section 4.3.2 with a
Softmax layer.

Alignment

192

the sequence. For example, segments could be defined where there is a significant change in the
reward score, which might correspond to shifts in the task being modeled.

4.4.1.4 Combination of Reward Models
A reward model can be viewed as a proxy for the environment. Since the true environment is often
too complex or unknown, developing a perfect proxy for the environment is generally not possible.
As a result, over-aligning LLMs with this imperfect proxy might lead to decreased performance,
known as the overoptimization problem [Stiennon et al., 2020; Gao et al., 2023a]7 . We can also
explain this through Goodhart’s law, which states: when a measure becomes a target, it ceases to
be a good measure [Goodhart, 1984].
Addressing the overoptimization problem is not easy, and there is no mature solution yet. The
ideal approach might be to develop an oracle reward model that can perfectly capture the true
objectives of the task and prevent the agent from “tricking”. However, creating such a model is
extremely difficult due to the complexity of the real-world environment, as well as the challenge
of defining all the relevant factors that contribute to the desired outcome. Instead, a more practical
approach is to combine multiple reward models, thereby alleviating the misalignment between
the training objective and the true objective that arises from using a single, specific reward model
[Coste et al., 2024].
Given a set of reward models, combining them is straightforward, and in some cases, we can
simply treat this problem as an ensemble learning problem. A simple yet common approach is to
average the outputs of these models to obtain a more precise reward estimation:
rcombine =

K
1 X
wk · rk (x, y)
N k=1

(4.72)

where rk (·) is the k-th reward model in the ensemble, wk is the weight of rk (·), and K is the
number of reward models. This combined reward can then be used to supervise the training of
a policy. In fact, there are many ways to combine different models, for example, one can make
predictions using Bayesian model averaging or develop a fusion network to learn to combine the
predictions from different models. Alternatively, one can frame this task as a multi-objective
optimization problem, and use multiple reward models to train the policy simultaneously. These
methods have been intensively discussed in the literature on optimization and machine learning
[Miettinen, 1999; Bishop, 2006].
In addition to model combination methods, another important issue is how to collect or construct multiple different reward models. One of the simplest approaches is to employ ensemble
learning techniques, such as developing diverse reward models from different subsets of a given
dataset or from various data sources. For RLHF, it is also possible to construct reward models
based on considerations of different aspects of alignment. For example, we can develop a reward
model to evaluate the factual accuracy of the output and another reward model to evaluate the
7
This problem is also called reward hacking or reward gaming [Krakovna et al., 2020; Skalse et al., 2022;
Pan et al., 2022], which refers to the phenomenon where the agent attempts to trick the reward model but fails to
align its actions with the true intended objectives of the task. Imagine a student who is assigned homework and is rewarded with points or praise for completing it. The student might then find ways to finish the homework with minimal
effort to maximize the reward, such as copying and pasting solutions from the internet or previous assignments, rather
than solving the problems themselves.

4.4 Improved Human Preference Alignment

Value Function

Preference
Data

ya ≻ yb

193

training with MLE

Training
with PPO

Reward Model
Policy
(a) RLHF (PPO)

Preference
Data

ya ≻ yb

training with MLE

Policy

(b) DPO
Fig. 4.10: Standard RLHF (PPO) vs. DPO. In RLHF, the human preference data is used to train a reward model, which
is then employed in training the policy as well as the value function. In DPO, the use of human preference data is more
direct, and the policy is trained on this data without the need for reward model training.

completeness of the output. These two models are complementary to each other, and can be combined to improve the overall evaluation of the output. Another approach is to employ different
off-the-shelf LLMs as reward models. This approach is simple and practical, as there have been
a lot of well-developed LLMs and we just need to use them with no or little modification. An
interesting issue, though not closely related to the discussion here, arises: can an LLM that aligns
with other LLMs outperform those LLMs? Probably not at first glance. In part, this is because
the target LLM merely imitates other LLMs based on limited supervision and thus cannot capture
well the nuances of the behaviors of these supervisors. However, given the strong generalization
ability of LLMs, this approach can, in fact, be quite beneficial. For example, using open-sourced
or commercial LLMs as reward models has demonstrated strong performance in aligning LLMs,
even achieving state-of-the-art results on several popular tasks [Lambert et al., 2024].

4.4.2 Direct Preference Optimization
Although learning reward models is a standard step in reinforcement learning, it makes the entire
training process much more complex than supervised training. Training a reliable reward model
is itself not an easy task, and a poorly trained reward model can greatly affect the outcome of
policy learning. We now consider an alternative alignment method, called direct preference optimization (DPO), which simplifies the training framework by eliminating the need to explicitly
model rewards [Rafailov et al., 2024]. This method directly optimizes the policy based on user
preferences, rather than developing a separate reward model. As a result, we can achieve human
preference alignment in a supervised learning-like fashion. Figure 4.10 shows a comparison of
the standard RLHF method and the DPO method.
Before deriving the DPO objective, let us first review the objective of policy training used in
RLHF. As discussed in Section 4.3.3, the policy is typically trained by optimizing a loss function

Alignment

194

with a penalty term. The DPO method assumes a simple loss function where the quality of the
output y given the input x is evaluated by the reward model r(x, y). The training objective is thus
given by


θ̃ = arg min Ex∼D Ey∼πθ (·|x) −r(x, y) +β (log πθ (y|x) − log πθref (y|x))
θ

|

{z

loss

}

|

{z

}

penalty



(4.73)

Note that in this optimization problem, only the term πθ (y|x) depends on the target policy πθ (·).
Both the reward model r(x, y) and the reference model πθref (y|x) are assumed to be fixed given
x and y. This is a strong assumption compared with PPO, but as will be shown later, it simplifies
the problem and crucial for deriving the DPO objective.
Since θ is the variable we want to optimize, we rearrange the right-hand side of Eq. (4.73) to
isolate πθ (y|x) as an independent term:




θ̃ = arg min Ex∼D Ey∼πθ (·|x) β log πθ (y|x) − β log πθref (y|x) − r(x, y)
θ



1
r(x, y)
β
θ



1
= arg min Ex∼D Ey∼πθ (·|x) log πθ (y|x) − log πθref (y|x) exp r(x, y)
|
{z
}
β
θ


= arg min Ex∼D Ey∼πθ (·|x) log πθ (y|x) − log πθref (y|x) +

dependent on θ

|

{z

not dependent on θ

(4.74)

}

This equation defines the objective function as the difference between the log-probability distribution function of y and another function of y. This form of the objective function seems not
“ideal”, as we usually prefer to see the difference between two distributions, so that we can interpret this difference as some kind of divergence between the distributions. A simple idea is
to convert the second term (i.e., log πθref (y|x) exp( β1 r(x, y))) into a log-probability distribution
over the domain of y. If we treat πθref (y|x) exp( β1 r(x, y)) as an unnormalized probability of y,
we can convert it into a normalized probability by dividing it by a normalization factor:
Z(x) =

X

πθref (y|x) exp

y



1
r(x, y)
β

(4.75)

Hence we can define a probability distribution by
∗

π (y|x) =

πθref (y|x) exp
Z(x)



1
β r(x, y)

(4.76)

4.4 Improved Human Preference Alignment

195

We then rewrite Eq. (4.74) as
h

θ̃ = arg min Ex∼D Ey∼πθ (·|x) log πθ (y|x) − log
θ

h

πθref (y|x) exp
Z(x

i

− log Z(x)







1
β r(x, y)

i

= arg min Ex∼D Ey∼πθ (·|x) log πθ (y|x) − log π ∗ (y|x) − log Z(x)
θ



h

i

= arg min Ex∼D Ey∼πθ (·|x) log πθ (y|x) − log π ∗ (y|x)
θ

h






−Ey∼πθ (·|x) log Z(x)




i

= arg min Ex∼D KL πθ (·|x) || π ∗ (·|x) − log Z(x)
θ

|

{z

KL divergence

}

|

{z

}

constant wrt. θ

(4.77)

Since log Z(x) is independent of θ, it does not affect the result of the arg minθ operation,
and can be removed from the objective. Now we obtain a new training objective which finds the
optimal policy πθ by minimizing the KL divergence between πθ (·|x) and π ∗ (·|x)
h


i

θ̃ = arg min Ex∼D KL πθ (·|x) || π ∗ (·|x)
θ

(4.78)

Clearly, the solution to this optimization problem is given by
πθ (y|x) = π ∗ (y|x)
πθref (y|x) exp β1 r(x, y))


=
Z(x

(4.79)

Given this equation, we can express the reward r(x, y) using the target model πθ (y|x), the
reference model πθref (y|x), and the normalization factor Z(x):
!

πθ (y|x)
+ log Z(x)
r(x, y) = β log
πθref (y|x)

(4.80)

This is interesting because we initially seek to learn the policy πθ (·) using the reward model
r(x, y), but eventually obtain a representation of the reward model based on the policy. Given the
reward model defined in Eq. (4.80), we can apply it to the Bradley-Terry model to calculate the
preference probability (also see Section 4.3.2):
Prθ (ya ≻ yb |x) = Sigmoid(r(x, ya ) − r(x, yb ))
 

πθ (ya |x)
+ log Z(x) −
= Sigmoid β log
πθref (ya |x)


πθ (yb |x)
+ log Z(x)
β log
πθref (yb |x)


πθ (yb |x)
πθ (ya |x)
− β log
= Sigmoid β log
πθref (ya |x)
πθref (yb |x)

(4.81)

Alignment

196

This formula is elegant because it converts the difference in rewards into the difference in
ratio functions, and we do not need to calculate the value of Z(x). A direct result is that we no
longer need a reward model, but only need the target policy and reference model to calculate the
probability of preferences. Finally, we can train the target policy by minimizing the following
DPO loss function




Ldpo (θ) = −E(x,ya ,yb )∼Dr log Prθ (ya ≻ yb |x)

(4.82)

The form of this loss function is very similar to that used in training reward models in RLHF (see
Eq. (4.36)). But it should be noted that the loss function here depends on the parameters of the
policy (i.e., θ) rather than the parameters of the reward model (i.e., φ).
The main advantage of DPO lies in its simplicity and efficiency. The DPO objective is very
straightforward — it directly optimizes for preference-based feedback, rather than relying on separately developed reward models. Moreover, DPO is generally more sample-efficient, as it learns
from a fixed dataset without the need for the computationally expensive sampling process used
in PPO. This makes DPO a popular method for human preference alignment, especially when
developing and applying reward models via reinforcement learning is challenging.
DPO can broadly be viewed as an offline reinforcement learning method, where the training
data is pre-collected and fixed, and there is no exploration. In contrast, online reinforcement learning methods like PPO, which require exploring new states through interaction with the environment (using the reward model as a proxy), also have their unique advantages. One of the benefits
of online reinforcement learning is that it allows the agent to continuously adapt to changes in
the environment by learning from real-time feedback. This means that, unlike offline methods,
online methods are not constrained by the static nature of pre-collected data and can discover
new problem-solving strategies. In addition, exploration can help the agent cover a wider range of
state-action pairs, thus improving generalization. This could be an important advantage for LLMs,
as generalization is considered a critical aspect in applying such large models.

4.4.3 Automatic Preference Data Generation
Although learning from human preferences is an effective and popular method for aligning LLMs,
annotating preference data is costly. Using human feedback does not only faces the problem of
limited scalability, but it may also introduce bias because human feedback is inherently subjective.
As a result, one can turn to AI feedback methods to address these scalability and consistency issues
without the limitations associated with human annotators.
As with data generation for instruction fine-tuning, generating preference data using LLMs is
straightforward. Given a set of inputs, we first use an LLM to generate pairs of outputs. Then, we
prompt the LLM to label the preference between each pair of outputs, along with its corresponding
input. Below is an example of prompting the LLM to generate a preference label for a pair of
consumer service responses.

4.4 Improved Human Preference Alignment

197

Consider a customer service scenario where a customer poses a request. You
will review two responses to this request. Please indicate which response is
preferred. Note that a good response should be courteous, clear, and concise. It
should address the customer’s concern directly, provide helpful information or a
solution, and maintain a positive tone.
Request:
Hello, I noticed that my order hasn’t arrived yet, though it was scheduled to
arrive several days ago. Could you please update me on its status? Thank you!
Response A:
I’m very sorry for the delay and understand how disappointing this can be. We’re
doing our best to sort this out quickly for you.
Response B:
Hey, stuff happens! Your package will get there when it gets there, no need to
stress.
Response A is preferred.

Once we collect such preference labels, we can use them, along with the output pair and input,
to train the reward model. Of course, we can consider demonstrating a few examples or using
advanced prompting techniques, such as CoT, to improve labeling performance. For example, we
can include in the prompt an example showing how and why one of the two responses is preferred
based on a CoT rationale.
In addition to preference labels, we can also obtain the probability associated with each label
[Lee et al., 2023]. A simple method is to extract the probabilities for the label tokens, such as “A”
and “B”, from the probabilities output by the LLM. We can then use the Softmax function or other
normalization techniques to re-normalize these probabilities into a distribution over the labels.
These probabilities of preferred labels can serve as pointwise supervision signals for training the
reward model, as discussed in Section 4.4.1.
For data generation, although it is easy to scale up, it is often necessary to ensure the data is
accurate and diverse. Here, the data quality and diversity issues involve not only the labeling of
preferences but also the inputs and outputs of the model. Therefore, we often need to use a variety
of techniques to obtain large-scale, high-quality data. For example, one can generate diverse
model outputs and annotations by using different LLMs, prompts, in-context demonstrations, and
so on [Cui et al., 2024]. Dubois et al. [2024] report that the variability in pairwise preference data
is important for training LLMs from either human or AI feedback.
While learning from AI feedback is highly scalable and generally objective, this method is
more suited to well-defined tasks where objective performance metrics are available. By contrast,
learning from human feedback is more advantageous when aligning AI systems with human values, preferences, and complex real-world tasks that require understanding of subtle or subjective
context. These methods can be combined to train LLMs that benefit from both human insights
and the scalability of AI feedback.

Alignment

198

4.4.4 Step-by-step Alignment
So far, our discussion of alignment has primarily focused on the use of reward models for evaluating entire input-output sequence pairs. These methods can be easily adapted to scenarios where
the correctness of an output can be examined by checking whether the desired result is included.
For example, in the task of calculating a mathematical expression, a reward model can provide
positive feedback if the answer is correct, and negative feedback if the answer is wrong. However, in many problems that require complex reasoning, simply examining the correctness of the
final result is insufficient for learning. Imagine a student who is only given the final answer to
a challenging math problem. Knowing whether the final answer is right or wrong does not help
the student figure out where they went wrong and how to calculate the correct answer. A better
approach would be to guide the student with a step-by-step breakdown of the problem-solving
process and encourage understanding of the underlying concepts and logic behind these steps.
In Chapter 3, we studied CoT methods to prompt LLMs to explicitly write out intermediate
steps or the reasoning process needed to reach a conclusion or solve a problem. We saw that
breaking down a problem into smaller parts could make it easier to understand the solution path
and increase the accuracy of the output. These methods can be naturally extended to the alignment
of LLMs, that is, we supervise the model during the intermediate steps of reasoning. Consider a
reasoning task where an LLM produces a sequence of reasoning steps y = {ȳ1 , ..., ȳns } for the
given input. The result of the reasoning is assumed to be included in the last step ȳns , and can
be easily verified. For this reasoning problem, Uesato et al. [2022] categorize LLM fine-tuning
approaches into two classes:

• Outcome-based Approaches. Supervision occurs only when the end result is verified. This
is a standard method for learning from human feedback we have discussed in this chapter.
For example, the LLM is optimized to maximize some form of the reward r(x, y).
• Process-based Approaches. Supervision is involved in all intermediate steps in addition to
the last step. To do this, we need to develop a model to give a supervision signal at each
step, and develop loss functions that can make use of such supervision signals.

Figure 4.11 shows two LLM outputs for an example math problem. Although the LLM gives
the correct final answer in both cases, it makes mistakes during the problem-solving process in the
second output. Outcome-based approaches overlook these mistakes and give positive feedback for
the entire solution. By contrast, process-based approaches can take these mistakes into account
and provide additional guidance on the detailed reasoning steps.
An important issue for process-based approaches is that we need to get step-level feedback
during a (potentially) long reasoning path. We can collect or generate reasoning paths corresponding to problems from existing datasets. Human experts then annotate each step in these paths for
correctness. These annotations can be used to directly train LLMS or as rewards in reward modeling. However, in practice, richer annotations are often introduced [Lightman et al., 2024]. In
addition to the correct and incorrect labels, a step can also be labeled as neutral to indicate that
while the step may be technically correct, it might still be problematic within the overall reasoning process. Furthermore, to improve the efficiency of data annotation, techniques such as active
learning can be employed. Identifying obvious errors usually does not significantly contribute to

4.4 Improved Human Preference Alignment

199

Question (from The International Junior Math Olympiad):
Jessica has a lot of apps on her phone. The number of apps she has is divisible by 5. If she
downloaded 10 more apps, she would still have less than 50 apps. What is the maximum
original number of apps Jessica could have?
Response 1:

Response 2:

ȳ1 : x must be a multiple of 5, i.e.,

x

x ≡ 0 (mod 5).

total number of apps should still be less
x
than 50: x + 10 < 50. Solving this
inequality: x < 40.
ȳ3 : Since x must be a multiple of 5 and

x

less than 40, the largest value of x is 35.

apps is 35.

x

new total will be x + 10, and this must
x
be less than 50. Therefore, x + 10 < 50.
Simplifying this gives x < 40.
ȳ3 : x must be less than or equal to

x

problematic

the largest multiple of 5, which is 40.
ȳ4 : But x + 10 should not be more

ȳ4 : Verify the result. If x = 35, then after

ȳ5 : The maximum original number of

so x must be a multiple of 5.
ȳ2 : If she downloads 10 more apps, her

ȳ2 : After downloading 10 apps, the total

downloading 10 apps, Jessica would
have: 35 + 10 = 45 which satisfy
45 < 50.

ȳ1 : The number of apps is divisible by 5,

problematic

than or equal to 50. So we need to
subtract 5 from 40.
ȳ5 : Therefore, the final result is 35.

x

x

x
x

correct

correct

Fig. 4.11: Two LLM responses to a math problem. In response 1, both the final result and all the reasoning steps are
correct. In response 2, the final result is correct, but there are mistakes in the reasoning process (highlighted in red).
For outcome-based approaches, both responses are considered correct. For process-based approaches, the mistakes in
response 2 can be considered in reward modeling.

learning from reasoning mistakes. Instead, annotating steps that the model confidently considers
correct but are actually problematic is often more effective.
Given a set of step-level annotated reasoning paths and corresponding inputs, we can train
a reward model to provide feedback for supervising policy learning. The reward model can be
treated as a classification model, and so its architecture can be a Transformer decoder with a
Softmax layer stacked on top. At step k, the reward model takes both the problem description
(denoted by x) and the reasoning steps generated so far (denoted by ȳ≤k ) as input and outputs
a probability distribution over the label set {correct, incorrect} or {correct, incorrect, neutral}.
Then the learned reward model is used to evaluate reasoning paths by assessing the correctness of
each step. A simple method to model correctness is to count the number of steps that are classified
as correct, given by
r(x, y) =

ns
X

δ(correct, C(x, ȳ≤k ))

(4.83)

k=1

where C(x, ȳ≤k ) denotes the label with the maximum probability. We can also use log-probabilities

Alignment

200
of classification to define the reward of the entire path
r(x, y) =

ns
X

log Pr(correct|x, ȳ≤k )

(4.84)

k=1

where Pr(correct|x, ȳ≤k ) denotes the probability of the correct label generated by the reward
model. The reward score r(x, y) can then be used to train the policy in RLHF as usual.
While we restrict our discussion to math problems, the approaches described here are general
and can be applied to a wide variety of tasks that involve multi-step reasoning and decisionmaking. Moreover, we can consider various aspects when assessing the quality of a step, rather
than just its correctness. For example, in dialogue systems, responses must not only be accurate
but also contextually appropriate across multiple turns of conversation. If a model provides a
correct response but fails to maintain coherence in the context of the ongoing dialogue, steplevel feedback could help the model identify and correct such discrepancies. Also note that the
process-based approaches are related to the fine-grained reward modeling approaches discussed
in Section 4.4.1.3. All these approaches essentially aim to provide more detailed supervision to
LLMs by breaking their outputs into smaller, more manageable steps. However, process-based
feedback focuses more on evaluating the correctness of a step based on its preceding steps, while
the approaches in Section 4.4.1.3 emphasize evaluating each step independently.
The idea of aligning LLMs step by step has great application potential, especially considering
the recent shift towards more complex reasoning tasks in the use of LLMs. For example, both
the GPT-o1 and GPT-o3 models are designed with more advanced reasoning techniques (such
as long internal CoT) to solve challenging problems like scientific and mathematical reasoning
[OpenAI, 2024]. These tasks often rely on long and complex reasoning paths, and therefore, it
seems essential to introduce detailed supervision signals in the reasoning process. Moreover, from
a practical perspective, effective supervision on long reasoning paths not only improves reasoning
performance, but it also helps the model eliminate redundant or unnecessary reasoning steps,
thereby reducing reasoning complexity and improving efficiency.

4.4.5 Inference-time Alignment
In this section we explored a variety of methods to align models with human preferences and annotations. However, one of the significant limitations of many such methods is that LLMs must
be fine-tuned. For RLHF and its variants, training LLMs with reward models can be computationally expensive and unstable, leading to increased complexity and costs when applying these
approaches. In this case, we can consider aligning models at inference time, thus avoiding the
additional complexity and effort involved.
One simple way to achieve inference-time alignment is to use the reward model to select
the best one from N alternative outputs generated by the LLM, a method known as Best-of-N
sampling (BoN sampling). We can consider BoN sampling as a form of reranking. In fact,
reranking methods have been widely used in NLP tasks, such as machine translation, for a long
time. They are typically applied in situations where training complex models is costly. In such
cases, directly reranking the outputs allows for the incorporation of these complex models at a
very low cost8 .
8

Reranking methods can also help us explore what are known as model errors and search errors, although these

4.5 Summary

201

In the BoN sampling process, the LLM takes the input sequence x and generates N different
output sequences {ŷ1 , ..., ŷN }:
{ŷ1 , ..., ŷN } = argTopN [Pr(y|x)]

(4.85)

y

where the argTopN operation returns the top-N outputs that maximize the function Pr(y|x).
These outputs can be generated in a variety of ways, depending on the search algorithm used by
the model (e.g., sampling or beam search). Once the N -best output candidates are generated, the
reward model is used to evaluate and select the best one:
ŷbest = max{r(x, ŷ1 ), ..., r(x, ŷN )}

(4.86)

It is worth noting that the result of BoN sampling is also influenced by the diversity of the
N -best list. This is a common issue with most reranking methods. Typically, we wish the N -best
output candidates to have relatively high quality but be sufficiently different from each other. In
many text generation systems, the N -best outputs are very similar, often differing by just one
or two words. The diversity issue is even more challenging in LLMs, as the N -best outputs
generated by an LLM can be different in their wordings, yet their semantic meanings are often
quite similar. In practice, one can adjust the model hyperparameters and/or adopt different LLMs
to generate more diverse output candidates for reranking. Nevertheless, as with many practical
systems, we need to make a trade-off between selecting high-quality candidates and ensuring
sufficient variation in the generated outputs.
BoN sampling can be used for training LLMs as well. A closely related method is rejection
sampling. In this method, we first select the “best” outputs from the N -best lists via the reward
model, and then take these selected outputs to fine-tune the LLM. In this way, we can introduce
human preferences into the training of LLMs via a much simpler approach compared to RLHF.
Many LLMs have adopted rejection sampling for fine-tuning [Nakano et al., 2021; Touvron et al.,
2023b].

4.5 Summary
In this chapter, we have explored a range of techniques for aligning LLMs. In particular, we
have discussed fine-tuning methods that enable LLMs to follow instructions and align them with
human preferences. One of the benefits of fine-tuning LLMs is computation efficiency. Unlike
pre-training based on large-scale neural network optimization, fine-tuning is a post-training step
and so is less computationally expensive. Moreover, it is better suited to address problems that are
not easily solved in pre-training, such as human value alignment. The widespread attention to the
alignment issue has also led to a surge of research papers on this topic, which has posed challenges
in writing this chapter, as it is difficult to cover all the latest techniques. However, we have tried
to provide a relatively detailed introduction to the fundamental approaches to alignment, such as
issues are not often discussed in the context of LLMs. For example, suppose we have an old model and a new, more
powerful model. We can use the new model to select the best output from the N -best list of the old model as the oracle
output. The performance difference between the oracle output and the top-1 output of the original N -best list reflects
the performance gain brought by the new model. If the performance gain is significant, we can say that the old model
has more model errors. If the gain is small, it may indicate that the issue lies in search errors, as the best candidates
were not found.

Alignment

202
instruction fine-tuning and RLHF.

While we have focused on LLM alignment techniques in this chapter, the term AI alignment
is a wide-ranging concept. It generally refers to the process of ensuring that the behavior of an AI
system aligns with human values, goals, and expectations. The idea of AI alignment can be traced
back to the early days of AI. A widely cited description of AI alignment comes from an article by
the mathematician and computer scientist Norbert Wiener [Wiener, 1960]. The quote is as follows
If we use, to achieve our purposes, a mechanical agency with whose operation
we cannot efficiently interfere ... we had better be quite sure that the purpose
put into the machine is the purpose which we really desire.
At that time, AI alignment was a distant concern for researchers. But today, it greatly influences the design of various AI systems. For example, in robotics, alignment is critical to ensuring that autonomous robots safely interact with humans and their environments. In autonomous
driving, cars must not only follow traffic laws but also make complex, real-time decisions that
prioritize human safety, avoid accidents, and navigate ethical dilemmas.
In current AI research, alignment is usually achieved by developing a surrogate objective that
is analogous to the real goal and steering the AI system towards this objective. However, designing
the objective of AI alignment is very difficult. One reason is that human values are diverse and
often context-dependent, making it difficult to distill them into a single, universally applicable
objective function. Also, the complexity of real-world environments, where values and goals often
conflict or evolve over time, further complicates alignment efforts. Even if we could define an
appropriate objective, AI systems may find unintended ways to achieve it, leading to “misaligned”
outcomes that still technically satisfy the objective but in a harmful or counterproductive way.
These challenges have motivated and are motivating AI research towards more aligned systems, either through developing new mechanisms for perceiving the world or more efficient and
generalizable methods to adapt these systems to given tasks. More importantly, as AI systems
become more powerful and intelligent, especially given that recent advances in LLMs have shown
remarkable capabilities in dealing with many challenging problems, the need for AI alignment
has become more urgent. Researchers have started to be concerned with AI safety and warn the
community that they need to develop and release AI systems with great caution to prevent these
systems from being misaligned [Russell, 2019; Bengio et al., 2024].

Bibliography
[Ainslie et al., 2020] Joshua Ainslie, Santiago Ontanon, Chris Alberti, Vaclav Cvicek, Zachary Fisher,
Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. Etc: Encoding long and
structured inputs in transformers. In Proceedings of the 2020 Conference on Empirical Methods in
Natural Language Processing (EMNLP), pages 268–284, 2020.
[Ainslie et al., 2023] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico
Lebron, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multihead checkpoints. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language
Processing, pages 4895–4901, 2023.
[Akyürek et al., 2023] Ekin Akyürek, Dale Schuurmans, Jacob Andreas, Tengyu Ma, and Denny Zhou.
What learning algorithm is in-context learning? investigations with linear models. In Proceedings of
The Eleventh International Conference on Learning Representations, 2023.
[Alabdulmohsin et al., 2022] Ibrahim M Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. Advances in Neural Information Processing Systems, 35:
22300–22312, 2022.
[Allal et al., 2024] Loubna Ben Allal, Anton Lozhkov, and Daniel van Strien. cosmopedia: how to create
large-scale synthetic data for pre-training. https://huggingface.co/blog/cosmopedia, 2024.
[Almazrouei et al., 2023] Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Mérouane Debbah, Étienne Goffinet, Daniel Hesslow, Julien Launay, Quentin
Malartic, Daniele Mazzotta, Badreddine Noune, Baptiste Pannier, and Guilherme Penedo. The falcon
series of open language models. arXiv preprint arXiv:2311.16867, 2023.
[Andreas et al., 2016] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. Neural module
networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
39–48, 2016.
[Arjovsky et al., 2016] Martin Arjovsky, Amar Shah, and Yoshua Bengio. Unitary evolution recurrent
neural networks. In International conference on machine learning, pages 1120–1128, 2016.
[Aschenbrenner, 2024] Leopold Aschenbrenner. Situational awareness: The decade ahead, 2024. URL
https://situational-awareness.ai/.
[Askell et al., 2021] Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan,
Andy Jones, Nicholas Joseph, Benjamin Mann, Nova DasSarma, Nelson Elhage, Zac Hatfield-Dodds,
Danny Hernandez, Jackson Kernion, Kamal Ndousse, Catherine Olsson, Dario Amodei, Tom B. Brown,
Jack Clark, Sam McCandlish, Chris Olah, and Jared Kaplan. A general language assistant as a laboratory
for alignment. arXiv preprint arXiv:2112.00861, 2021.
[Bach et al., 2022] Stephen H. Bach, Victor Sanh, Zheng Xin Yong, Albert Webson, Colin Raffel, Nihal V.
Nayak, Abheesht Sharma, Taewoon Kim, M. Saiful Bari, Thibault Févry, Zaid Alyafeai, Manan Dey,
Andrea Santilli, Zhiqing Sun, Srulik Ben-David, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Alan
Fries, Maged Saeed AlShaibani, Shanya Sharma, Urmish Thakker, Khalid Almubarak, Xiangru Tang,
Dragomir R. Radev, Mike Tian-Jian Jiang, and Alexander M. Rush. Promptsource: An integrated development environment and repository for natural language prompts. In Proceedings of the 60th Annual
Meeting of the Association for Computational Linguistics: System Demonstrations, pages 93–104, 2022.
[Bengio et al., 2003] Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. A neural
probabilistic language model. Journal of Machine Learning Research, 3:1137–1155, 2003.
[Bengio et al., 2006] Yoshua Bengio, Pascal Lamblin, Dan Popovici, and Hugo Larochelle. Greedy layerwise training of deep networks. Advances in neural information processing systems, 19, 2006.
[Bengio et al., 2024] Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor
Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, Gillian K. Hadfield, Jeff

203

204

Alignment

Clune, Tegan Maharaj, Frank Hutter, Atilim Gunes Baydin, Sheila A. McIlraith, Qiqi Gao, Ashwin
Acharya, David Krueger, Anca Dragan, Philip Torr, Stuart Russell, Daniel Kahneman, Jan Markus
Brauner, and Sören Mindermann. Managing extreme ai risks amid rapid progress. Science, 384(6698):
842–845, 2024.
[Bentivogli and Giampiccolo, 2011] Luisa Bentivogli and Danilo Giampiccolo. Pascal recognizing textual
entailment challenge (rte-7) at tac 2011. https://tac.nist.gov/2011/RTE/, 2011.
[Besta et al., 2024] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski,
Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, and Torsten
Hoefler. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of
the AAAI Conference on Artificial Intelligence, volume 38, pages 17682–17690, 2024.
[Biderman et al., 2021] Stella Biderman, Sid Black, Charles Foster, Leo Gao, Eric Hallahan, Horace He,
Ben Wang, and Phil Wang. Rotary embeddings: A relative revolution. https://blog.eleuther.ai/
rotary-embeddings/, 2021.
[Bishop, 2006] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
[Blum and Mitchell, 1998] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with
co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages
92–100, 1998.
[Bradley and Terry, 1952] Ralph Allan Bradley and Milton E. Terry. Rank analysis of incomplete block
designs: I. the method of paired comparisons. Biometrika, 39(3/4):324–345, 1952.
[Brandon et al., 2024] William Brandon, Mayank Mishra, Aniruddha Nrusimha, Rameswar Panda, and
Jonathan Ragan Kelly. Reducing transformer key-value cache size with cross-layer attention. arXiv
preprint arXiv:2405.12981, 2024.
[Brill, 1992] Eric Brill. A simple rule-based part of speech tagger. In Speech and Natural Language:
Proceedings of a Workshop Held at Harriman, New York, February 23-26, 1992, 1992.
[Brown et al., 1993] Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, and Robert L. Mercer.
The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics,
19(2):263–311, 1993.
[Brown et al., 2020] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla
Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel
Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey
Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess,
Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei.
Language models are few-shot learners. Advances in neural information processing systems, 33:1877–
1901, 2020.
[Bubeck et al., 2023] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric
Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott M. Lundberg, Harsha Nori, Hamid
Palangi, Marco Túlio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments
with gpt-4. arXiv preprint arXiv:2303.12712, 2023.
[Bulatov et al., 2022] Aydar Bulatov, Yury Kuratov, and Mikhail Burtsev. Recurrent memory transformer.
Advances in Neural Information Processing Systems, 35:11079–11091, 2022.
[Burges et al., 2005] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton,
and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the 22nd international
conference on Machine learning, pages 89–96, 2005.
[Burns et al., 2023] Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold
Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, and Jeff Wu.
Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. arXiv preprint
arXiv:2312.09390, 2023a.

4.5 Summary

205

[Burns et al., 2023] Collin Burns, Jan Leike, Leopold Aschenbrenner, Jeffrey Wu, Pavel Izmailov, Leo
Gao, Bowen Baker, and Jan Hendrik Kirchner. Weak-to-strong generalization, 2023b. URL https://
https://openai.com/index/weak-to-strong-generalization.
[Caballero et al., 2023] Ethan Caballero, Kshitij Gupta, Irina Rish, and David Krueger. Broken neural
scaling laws. In ICLR 2023 Workshop on Mathematical and Empirical Understanding of Foundation
Models, 2023.
[Cao et al., 2007] Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. Learning to rank: from
pairwise approach to listwise approach. In Proceedings of the 24th international conference on Machine
learning, pages 129–136, 2007.
[Chang et al., 2024] Kaiyan Chang, Songcheng Xu, Chenglong Wang, Yingfeng Luo, Tong Xiao, and
Jingbo Zhu. Efficient prompting methods for large language models: A survey. arXiv preprint
arXiv:2404.01077, 2024.
[Charniak, 1997] Eugene Charniak. Statistical parsing with a context-free grammar and word statistics.
AAAI/IAAI, 2005(598-603):18, 1997.
[Chen et al., 2023] Banghao Chen, Zhaofeng Zhang, Nicolas Langrené, and Shengxin Zhu. Unleashing
the potential of prompt engineering in large language models: a comprehensive review. arXiv preprint
arXiv:2310.14735, 2023a.
[Chen et al., 2023] Lichang Chen, Shiyang Li, Jun Yan, Hai Wang, Kalpa Gunaratna, Vikas Yadav, Zheng
Tang, Vijay Srinivasan, Tianyi Zhou, Heng Huang, and Hongxia Jin. Alpagasus: Training a better alpaca
with fewer data. arXiv preprint arXiv:2307.08701, 2023b.
[Chen et al., 2024] Lichang Chen, Shiyang Li, Jun Yan, Hai Wang, Kalpa Gunaratna, Vikas Yadav, Zheng
Tang, Vijay Srinivasan, Tianyi Zhou, Heng Huang, and Hongxia Jin. Alpagasus: Training a better alpaca
with fewer data. In The Twelfth International Conference on Learning Representations, 2024a.
[Chen et al., 2023] Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. Extending
context window of large language models via positional interpolation. arXiv preprint arXiv:2306.15595,
2023c.
[Chen et al., 2020] Tianlong Chen, Jonathan Frankle, Shiyu Chang, Sijia Liu, Yang Zhang, Zhangyang
Wang, and Michael Carbin. The lottery ticket hypothesis for pre-trained bert networks. Advances in
neural information processing systems, 33:15834–15846, 2020.
[Chen et al., 2024] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play finetuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335,
2024b.
[Chevalier et al., 2023] Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. Adapting
language models to compress contexts. In Proceedings of the 2023 Conference on Empirical Methods
in Natural Language Processing, pages 3829–3846, 2023.
[Chi et al., 2022] Ta-Chung Chi, Ting-Han Fan, Peter J Ramadge, and Alexander Rudnicky. Kerple:
Kernelized relative positional embedding for length extrapolation. Advances in Neural Information Processing Systems, 35:8386–8399, 2022.
[Chi et al., 2023] Ta-Chung Chi, Ting-Han Fan, Alexander Rudnicky, and Peter Ramadge. Dissecting
transformer length extrapolation via the lens of receptive field analysis. In Proceedings of the 61st Annual
Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13522–13537,
2023.
[Chiang et al., 2023] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna:
An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023. URL https://
lmsys.org/blog/2023-03-30-vicuna/.
[Chowdhery et al., 2022] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav

206

Alignment

Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker
Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam
Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury,
Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus,
Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan
Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov,
Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta,
Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. Palm: Scaling
language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.
[Christiano et al., 2017] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario
Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017.
[Chu et al., 2023] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang,
Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. A survey of chain of thought reasoning: Advances,
frontiers and future. arXiv preprint arXiv:2309.15402, 2023.
[Chung et al., 2022] Hyung Won Chung, Le Hou, S. Longpre, Barret Zoph, Yi Tay, William Fedus,
Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu,
Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Wei Yu, Vincent Zhao, Yanping Huang, Andrew M. Dai, Hongkun Yu, Slav Petrov,
Ed Huai hsin Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei.
Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.
[Clark et al., 2019] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra:
Pre-training text encoders as discriminators rather than generators. In Proceedings of International
Conference on Learning Representations, 2019.
[Cobbe et al., 2021] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz
Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John
Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021.
[Conneau et al., 2020] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Édouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov.
Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting
of the Association for Computational Linguistics, pages 8440–8451, 2020.
[Coste et al., 2024] Thomas Coste, Usman Anwar, Robert Kirk, and David Krueger. Reward model ensembles help mitigate overoptimization. In The Twelfth International Conference on Learning Representations, 2024.
[Cui et al., 2024] Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan
Ni, Guotong Xie, Ruobing Xie, Yankai Lin, Zhiyuan Liu, and Maosong Sun. ULTRAFEEDBACK:
Boosting language models with scaled AI feedback. In Proceedings of the 41st International Conference
on Machine Learning, volume 235, pages 9722–9744, 2024.
[Dai et al., 2023] Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Shuming Ma, Zhifang Sui, and Furu Wei.
Why can gpt learn in-context? language models secretly perform gradient descent as meta-optimizers.
In Findings of the Association for Computational Linguistics: ACL 2023, pages 4005–4019, 2023.
[Dai et al., 2019] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G Carbonell, Quoc Le, and Ruslan
Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2978–2988,
2019.
[Dao et al., 2022] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast
and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing

4.5 Summary

207

Systems, 35:16344–16359, 2022.
[Dehghani et al., 2018] Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz
Kaiser. Universal transformers. arXiv preprint arXiv:1807.03819, 2018.
[Deletang et al., 2024] Gregoire Deletang, Anian Ruoss, Paul-Ambroise Duquenne, Elliot Catt, Tim
Genewein, Christopher Mattern, Jordi Grau-Moya, Li Kevin Wenliang, Matthew Aitchison, Laurent
Orseau, Marcus Hutter, and Joel Veness. Language modeling is compression. In The Twelfth International Conference on Learning Representations, 2024.
[Deng et al., 2022] Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu,
Meng Song, Eric Xing, and Zhiting Hu. Rlprompt: Optimizing discrete text prompts with reinforcement
learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing,
pages 3369–3391, 2022.
[Devlin et al., 2019] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pretraining of deep bidirectional transformers for language understanding. In Proceedings of the 2019
Conference of the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, 2019.
[Ding et al., 2024] Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang
Xu, Fan Yang, and Mao Yang. Longrope: Extending llm context window beyond 2 million tokens. arXiv
preprint arXiv:2402.13753, 2024.
[Dolan and Brockett, 2005] Bill Dolan and Chris Brockett. Automatically constructing a corpus of sentential paraphrases. In Proceedings of Third International Workshop on Paraphrasing (IWP2005), 2005.
[Dong et al., 2019] Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng
Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language
understanding and generation. Advances in neural information processing systems, 32, 2019.
[Dong et al., 2022] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun,
Jingjing Xu, and Zhifang Sui. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022.
[Dong et al., 2021] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you
need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine
Learning, pages 2793–2803. PMLR, 2021.
[Drozdov et al., 2022] Andrew Drozdov, Nathanael Schärli, Ekin Akyürek, Nathan Scales, Xinying Song,
Xinyun Chen, Olivier Bousquet, and Denny Zhou. Compositional semantic parsing with large language
models. In Proceedings of The Eleventh International Conference on Learning Representations, 2022.
[Dua et al., 2022] Dheeru Dua, Shivanshu Gupta, Sameer Singh, and Matt Gardner. Successive prompting
for decomposing complex questions. In Proceedings of the 2022 Conference on Empirical Methods in
Natural Language Processing, pages 1251–1265, 2022.
[Dubey et al., 2024] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad AlDahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of
models. arXiv preprint arXiv:2407.21783, 2024.
[Dubois et al., 2024] Yann Dubois, Chen Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy
Ba, Carlos Guestrin, Percy S Liang, and Tatsunori B Hashimoto. Alpacafarm: A simulation framework
for methods that learn from human feedback. Advances in Neural Information Processing Systems, 36,
2024.
[Eisenstein et al., 2023] Jacob Eisenstein, Chirag Nagpal, Alekh Agarwal, Ahmad Beirami, Alex D’Amour,
DJ Dvijotham, Adam Fisch, Katherine Heller, Stephen Pfohl, Deepak Ramachandran, and Peter Shaw.
Helping or herding? reward model ensembles mitigate but do not eliminate reward hacking. arXiv
preprint arXiv:2312.09244, 2023.
[Elsken et al., 2019] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search:
A survey. Journal of Machine Learning Research, 20(55):1–21, 2019.

208

Alignment

[Erhan et al., 2010] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does
unsupervised pre-training help deep learning? In Proceedings of the thirteenth international conference
on artificial intelligence and statistics, pages 201–208, 2010.
[Fan et al., 2019] Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand
with structured dropout. In Proceedings of International Conference on Learning Representations, 2019.
[Fedus et al., 2022] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to
trillion parameter models with simple and efficient sparsity. The Journal of Machine Learning Research,
23(1):5232–5270, 2022.
[Fernandes et al., 2023] Patrick Fernandes, Aman Madaan, Emmy Liu, António Farinhas, Pedro Henrique
Martins, Amanda Bertsch, José G. C. de Souza, Shuyan Zhou, Tongshuang Wu, Graham Neubig, and
André F. T. Martins. Bridging the gap: A survey on integrating (human) feedback for natural language
generation. Transactions of the Association for Computational Linguistics, 11:1643–1668, 2023.
[Franklin and Graesser, 1996] Stan Franklin and Art Graesser. Is it an agent, or just a program?: A taxonomy for autonomous agents. In International workshop on agent theories, architectures, and languages,
pages 21–35. Springer, 1996.
[Frensch and Funke, 2014] Peter A Frensch and Joachim Funke. Complex problem solving: The European
perspective. Psychology Press, 2014.
[Gale et al., 2019] Trevor Gale, Erich Elsen, and Sara Hooker. The state of sparsity in deep neural networks.
arXiv preprint arXiv:1902.09574, 2019.
[Ganguli et al., 2023] Deep Ganguli, Amanda Askell, Nicholas Schiefer, Thomas I. Liao, Kamile Lukosiute, Anna Chen, Anna Goldie, Azalia Mirhoseini, Catherine Olsson, Danny Hernandez, Dawn Drain,
Dustin Li, Eli Tran-Johnson, Ethan Perez, Jackson Kernion, Jamie Kerr, Jared Mueller, Joshua Landau,
Kamal Ndousse, Karina Nguyen, Liane Lovitt, Michael Sellitto, Nelson Elhage, Noemí Mercado, Nova
DasSarma, Oliver Rausch, Robert Lasenby, Robin Larson, Sam Ringer, Sandipan Kundu, Saurav Kadavath, Scott Johnston, Shauna Kravec, Sheer El Showk, Tamera Lanham, Timothy Telleen-Lawton, Tom
Henighan, Tristan Hume, Yuntao Bai, Zac Hatfield-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph,
Sam McCandlish, Tom Brown, Christopher Olah, Jack Clark, Samuel R. Bowman, and Jared Kaplan.
The capacity for moral self-correction in large language models. arXiv preprint arXiv:2302.07459, 2023.
[Gao et al., 2023] Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pages 10835–10866. PMLR, 2023a.
[Gao et al., 2023] Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie
Callan, and Graham Neubig. Pal: Program-aided language models. In International Conference on
Machine Learning, pages 10764–10799. PMLR, 2023b.
[Gao et al., 2023] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei
Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv
preprint arXiv:2312.10997, 2023c.
[Garg et al., 2022] Shivam Garg, Dimitris Tsipras, Percy S Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes. Advances in Neural Information
Processing Systems, 35:30583–30598, 2022.
[Ge et al., 2024] Yuan Ge, Yilun Liu, Chi Hu, Weibin Meng, Shimin Tao, Xiaofeng Zhao, Hongxia
Ma, Li Zhang, Boxing Chen, Hao Yang, Bei Li, Tong Xiao, and Jingbo Zhu. Clustering and ranking: Diversity-preserved instruction selection through expert-aligned quality estimation. arXiv preprint
arXiv:2402.18191, 2024.
[Gemma Team, 2024] Google DeepMind Gemma Team. Gemma: Open Models Based on Gemini Research and Technology, 2024.
[Goodhart, 1984] Charles AE Goodhart. Problems of monetary management: the UK experience. Springer,
1984.

4.5 Summary

209

[Gordon et al., 2021] Mitchell A Gordon, Kevin Duh, and Jared Kaplan. Data and parameter scaling laws
for neural machine translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural
Language Processing, pages 5915–5922, 2021.
[Gu and Dao, 2023] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state
spaces. arXiv preprint arXiv:2312.00752, 2023.
[Gunasekar et al., 2023] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del
Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil
Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman
Kalai, Yin Tat Lee, and Yuanzhi Li. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023.
[Guo et al., 2024] Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang
Bian, and Yujiu Yang. Connecting large language models with evolutionary algorithms yields powerful
prompt optimizers. In The Twelfth International Conference on Learning Representations, 2024.
[Gupta and Berant, 2020] Ankit Gupta and Jonathan Berant. Gmat: Global memory augmentation for
transformers. arXiv preprint arXiv:2006.03274, 2020.
[Gupta et al., 2021] Ankit Gupta, Guy Dar, Shaya Goodman, David Ciprut, and Jonathan Berant. Memoryefficient transformers via top-k attention. In Proceedings of the Second Workshop on Simple and Efficient
Natural Language Processing, pages 39–52, 2021.
[Han et al., 2021] Xu Han, Zhengyan Zhang, Ning Ding, Yuxian Gu, Xiao Liu, Yuqi Huo, Jiezhong Qiu,
Liang Zhang, Wentao Han, Minlie Huang, Qin Jin, Yanyan Lan, Yang Liu, Zhiyuan Liu, Zhiwu Lu,
Xipeng Qiu, Ruihua Song, Jie Tang, Ji-Rong Wen, Jinhui Yuan, Wayne Xin Zhao, and Jun Zhu. Pretrained models: Past, present and future. AI Open, 2:225–250, 2021.
[Han et al., 2024] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient
fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024.
[Harlap et al., 2018] Aaron Harlap, Deepak Narayanan, Amar Phanishayee, Vivek Seshadri, Nikhil Devanur, Greg Ganger, and Phil Gibbons. Pipedream: Fast and efficient pipeline parallel dnn training. arXiv
preprint arXiv:1806.03377, 2018.
[He et al., 2019] Kaiming He, Ross Girshick, and Piotr Dollár. Rethinking imagenet pre-training. In
Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4918–4927, 2019.
[He et al., 2021] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decodingenhanced bert with disentangled attention. In Proceedings of International Conference on Learning
Representations, 2021.
[Hendrycks and Gimpel, 2016] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus).
arXiv preprint arXiv:1606.08415, 2016.
[Hendrycks et al., 2020] Dan Hendrycks, Xiaoyuan Liu, Eric Wallace, Adam Dziedzic, Rishabh Krishnan,
and Dawn Song. Pretrained transformers improve out-of-distribution robustness. In Proceedings of the
58th Annual Meeting of the Association for Computational Linguistics, pages 2744–2751, 2020.
[Hendrycks et al., 2021] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn
Song, and Jacob Steinhardt. Measuring massive multitask language understanding. In Proceedings of
International Conference on Learning Representations, 2021.
[Hestness et al., 2017] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun,
Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is
predictable, empirically. arXiv preprint arXiv:1712.00409, 2017.
[Hewitt, 2024] John Hewitt. Instruction following without instruction tuning, 2024. URL https://nlp.
stanford.edu/~johnhew/instruction-following.html.
[Hewitt et al., 2024] John Hewitt, Nelson F Liu, Percy Liang, and Christopher D Manning. Instruction
following without instruction tuning. arXiv preprint arXiv:2409.14254, 2024.

210

Alignment

[Hochreiter and Schmidhuber, 1997] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.
Neural computation, 9(8):1735–1780, 1997.
[Hoffmann et al., 2022] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor
Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom
Hennigan, Eric Noland, Katie Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Jack W. Rae, Oriol Vinyals, and Laurent Sifre. Training
compute-optimal large language models. arXiv preprint arXiv:2203.15556, 2022.
[Honovich et al., 2023] Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language models with (almost) no human labor. In Proceedings of the 61st Annual Meeting
of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14409–14428, 2023.
[Houlsby et al., 2019] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin
De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer
learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, pages
2790–2799. PMLR, 2019.
[Hu et al., 2022] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang,
Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International
Conference on Learning Representations, 2022.
[Huang, 2009] Liang Huang. Dynamic programming-based search algorithms in NLP. In Proceedings
of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the
Association for Computational Linguistics, Companion Volume: Tutorial Abstracts, 2009.
[Huang et al., 2019] Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan Firat, Mia Xu Chen, Dehao
Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V Le, Yonghui Wu, and Zhifeng Chen. Gpipe: Efficient
training of giant neural networks using pipeline parallelism. Advances in neural information processing
systems, 32, 2019.
[Hutchins et al., 2022] DeLesley Hutchins, Imanol Schlag, Yuhuai Wu, Ethan Dyer, and Behnam
Neyshabur. Block-recurrent transformers. Advances in neural information processing systems, 35:
33248–33261, 2022.
[Jelinek, 1998] Frederick Jelinek. Statistical methods for speech recognition. MIT Press, 1998.
[Jiang et al., 2023] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile
Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril,
Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b. arXiv preprint arXiv:2310.06825,
2023a.
[Jiang et al., 2023] Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin, Yuqing Yang, and Lili Qiu. Llmlingua:
Compressing prompts for accelerated inference of large language models. In Proceedings of the 2023
Conference on Empirical Methods in Natural Language Processing, pages 13358–13376, 2023b.
[Jiang et al., 2020] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what
language models know? Transactions of the Association for Computational Linguistics, 8:423–438,
2020.
[Jiao et al., 2020] Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang,
and Qun Liu. Tinybert: Distilling bert for natural language understanding. In Findings of the Association
for Computational Linguistics: EMNLP 2020, pages 4163–4174, 2020.
[Joshi et al., 2017] Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large
scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601–
1611, 2017.
[Joshi et al., 2020] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer
Levy. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the

4.5 Summary

211

association for computational linguistics, 8:64–77, 2020.
[Jurafsky and Martin, 2008] Dan Jurafsky and James H. Martin. Speech and Language Processing (2nd
ed.). Prentice Hall, 2008.
[Kahneman, 2011] Daniel Kahneman. Thinking, fast and slow. macmillan, 2011.
[Kaplan et al., 2020] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language
models. arXiv preprint arXiv:2001.08361, 2020.
[Katharopoulos et al., 2020] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret.
Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference
on machine learning, pages 5156–5165. PMLR, 2020.
[Khandelwal et al., 2020] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike
Lewis. Generalization through memorization: Nearest neighbor language models. In International
Conference on Learning Representations, 2020.
[Khot et al., 2023] Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark,
and Ashish Sabharwal. Decomposed prompting: A modular approach for solving complex tasks. In
Proceedings of The Eleventh International Conference on Learning Representations, 2023.
[Kim et al., 2023] Sehoon Kim, Coleman Hooper, Thanakul Wattanawong, Minwoo Kang, Ruohan
Yan, Hasan Genc, Grace Dinh, Qijing Huang, Kurt Keutzer, Michael W. Mahoney, Yakun Sophia
Shao, and Amir Gholami. Full stack optimization of transformer inference: a survey. arXiv preprint
arXiv:2302.14017, 2023.
[Kirkpatrick et al., 2017] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume
Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska,
Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic
forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521–3526,
2017.
[Koehn, 2010] Philipp Koehn. Statistical Machine Translation. Cambridge University Press, 2010.
[Kojima et al., 2022] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke
Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing
systems, 35:22199–22213, 2022.
[Korthikanti et al., 2023] Vijay Anand Korthikanti, Jared Casper, Sangkug Lym, Lawrence McAfee,
Michael Andersch, Mohammad Shoeybi, and Bryan Catanzaro. Reducing activation recomputation in
large transformer models. Proceedings of Machine Learning and Systems, 5, 2023.
[Krakovna et al., 2020] Victoria Krakovna, Jonathan Uesato, Vladimir Mikulik, Matthew
Rahtz, Tom Everitt, Ramana Kumar, Zac Kenton, Jan Leike, and Shane Legg.
Specification gaming: the flip side of ai ingenuity. https://deepmind.google/discover/blog/
specification-gaming-the-flip-side-of-ai-ingenuity, 2020.
[Kung and Peng, 2023] Po-Nien Kung and Nanyun Peng. Do models really learn to follow instructions?
an empirical study of instruction tuning. arXiv preprint arXiv:2305.11383, 2023.
[Kwon et al., 2023] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao
Yu, Joseph E Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language
model serving with pagedattention. arXiv preprint arXiv:2309.06180, 2023.
[Lake and Baroni, 2018] Brenden Lake and Marco Baroni. Generalization without systematicity: On
the compositional skills of sequence-to-sequence recurrent networks. In International conference on
machine learning, pages 2873–2882. PMLR, 2018.
[Lambert et al., 2024] Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen
Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, Noah A. Smith, and Hannaneh Hajishirzi. Rewardbench: Evaluating reward models for language modeling. arXiv preprint

212

Alignment

arXiv:2403.13787, 2024.
[Lample and Conneau, 2019] Guillaume Lample and Alexis Conneau. Cross-lingual language model
pretraining. arXiv preprint arXiv:1901.07291, 2019.
[Lan et al., 2020] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and
Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. In Proceedings
of International Conference on Learning Representations, 2020.
[Lee et al., 2023] Harrison Lee, Samrat Phatale, Hassan Mansoor, Kellie Ren Lu, Thomas Mesnard, Johan
Ferret, Colton Bishop, Ethan Hall, Victor Carbune, and Abhinav Rastogi. Rlaif: Scaling reinforcement
learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267, 2023.
[Lester et al., 2021] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameterefficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, 2021.
[Lewis et al., 2020] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence
pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th
Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, 2020.
[Li et al., 2023] Bei Li, Rui Wang, Junliang Guo, Kaitao Song, Xu Tan, Hany Hassan, Arul Menezes, Tong
Xiao, Jiang Bian, and JingBo Zhu. Deliberate then generate: Enhanced prompting framework for text
generation. arXiv preprint arXiv:2305.19835, 2023a.
[Li, 2011] Hang Li. Learning to Rank for Information Retrieval and Natural Language Processing. Online
access: Morgan & Claypool Synthesis Collection Five. Morgan & Claypool Publishers, 2011. ISBN
9781608457076.
[Li et al., 2022] Huayang Li, Yixuan Su, Deng Cai, Yan Wang, and Lemao Liu. A survey on retrievalaugmented text generation. arXiv preprint arXiv:2202.01110, 2022.
[Li et al., 2024] Shanda Li, Chong You, Guru Guruganesh, Joshua Ainslie, Santiago Ontanon, Manzil
Zaheer, Sumit Sanghai, Yiming Yang, Sanjiv Kumar, and Srinadh Bhojanapalli. Functional interpolation
for relative positions improves long context transformers. In The Twelfth International Conference on
Learning Representations, 2024.
[Li et al., 2023] Shenggui Li, Fuzhao Xue, Chaitanya Baranwal, Yongbin Li, and Yang You. Sequence
parallelism: Long sequence training from system perspective. In Proceedings of the 61st Annual Meeting
of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2391–2404, 2023b.
[Li and Liang, 2021] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for
generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics
and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers),
pages 4582–4597, 2021.
[Li, 2023] Yinheng Li. A practical survey on zero-shot prompt design for in-context learning. In Proceedings of the 14th International Conference on Recent Advances in Natural Language Processing, pages
641–647, 2023.
[Li et al., 2023] Yucheng Li, Bo Dong, Frank Guerin, and Chenghua Lin. Compressing context to enhance
inference efficiency of large language models. In Proceedings of the 2023 Conference on Empirical
Methods in Natural Language Processing, pages 6342–6353, 2023c.
[Lialin et al., 2023] Vladislav Lialin, Vijeta Deshpande, and Anna Rumshisky. Scaling down to scale up:
A guide to parameter-efficient fine-tuning. arXiv preprint arXiv:2303.15647, 2023.
[Lightman et al., 2024] Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker,
Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The
Twelfth International Conference on Learning Representations, 2024.
[Liu et al., 2024] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang

4.5 Summary

213

Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint
arXiv:2412.19437, 2024a.
[Liu et al., 2022] Jiachang Liu, Dinghan Shen, Yizhe Zhang, William B Dolan, Lawrence Carin, and
Weizhu Chen. What makes good in-context examples for gpt-3? In Proceedings of Deep Learning Inside
Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning
Architectures, pages 100–114, 2022.
[Liu et al., 2023] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham
Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language
processing. ACM Computing Surveys, 55(9):1–35, 2023a.
[Liu et al., 2024] Tianqi Liu, Yao Zhao, Rishabh Joshi, Misha Khalman, Mohammad Saleh, Peter J Liu, and
Jialu Liu. Statistical rejection sampling improves preference optimization. In The Twelfth International
Conference on Learning Representations, 2024b.
[Liu, 2009] Tie-Yan Liu. Learning to rank for information retrieval. Foundations and Trends® in Information Retrieval, 3(3):225–331, 2009.
[Liu et al., 2023] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie
Tang. Gpt understands, too. AI Open, 2023b.
[Liu et al., 2023] Xiaoxia Liu, Jingyi Wang, Jun Sun, Xiaohan Yuan, Guoliang Dong, Peng Di, Wenhai
Wang, and Dongxia Wang. Prompting frameworks for large language models: A survey. arXiv preprint
arXiv:2311.12785, 2023c.
[Liu et al., 2024] Xinyu Liu, Runsong Zhao, Pengcheng Huang, Chunyang Xiao, Bei Li, Jingang Wang,
Tong Xiao, and Jingbo Zhu. Forgetting curve: A reliable method for evaluating memorization capability
for long-context models. In Proceedings of the 2024 Conference on Empirical Methods in Natural
Language Processing, pages 4667–4682, 2024c.
[Liu et al., 2019] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy,
Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining
approach. arXiv preprint arXiv:1907.11692, 2019.
[Longpre et al., 2023] Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay,
Denny Zhou, Quoc V. Le, Barret Zoph, Jason Wei, and Adam Roberts. The flan collection: Designing
data and methods for effective instruction tuning. In International Conference on Machine Learning,
pages 22631–22648. PMLR, 2023.
[Ma et al., 2023] Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig,
Jonathan May, and Luke Zettlemoyer. Mega: Moving average equipped gated attention. In The Eleventh
International Conference on Learning Representations, 2023.
[Ma et al., 2024] Xuezhe Ma, Xiaomeng Yang, Wenhan Xiong, Beidi Chen, Lili Yu, Hao Zhang, Jonathan
May, Luke Zettlemoyer, Omer Levy, and Chunting Zhou. Megalodon: Efficient llm pretraining and
inference with unlimited context length. arXiv preprint arXiv:2404.08801, 2024.
[Madaan et al., 2024] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao,
Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark.
Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36, 2024.
[Manning, 2022] Christopher D Manning. Human language understanding & reasoning. Daedalus, 151
(2):127–138, 2022.
[Marcus, 1993] Gary F Marcus. Negative evidence in language acquisition. Cognition, 46(1):53–85, 1993.
[Martins et al., 2022] Pedro Henrique Martins, Zita Marinho, and André FT Martins. ∞-former: Infinite
memory transformer-former: Infinite memory transformer. In Proceedings of the 60th Annual Meeting
of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5468–5485, 2022.

214

Alignment

[Mavi et al., 2024] Vaibhav Mavi, Anubhav Jangra, and Adam Jatowt. Multi-hop question answering.
Foundations and Trends® in Information Retrieval, 17(5):457–586, 2024.
[Michel et al., 2019] Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than
one? Advances in neural information processing systems, 32, 2019.
[Micikevicius et al., 2018] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich
Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, and Hao
Wu. Mixed precision training. In Proceedings of International Conference on Learning Representations,
2018.
[Miettinen, 1999] Kaisa Miettinen. Nonlinear multiobjective optimization, volume 12. Springer Science
& Business Media, 1999.
[Mikolov et al., 2013] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation
of word representations in vector space. In Proceedings of the International Conference on Learning
Representations (ICLR 2013), 2013a.
[Mikolov et al., 2013] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. Distributed representations of words and phrases and their compositionality. In Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2, pages 3111–3119, 2013b.
[Min et al., 2019] Sewon Min, Victor Zhong, Luke Zettlemoyer, and Hannaneh Hajishirzi. Multi-hop reading comprehension through question decomposition and rescoring. In Proceedings of the 57th Annual
Meeting of the Association for Computational Linguistics, pages 6097–6109, 2019.
[Minaee et al., 2024] Shervin Minaee, Tomas Mikolov, Narjes Nikzad, Meysam Chenaghlu, Richard
Socher, Xavier Amatriain, and Jianfeng Gao. Large language models: A survey. arXiv preprint
arXiv:2402.06196, 2024.
[Mishra et al., 2022] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Crosstask generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual
Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3470–3487,
2022.
[Mnih et al., 2016] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Tim Harley,
Timothy P Lillicrap, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In Proceedings of the 33rd International Conference on International Conference on
Machine Learning, pages 1928–1937, 2016.
[Mohtashami and Jaggi, 2024] Amirkeivan Mohtashami and Martin Jaggi. Random-access infinite context
length for transformers. Advances in Neural Information Processing Systems, 36, 2024.
[Mu et al., 2024] Jesse Mu, Xiang Li, and Noah Goodman. Learning to compress prompts with gist tokens.
Advances in Neural Information Processing Systems, 36, 2024.
[Munkhdalai et al., 2024] Tsendsuren Munkhdalai, Manaal Faruqui, and Siddharth Gopal. Leave no context
behind: Efficient infinite context transformers with infini-attention. arXiv preprint arXiv:2404.07143,
2024.
[Nakano et al., 2021] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina
Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna
Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman.
Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332,
2021.
[Narayanan et al., 2021] Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley,
Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan
Catanzaro, Amar Phanishayee, and Matei Zaharia. Efficient large-scale language model training on
gpu clusters using megatron-lm. In Proceedings of the International Conference for High Performance
Computing, Networking, Storage and Analysis, pages 1–15, 2021.

4.5 Summary

215

[Ng et al., 1999] Andrew Y Ng, Daishi Harada, and Stuart J Russell. Policy invariance under reward
transformations: Theory and application to reward shaping. In Proceedings of the Sixteenth International
Conference on Machine Learning, pages 278–287, 1999.
[OpenAI, 2024] OpenAI. Learning to reason with llms, September 2024. URL https://openai.com/
index/learning-to-reason-with-llms/.
[Ouyang et al., 2022] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela
Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton,
Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan
Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. Advances
in Neural Information Processing Systems, 35:27730–27744, 2022.
[Pal et al., 2023] Koyena Pal, Jiuding Sun, Andrew Yuan, Byron C Wallace, and David Bau. Future lens:
Anticipating subsequent tokens from a single hidden state. In Proceedings of the 27th Conference on
Computational Natural Language Learning (CoNLL), pages 548–560, 2023.
[Pan et al., 2022] Alexander Pan, Kush Bhatia, and Jacob Steinhardt. The effects of reward misspecification: Mapping and mitigating misaligned models. In International Conference on Learning Representations, 2022.
[Pan et al., 2024] Liangming Pan, Michael Saxon, Wenda Xu, Deepak Nathani, Xinyi Wang, and
William Yang Wang. Automatically correcting large language models: Surveying the landscape of
diverse automated correction strategies. Transactions of the Association for Computational Linguistics,
12:484–506, 2024.
[Parisi et al., 2022] Aaron Parisi, Yao Zhao, and Noah Fiedel. Talm: Tool augmented language models.
arXiv preprint arXiv:2205.12255, 2022.
[Parisi et al., 2019] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter.
Continual lifelong learning with neural networks: A review. Neural networks, 113:54–71, 2019.
[Parmar et al., 2018] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer,
Alexander Ku, and Dustin Tran. Image transformer. In International conference on machine learning, pages 4055–4064. PMLR, 2018.
[Penedo et al., 2023] Guilherme Penedo, Quentin Malartic, Daniel Hesslow, Ruxandra Cojocaru, Alessandro Cappelli, Hamza Alobeidli, Baptiste Pannier, Ebtesam Almazrouei, and Julien Launay. The refinedweb dataset for falcon llm: outperforming curated corpora with web data, and web data only. arXiv
preprint arXiv:2306.01116, 2023.
[Peng et al., 2024] Bowen Peng, Jeffrey Quesnelle, Honglu Fan, and Enrico Shippole. YaRN: Efficient context window extension of large language models. In The Twelfth International Conference on Learning
Representations, 2024.
[Pennington et al., 2014] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global
vectors for word representation. In Proceedings of Empirical Methods in Natural Language Processing
(EMNLP), pages 1532–1543, 2014.
[Peters et al., 2018] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark,
Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In Proceedings of the
2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), 2018.
[Plackett, 1975] Robin L Plackett. The analysis of permutations. Journal of the Royal Statistical Society
Series C: Applied Statistics, 24(2):193–202, 1975.
[Prasad et al., 2023] Archiki Prasad, Peter Hase, Xiang Zhou, and Mohit Bansal. Grips: Gradient-free, editbased instruction search for prompting large language models. In Proceedings of the 17th Conference of
the European Chapter of the Association for Computational Linguistics, pages 3845–3864, 2023.

216

Alignment

[Press et al., 2022] Ofir Press, Noah Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. In Proceedings of International Conference on Learning
Representations, 2022.
[Press et al., 2023] Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis.
Measuring and narrowing the compositionality gap in language models. In Findings of the Association
for Computational Linguistics: EMNLP 2023, pages 5687–5711, 2023.
[Pryzant et al., 2023] Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng.
Automatic prompt optimization with "gradient descent" and beam search. In The 2023 Conference on
Empirical Methods in Natural Language Processing, 2023.
[Qiu et al., 2020] Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang.
Pre-trained models for natural language processing: A survey. Science China Technological Sciences,
63(10):1872–1897, 2020.
[Radford et al., 2018] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving
language understanding by generative pre-training. OpenAI, 2018.
[Radford et al., 2019] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya
Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 1(8), 2019.
[Radford et al., 2021] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya
Sutskever. Learning transferable visual models from natural language supervision. In International
conference on machine learning, pages 8748–8763. PMLR, 2021.
[Rae et al., 2019] Jack W Rae, Anna Potapenko, Siddhant M Jayakumar, Chloe Hillier, and Timothy P
Lillicrap. Compressive transformers for long-range sequence modelling. In International Conference on
Learning Representations, 2019.
[Rafailov et al., 2024] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano
Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward
model. Advances in Neural Information Processing Systems, 36, 2024.
[Raffel et al., 2020] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael
Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified
text-to-text transformer. Journal of Machine Learning Research, 21(140):1–67, 2020.
[Ramachandran et al., 2017] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation
functions. arXiv preprint arXiv:1710.05941, 2017.
[Rolnick et al., 2019] David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory
Wayne. Experience replay for continual learning. Advances in Neural Information Processing Systems,
32, 2019.
[Rosenfeld et al., 2020] Jonathan S Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales. In Proceedings of International Conference
on Learning Representations, 2020.
[Ruan et al., 2024] Junhao Ruan, Long Meng, Weiqiao Shan, Tong Xiao, and Jingbo Zhu. A survey of llm
surveys. https://github.com/NiuTrans/ABigSurveyOfLLMs, 2024.
[Rubin et al., 2022] Ohad Rubin, Jonathan Herzig, and Jonathan Berant. Learning to retrieve prompts
for in-context learning. In Proceedings of the 2022 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies, pages 2655–2671, 2022.
[Russell, 2019] Stuart Russell. Human Compatible: Artificial Intelligence and the Problem of Controls.
Viking, 2019.
[Sanh et al., 2020] Victor Sanh, Thomas Wolf, and Alexander Rush. Movement pruning: Adaptive sparsity
by fine-tuning. Advances in Neural Information Processing Systems, 33:20378–20389, 2020.

4.5 Summary

217

[Sanh et al., 2022] Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid
Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish
Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak,
Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen,
Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht
Sharma, Andrea Santilli, Thibault Fevry, Jason Alan Fries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M Rush. Multitask prompted training enables zero-shot
task generalization. In Proceedings of International Conference on Learning Representations, 2022.
[Schick et al., 2023] Timo Schick, Jane A. Yu, Zhengbao Jiang, Fabio Petroni, Patrick Lewis, Gautier Izacard, Qingfei You, Christoforos Nalmpantis, Edouard Grave, and Sebastian Riedel. PEER: A collaborative
language model. In Proceedings of The Eleventh International Conference on Learning Representations,
2023.
[Schick et al., 2024] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric
Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can
teach themselves to use tools. Advances in Neural Information Processing Systems, 36, 2024.
[Schmidhuber, 2015] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural networks, 61:85–117, 2015.
[Schulman et al., 2015] John Schulman, Sergey Levine, Philipp Moritz, Michael Jordan, and Pieter Abbeel.
Trust region policy optimization. In Proceedings of the 32nd International Conference on International
Conference on Machine Learning-Volume 37, pages 1889–1897, 2015.
[Schulman et al., 2017] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov.
Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
[Sennrich et al., 2016] Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine
translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Papers), pages 86–96, 2016.
[Seo et al., 2017] Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. Bidirectional
attention flow for machine comprehension. In Proceedings of International Conference on Learning
Representations, 2017.
[Shannon, 1951] Claude E Shannon. Prediction and entropy of printed english. Bell system technical
journal, 30(1):50–64, 1951.
[Shaw et al., 2018] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position
representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages
464–468, 2018.
[Shazeer, 2019] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint
arXiv:1911.02150, 2019.
[Shazeer, 2020] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020.
[Shen et al., 2020] Sheng Shen, Zhen Dong, Jiayu Ye, Linjian Ma, Zhewei Yao, Amir Gholami, Michael W
Mahoney, and Kurt Keutzer. Q-bert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8815–8821, 2020.
[Shoeybi et al., 2019] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper,
and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using model
parallelism. arXiv preprint arXiv:1909.08053, 2019.
[Skalse et al., 2022] Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining
and characterizing reward gaming. Advances in Neural Information Processing Systems, 35:9460–9471,
2022.
[Snell et al., 2022] Charlie Snell, Dan Klein, and Ruiqi Zhong. Learning by distilling context. arXiv

218

Alignment

preprint arXiv:2209.15189, 2022.
[Socher et al., 2013] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning,
Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a
sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language
processing, pages 1631–1642, 2013.
[Song et al., 2019] Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mass: Masked sequence
to sequence pre-training for language generation. In International Conference on Machine Learning,
pages 5926–5936. PMLR, 2019.
[Stiennon et al., 2020] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea
Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback.
Advances in Neural Information Processing Systems, 33:3008–3021, 2020.
[Su et al., 2024] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024.
[Sun et al., 2020] Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
Mobilebert: a compact task-agnostic bert for resource-limited devices. In Proceedings of the 58th Annual
Meeting of the Association for Computational Linguistics, pages 2158–2170, 2020.
[Sutskever et al., 2014] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with
neural networks. Advances in neural information processing systems, 27, 2014.
[Sutton and Barto, 2018] Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction (2nd ed.). The MIT Press, 2018.
[Szepesvári, 2010] Csaba Szepesvári. Algorithms for reinforcement learning. Synthesis Lectures on Artificial Intelligence and Machine Learning, 4(1):1–103, 2010.
[Talmor and Berant, 2018] Alon Talmor and Jonathan Berant. The web as a knowledge-base for answering
complex questions. arXiv preprint arXiv:1803.06643, 2018.
[Taori et al., 2023] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos
Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama
model. https://github.com/tatsu-lab/stanford_alpaca, 2023.
[Tay et al., 2020] Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A
survey. CoRR, abs/2009.06732, 2020.
[Team et al., 2024] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin,
Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al.
Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024.
[Teknium, 2023] Teknium. Openhermes 2.5: An open dataset of synthetic data for generalist llm assistants,
2023. URL https://huggingface.co/datasets/teknium/OpenHermes-2.5.
[Touvron et al., 2023] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne
Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation
language models. arXiv preprint arXiv:2302.13971, 2023a.
[Touvron et al., 2023] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi,
Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel,
Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony
Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee,
Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor
Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor,

4.5 Summary

219

Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan,
Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas
Scialom. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288,
2023b.
[Uesato et al., 2022] Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa
Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. Solving math word problems with processand outcome-based feedback. arXiv preprint arXiv:2211.14275, 2022.
[Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N
Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in
Neural Information Processing Systems, volume 30, 2017.
[Von Oswald et al., 2023] Johannes Von Oswald, Eyvind Niklasson, Ettore Randazzo, João Sacramento,
Alexander Mordvintsev, Andrey Zhmoginov, and Max Vladymyrov. Transformers learn in-context by
gradient descent. In Proceedings of International Conference on Machine Learning, pages 35151–
35174. PMLR, 2023.
[Wang et al., 2024] Chenglong Wang, Hang Zhou, Yimin Hu, Yifu Huo, Bei Li, Tongran Liu, Tong Xiao,
and Jingbo Zhu. Esrl: Efficient sampling-based reinforcement learning for sequence generation. In
Proceedings of the AAAI Conference on Artificial Intelligence, pages 19107–19115, 2024.
[Wang et al., 2023] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of
continual learning: Theory, method and application. arXiv preprint arXiv:2302.00487, 2023a.
[Wang et al., 2019] Qiang Wang, Bei Li, Tong Xiao, Jingbo Zhu, Changliang Li, Derek F Wong, and
Lidia S Chao. Learning deep transformer models for machine translation. In Proceedings of the 57th
Annual Meeting of the Association for Computational Linguistics, pages 1810–1822, 2019.
[Wang et al., 2022] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou.
Rationale-augmented ensembles in language models. arXiv preprint arXiv:2207.00747, 2022a.
[Wang et al., 2023] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang,
Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. In Proceedings of The Eleventh International Conference on Learning Representations,
2023b.
[Wang et al., 2022] Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza
Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, Eshaan Pathak, Giannis Karamanolakis, Haizhi Gary Lai, Ishan Purohit, Ishani Mondal, Jacob Anderson,
Kirby Kuznia, Krima Doshi, Kuntal Kumar Pal, Maitreya Patel, Mehrad Moradshahi, Mihir Parmar, Mirali Purohit, Neeraj Varshney, Phani Rohitha Kaza, Pulkit Verma, Ravsehaj Singh Puri, Rushang Karia,
Savan Doshi, Shailaja Keyur Sampat, Siddhartha Mishra, Sujan Reddy A, Sumanta Patro, Tanay Dixit,
and Xudong Shen. Super-naturalinstructions: Generalization via declarative instructions on 1600+ nlp
tasks. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing,
pages 5085–5109, 2022b.
[Wang et al., 2023] Yizhong Wang, Hamish Ivison, Pradeep Dasigi, Jack Hessel, Tushar Khot, Khyathi Raghavi Chandu, David Wadden, Kelsey MacMillan, Noah A. Smith, Iz Beltagy, and Hannaneh
Hajishirzi. How far can camels go? exploring the state of instruction tuning on open resources. Advances in Neural Information Processing Systems, 36:74764–74786, 2023c.
[Wang et al., 2023] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel
Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers), pages 13484–13508, 2023d.
[Wang et al., 2023] Zhenyi Wang, Enneng Yang, Li Shen, and Heng Huang. A comprehensive survey of
forgetting in deep learning beyond continual learning. arXiv preprint arXiv:2307.09218, 2023e.

220

Alignment

[Warstadt et al., 2019] Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625–641, 2019.
[Wei et al., 2022] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan
Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. In Proceedings
of International Conference on Learning Representations, 2022a.
[Wei et al., 2022] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud,
Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol
Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. arXiv
preprint arXiv:2206.07682, 2022b.
[Wei et al., 2022] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia,
Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language
models. Advances in Neural Information Processing Systems, 35:24824–24837, 2022c.
[Welleck et al., 2023] Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel
Khashabi, and Yejin Choi. Generating sequences by learning to self-correct. In Proceedings of The
Eleventh International Conference on Learning Representations, 2023.
[Weng, 2021] Lilian Weng. How to train really large models on many gpus? lilianweng.github.io, Sep
2021. URL https://lilianweng.github.io/posts/2021-09-25-train-large/.
[Wiener, 1960] Norbert Wiener. Some moral and technical consequences of automation: As machines
learn they may develop unforeseen strategies at rates that baffle their programmers. Science, 131(3410):
1355–1358, 1960.
[Williams et al., 2018] Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge
corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language Technologies,
Volume 1 (Long Papers), pages 1112–1122, 2018.
[Williams, 1992] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist
reinforcement learning. Machine learning, 8:229–256, 1992.
[Wingate et al., 2022] David Wingate, Mohammad Shoeybi, and Taylor Sorensen. Prompt compression
and contrastive conditioning for controllability and toxicity reduction in language models. In Findings
of the Association for Computational Linguistics: EMNLP 2022, pages 5621–5634, 2022.
[Wu et al., 2024] Wilson Wu, John X Morris, and Lionel Levine. Do language models plan for future
tokens? arXiv preprint arXiv:2404.00859, 2024.
[Wu et al., 2021] Yuhuai Wu, Markus Norman Rabe, DeLesley Hutchins, and Christian Szegedy. Memorizing transformers. In Proceedings of International Conference on Learning Representations, 2021.
[Wu et al., 2023] Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu,
Noah A. Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better
rewards for language model training. In Thirty-seventh Conference on Neural Information Processing
Systems, 2023.
[Xia et al., 2024] Mengzhou Xia, Sadhika Malladi, Suchin Gururangan, Sanjeev Arora, and Danqi Chen.
Less: Selecting influential data for targeted instruction tuning. arXiv preprint arXiv:2402.04333, 2024.
[Xiao et al., 2024] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient
streaming language models with attention sinks. In Proceedings of The Twelfth International Conference
on Learning Representations, 2024.
[Xiao and Zhu, 2023] Tong Xiao and Jingbo Zhu. Introduction to transformers: an nlp perspective. arXiv
preprint arXiv:2311.17633, 2023.
[Xiao et al., 2013] Tong Xiao, Jingbo Zhu, and Tongran Liu. Bagging and boosting statistical machine
translation systems. Artificial Intelligence, 195:496–527, 2013.

4.5 Summary

221

[Xiao et al., 2019] Tong Xiao, Yinqiao Li, Jingbo Zhu, Zhengtao Yu, and Tongran Liu. Sharing attention
weights for fast transformer. In Proceedings of the Twenty-Eighth International Joint Conference on
Artificial Intelligence (IJCAI-19), pages 5292–5298, 2019.
[Xie et al., 2022] Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. An explanation
of in-context learning as implicit bayesian inference. In Proceedings of International Conference on
Learning Representations, 2022.
[Xin et al., 2020] Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. Deebert: Dynamic early
exiting for accelerating bert inference. In Proceedings of the 58th Annual Meeting of the Association for
Computational Linguistics, pages 2246–2251, 2020.
[Xu et al., 2024] Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao,
Qingwei Lin, and Daxin Jiang. Wizardlm: Empowering large pre-trained language models to follow
complex instructions. In The Twelfth International Conference on Learning Representations, 2024.
[Yang et al., 2024] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu,
Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint
arXiv:2412.15115, 2024.
[Yang et al., 2019] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and
Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in
neural information processing systems, 32, 2019.
[Yao et al., 2024] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik
Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in
Neural Information Processing Systems, 36, 2024.
[Yarowsky, 1995] David Yarowsky. Unsupervised word sense disambiguation rivaling supervised methods.
In Proceedings of the 33rd annual meeting of the association for computational linguistics, pages 189–
196, 1995.
[Yu et al., 2023] Zihan Yu, Liang He, Zhen Wu, Xinyu Dai, and Jiajun Chen. Towards better chain-ofthought prompting strategies: A survey. arXiv preprint arXiv:2310.04959, 2023.
[Zaheer et al., 2020] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, C. Alberti,
S. Ontañón, Philip Pham, Anirudh Ravula, Qifan Wang, L. Yang, and A. Ahmed. Big bird: Transformers
for longer sequences. Advances in neural information processing systems, 33:17283–17297, 2020.
[Zellers et al., 2018] Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. Swag: A large-scale
adversarial dataset for grounded commonsense inference. In Proceedings of the 2018 Conference on
Empirical Methods in Natural Language Processing, pages 93–104, 2018.
[Zhang and Sennrich, 2019] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019.
[Zhang et al., 2023] Zhuosheng Zhang, Yao Yao, Aston Zhang, Xiangru Tang, Xinbei Ma, Zhiwei He,
Yiming Wang, Mark Gerstein, Rui Wang, Gongshen Liu, and Hai Zhao. Igniting language intelligence: The hitchhiker’s guide from chain-of-thought reasoning to language agents. arXiv preprint
arXiv:2311.11797, 2023a.
[Zhang et al., 2023] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought
prompting in large language models. In The Eleventh International Conference on Learning Representations, 2023b.
[Zhao et al., 2024] Hao Zhao, Maksym Andriushchenko, Francesco Croce, and Nicolas Flammarion. Long
is more for alignment: A simple but tough-to-beat baseline for instruction fine-tuning. arXiv preprint
arXiv:2402.04833, 2024.
[Zhao et al., 2023] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou,
Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Z. Chen,
Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jianyun Nie, and Ji rong Wen.

222

Alignment

A survey of large language models. arXiv preprint arXiv:2303.18223, 2023.
[Zhou et al., 2023] Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma,
Avia Efrat, Ping Yu, Lili Yu, Susan Zhang, Gargi Ghosh, Mike Lewis, Luke Zettlemoyer, and Omer
Levy. Lima: Less is more for alignment. arXiv preprint arXiv:2305.11206, 2023a.
[Zhou et al., 2023] Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale
Schuurmans, Claire Cui, Olivier Bousquet, Quoc V. Le, and Ed H. Chi. Least-to-most prompting enables
complex reasoning in large language models. In Proceedings of The Eleventh International Conference
on Learning Representations, 2023b.
[Zhou et al., 2020] Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian McAuley, Ke Xu, and Furu Wei. Bert
loses patience: Fast and robust inference with early exit. Advances in Neural Information Processing
Systems, 33:18330–18341, 2020.
[Zhou et al., 2023] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris
Chan, and Jimmy Ba. Large language models are human-level prompt engineers. In The Eleventh
International Conference on Learning Representations, 2023c.
[Zoph and Le, 2016] Barret Zoph and Quoc Le. Neural architecture search with reinforcement learning. In
Proceedings of International Conference on Learning Representations, 2016.
[Zoph et al., 2020] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk,
and Quoc Le. Rethinking pre-training and self-training. Advances in neural information processing
systems, 33:3833–3845, 2020.

Index
k-NN, 74
k-NN LM, 76
k-NN language modeling, 76
k-nearest neighbors, 74
A2C, 175
action-value function, 171
advantage, 175
advantage actor-critic, 175
Agent, 47
ALiBi, 85
alignment, 46
attention with linear biases, 85
automated machine learning, 137
automatic prompt design, 137
AutoML, 137
autonomous agents, 134
BART, 19
BERT, 1
Best-of-N sampling, 197
BoN sampling, 197
Bradley-Terry model, 178
calculation annotation, 113
catastrophic forgetting, 34
causal language modeling, 9
chain of thought, 113
chain-of-thought prompting, 53
completion, 6
compositional generalization, 122
CoT, 113
COT prompting, 53
cross-lingual language models, 28
cumulative reward, 172
deliberate-then-generate, 126
demonstrations, 6
direct preference optimization, 190
Document Rotation, 20
DPO, 190
DTG, 126
emergent abilities, 63
external memories, 74
Extrapolation, 81

few-shot COT prompting, 54
gated linear unit, 58
gaussian error linear unit, 58
GeLU, 58
GLU, 58
GPT, 1
GQA, 80
Grouped query attention, 80
hard prompts, 140
human preference alignment, 152
ICL, 53
ICT, 6
importance sampling, 180
in-context learning, 6, 53, 95
input inversion, 163
instruction alignment, 152
instruction fine-tuning, 43, 154
interference, 30
internal memories, 74
Interpolation, 82
irreducible error, 63
key-value cache, 68
KV cache, 68
label mapping, 105
Learning from Human Feedback, 47
least-to-most prompting, 119
long-context LLMs, 66
masked language modeling, 1, 9
mBERT, 28
memory-based methods, 74
MQA, 79
multi-lingual BERT, 28
multi-query attention, 79
NAS, 137
neural architecture search, 137
next sentence prediction, 13
NSP, 13
offline reinforcement learning, 193
one-shot COT prompting, 54
Outcome-based Approaches, 195
223

Alignment

224
overoptimization problem, 189
Performance Estimation, 137
performance function, 172
performance gap recovered, 167
permuted language modeling, 11
PGR, 167
Plackett-Luce model, 184
PPO, 50, 181
prefix fine-tuning, 144
prefix language modeling, 16
problem decomposition, 116
Process-based Approaches, 195
prompt embeddings, 148
prompt engineering, 95
prompt optimization, 137
Prompt Search Space, 137
prompting engineering, 51
proximal policy optimization, 50, 181
Q-value function, 171
RAG, 76
ratio function, 180
rectified linear unit, 58
reinforcement learning from human feedback,
47, 153
rejection sampling, 198
relation extraction, 108
ReLU, 58
retrieval-augmented generation, 76
return, 172
reward gaming, 189
reward hacking, 189
Reward Model, 47
RLHF, 47, 153
RoBERTa, 26
sample efficient, 164
scaling laws, 63
self-consistency, 129
self-instruct, 160
self-supervised learning, 3
self-training, 3
Sentence Reordering, 19
Sequence Encoding Models, 3
Sequence Generation Models, 3
SFT, 47, 152

single-round prediction, 155
soft prompts, 140
Span Masking, 19
state-value function, 171
Strong Ceiling Performance, 167
Sub-problem Generation, 118
Sub-problem Solving, 118
superficial alignment hypothesis, 164
Supervised Fine-tuning, 47
supervised fine-tuning, 152
supervised learning, 2
surrogate objective, 180
T5, 15
TD, 176
temporal difference, 176
text completion, 109
text transformation, 109
Token Deletion, 19
Token Masking, 19
Transformers, 1
translation language modeling, 29
trust regions, 181
unsupervised learning, 2
Weak Performance, 167
weak-to-strong generalization, 166
Weak-to-strong Performance, 167
XLMs, 28
zero-shot COT, 54
zero-shot learning, 45