Foundational

Large Language
Models &
Text Generation
Authors: Mohammadamin Barektain,
Anant Nawalgaria, Daniel J. Mankowitz,
Majd Al Merey, Yaniv Leviathan, Massimo Mascaro,
Matan Kalman, Elena Buchatskaya,
Aliaksei Severyn, Irina Sigler, and Antonio Gulli
Foundational Large Language Models & Text Generation

Acknowledgements

Content contributors

Adam Sadvovsky

Yonghui Wu

Andrew Dai

Efi Kokiopolou

Chuck Sugnet

Aleksey Vlasenko

Erwin Huizenga

Aida Nematzadeh

Ira Ktena

Olivia Wiles

Lavi Nigam

Curators and Editors

Antonio Gulli

Anant Nawalgaria

Grace Mollison

Technical Writer

Mark Iverson

Designer

Michael Lanning

February 2025 2
Table of contents

Introduction 6

Why language models are important 7

Large language models 8

Transformer 9

Input preparation and embedding 11

Multi-head attention 12

Understanding self-attention 12

Multi-head attention: power in diversity 14

Layer normalization and residual connections 15

Feedforward layer 15

Encoder and decoder 16

Mixture of Experts (MoE) 17

Training the transformer 20

Data preparation 21

Training and loss function 21

The evolution of transformers 23

GPT-1 23

BERT 25
 GPT-2 25

GPT-3/3.5/4 27

LaMDA 28

Gopher 29

GLaM 31

Chinchilla 31

PaLM 33

PaLM 2 33

Gemini 34

Gemma 37

LLaMA 38

Mixtral 39

OpenAI O1 40

DeepSeek 40

Other open models 41

Comparison 43

Fine-tuning large language models 45

Supervised fine-tuning  46

Reinforcement learning from human feedback 47

Parameter Efficient Fine-Tuning 49

Using large language models 52

Prompt engineering  52

Sampling Techniques and Parameters 53

Task-based Evaluation 54

Accelerating inference 57
 Trade offs 58

The Quality vs Latency/Cost Tradeoff 58

The Latency vs Cost Tradeoff 59

Output-approximating methods 60

Quantization 60

Distillation 61

Output-preserving methods 62

Flash Attention 63

Prefix Caching 63

Speculative Decoding 65

Batching and Parallelization 67

Applications 68

Code and mathematics 71

Machine translation 72

Text summarization 73

Question-answering 73

Chatbots 74

Content generation 75

Natural language inference 75

Text classification 76

Text analysis 77

Multimodal applications 78

Summary 80

Endnotes 82
Foundational Large Language Models & Text Generation

We believe that this new crop of

technologies has the potential to
assist, complement, empower,
and inspire people at any time
across almost any field.

Introduction
The advent of Large Language Models (LLMs) represents a seismic shift in the world of
artificial intelligence. Their ability to process, generate, and understand user intent is
fundamentally changing the way we interact with information and technology.

An LLM is an advanced artificial intelligence system that specializes in processing,

understanding, and generating human-like text. These systems are typically implemented as
a deep neural network and are trained on massive amounts of text data. This allows them to
learn the intricate patterns of language, giving them the ability to perform a variety of tasks,
like machine translation, creative text generation, question answering, text summarization,
and many more reasoning and language oriented tasks. This whitepaper dives into the
timeline of the various architectures and approaches building up to the large language
models and the architectures being used at the time of publication. It also discusses fine-

February 2025 6
Foundational Large Language Models & Text Generation

tuning techniques to customize an LLM to a certain domain or task, methods to make the
training more efficient, as well as methods to accelerate inference. These are then followed
by various applications and code examples.

Why language models are important

LLMs achieve an impressive performance boost from the previous state of the art NLP
models across a variety of different and complex tasks which require answering questions
or complex reasoning, making feasible many new applications. These include language
translation, code generation and completion, text generation, text classification, and
question-answering, to name a few. Although foundational LLMs trained in a variety of
tasks on large amounts of data perform very well out of the box and display emergent
behaviors (e.g. the ability to perform tasks they have not been directly trained for) they can
also be adapted to solve specific tasks where performance out of the box is not at the level
desired through a process known as fine-tuning. This requires significantly less data and
computational resources than training an LLM from scratch. LLMs can be further nudged
and guided towards the desired behavior by the discipline of prompt engineering: the art and
science of composing the prompt and the parameters of an LLM to get the desired response.

The big question is: how do these large language models work? The next section explores the
core building blocks of LLMs, focusing on transformer architectures and their evolution from
the original ‘Attention is all you need’ paper1 to the latest models such as Gemini, Google’s
most capable LLM. We also cover training and fine-tuning techniques, as well as methods to
improve the speed of response generation. The whitepaper concludes with a few examples
of how language models are used in practice.

February 2025 7
Foundational Large Language Models & Text Generation

Large language models

A language model predicts the probability of a sequence of words. Commonly, when given
a prefix of text, a language model assigns probabilities to subsequent words. For example,
given the prefix “The most famous city in the US is…”, a language model might predict high
probabilities to the words “New York” and “Los Angeles” and low probabilities to the words
“laptop” or “apple”. You can create a basic language model by storing an n-gram table,2 while
modern language models are often based on neural models, such as transformers.

Before the invention of transformers1, recurrent neural networks (RNNSs) were the popular
approach for modeling sequences. In particular, “long short-term memory” (LSTM) and
“gated recurrent unit” (GRU) were common architectures.3 This area includes language
problems such as machine translation, text classification, text summarization, and question-
answering, among others. RNNs process input and output sequences sequentially. They
generate a sequence of hidden states based on the previous hidden state and the current
input. The sequential nature of RNNs makes them compute-intensive and hard to parallelize
during training (though recent work in state space modeling is attempting to overcome
these challenges).

Transformers, on the other hand, are a type of neural network that can process sequences
of tokens in parallel thanks to the self-attention mechanism.1 This means that transformers
can model long-term contexts more effectively and are easier to parallelize than RNNs. This
makes them significantly faster to train, and more powerful compared to RNNs for handling
long-term dependencies in long sequence tasks. However, the cost of self-attention in the
original transformers is quadratic in the context length which limits the size of the context,
while RNNs have a theoretically infinite context length. Although they have infinite context
length, in practice they struggle to utilize it due to vanishing gradient problem. Transformers
have become the most popular approach for sequence modeling and transfer learning
problems in recent years.

February 2025 8
Foundational Large Language Models & Text Generation

Herein, we discuss the first version of the transformer model and then move on to the more
recent advanced models and algorithms.

Transformer

The transformer architecture was developed at Google in 2017 for use in a translation model.1
It’s a sequence-to-sequence model capable of converting sequences from one domain
into sequences in another domain. For example, translating French sentences to English
sentences. The original transformer architecture consists of two parts: an encoder and a
decoder. The encoder converts the input text (e.g., a French sentence) into a representation,
which is then passed to the decoder. The decoder uses this representation to generate the
output text (e.g., an English translation) autoregressively.1 Notably, the size of the output of
the transformer encoder is linear in the size of its input. Figure 1 shows the design of the
original transformer architecture.

The transformer consists of multiple layers. A layer in a neural network comprises a set of
parameters that perform a specific transformation on the data. In the diagram you can see
an example of some layers which include Multi-Head Attention, Add & Norm, Feed-Forward,
Linear, Softmax etc. The layers can be sub-divided into the input, hidden and output layers.
The input layer (e.g., Input/Output Embedding) is the layer where the raw data enters the
network. Input embeddings are used to represent the input tokens to the model. Output
embeddings are used to represent the output tokens that the model predicts. For example, in
a machine translation model, the input embeddings would represent the words in the source
language, while the output embeddings would represent the words in the target language.
The output layer (e.g., Softmax) is the final layer that produces the output of the network. The
hidden layers (e.g., Multi-Head Attention) are between the input and output layers and are
where the magic happens!

February 2025 9
Foundational Large Language Models & Text Generation

Figure 1. Original Transformer1 (P.C:5)

February 2025 10
Foundational Large Language Models & Text Generation

To better understand the different layers in the transformer, let’s use a French-to-English
translation task as an example. Here, we explain how a French sentence is input into the
transformer and a corresponding English translation is output. We will also describe each of
the components inside the transformer from Figure 1.

Input preparation and embedding

To prepare language inputs for transformers, we convert an input sequence into tokens
and then into input embeddings. At a high level, an input embedding is a high-dimensional
vector68 that represents the meaning of each token in the sentence. This embedding is
then fed into the transformer for processing. Generating an input embedding involves the
following steps:

1. Normalization (Optional): Standardizes text by removing redundant whitespace,

accents, etc.

2. Tokenization: Breaks the sentence into words or subwords and maps them to integer
token IDs from a vocabulary.

3. Embedding: Converts each token ID to its corresponding high-dimensional vector,

typically using a lookup table. These can be learned during the training process.

4. Positional Encoding: Adds information about the position of each token in the sequence
to help the transformer understand word order.

These steps help to prepare the input for the transformers so that they can better
understand the meaning of the text.

February 2025 11
Foundational Large Language Models & Text Generation

Multi-head attention

After converting input tokens into embedding vectors, you feed these embeddings into
the multi-head attention module (see Figure 1). Self-attention is a crucial mechanism in
transformers; it enables them to focus on specific parts of the input sequence relevant to
the task at hand and to capture long-range dependencies within sequences more effectively
than traditional RNNs.

Understanding self-attention

Consider the following sentence: “The tiger jumped out of a tree to get a drink because it
was thirsty.” Self-attention helps to determine relationships between different words and
phrases in sentences. For example, in this sentence, “the tiger” and “it” are the same object,
so we would expect these two words to be strongly connected. Self-attention achieves this
through the following steps (Figure 2):

1. Creating queries, keys, and values: Each input embedding is multiplied by three learned
weight matrices (Wq, Wk, Wv) to generate query (Q), key (K), and value (V) vectors. These
are like specialized representations of each word.

• Query: The query vector helps the model ask, “Which other words in the sequence are
relevant to me?”

• Key: The key vector is like a label that helps the model identify how a word might be
relevant to other words in the sequence.

• Value: The value vector holds the actual word content information.

2. Calculating scores: Scores are calculated to determine how much each word should
‘attend’ to other words. This is done by taking the dot product of the query vector of one
word with the key vectors of all the words in the sequence.

February 2025 12
Foundational Large Language Models & Text Generation

3. Normalization: The scores are divided by the square root of the key vector dimension (dk)
for stability, then passed through a softmax function to obtain attention weights. These
weights indicate how strongly each word is connected to the others.

4. Weighted values: Each value vector is multiplied by its corresponding attention weight.
The results are summed up, producing a context-aware representation for each word.

Figure 2. The process of computing self-attention in the multi-head attention module1 (P.C:5)

February 2025 13
Foundational Large Language Models & Text Generation

In practice, these computations are performed at the same time, by stacking the query, key
and value vectors for all the tokens into Q, K and V matrices and multiplying them together as
shown in Figure 3.

Figure 3. The basic operation of attention,1 with Q=query, K=Keys and V=Value, Z=Attention, d_k = dimension
of queries and keys (P.C:5)

Multi-head attention: power in diversity

Multi-head attention employs multiple sets of Q, K, V weight matrices. These run in parallel,
each ‘head’ potentially focusing on different aspects of the input relationships. The outputs
from each head are concatenated and linearly transformed, giving the model a richer
representation of the input sequence.

The use of multi-head attention improves the model’s ability to handle complex language
patterns and long-range dependencies. This is crucial for tasks that require a nuanced
understanding of language structure and content, such as machine translation, text
summarization, and question-answering. The mechanism enables the transformer to consider
multiple interpretations and representations of the input, which enhances its performance on
these tasks.

February 2025 14
Foundational Large Language Models & Text Generation

Layer normalization and residual connections

Each layer in a transformer, consisting of a multi-head attention module and a feed-forward

layer, employs layer normalization and residual connections. This corresponds to the Add
and Norm layer in Figure 1, where ‘Add’ corresponds to the residual connection and ‘Norm’
corresponds to layer normalization. Layer normalization computes the mean and variance
of the activations to normalize the activations in a given layer. This is typically performed to
reduce covariate shift as well as improve gradient flow to yield faster convergence during
training as well as improved overall performance.

Residual connections propagate the inputs to the output of one or more layers. This has the
effect of making the optimization procedure easier to learn and also helps deal with vanishing
and exploding gradients.

The Add and Norm layer is applied to both the multi-head attention module and the feed-
forward layer described in the following section.

Feedforward layer

The output of the multi-head attention module and the subsequent ‘Add and Norm’ layer is
fed into the feedforward layer of each transformer block. This layer applies a position-wise
transformation to the data, independently for each position in the sequence, which allows the
incorporation of additional non-linearity and complexity into the model’s representations. The
feedforward layer typically consists of two linear transformations with a non-linear activation
function, such as ReLU or GELU, in between. This structure adds further representational
power to the model. After processing by the feedforward layer, the data undergoes
another ‘Add and Norm’ step, which contributes to the stability and effectiveness of deep
transformer models.

February 2025 15
Foundational Large Language Models & Text Generation

Encoder and decoder

The original transformer architecture relies on a combination of encoder and decoder

modules. Each encoder and decoder consists of a series of layers, with each layer
comprising key components: a multi-head self-attention mechanism, a position-wise feed-
forward network, normalization layers, and residual connections.

The encoder’s primary function is to process the input sequence into a continuous
representation that holds contextual information for each token. The input sequence is first
normalized, tokenized, and converted into embeddings. Positional encodings are added to
these embeddings to retain sequence order information. Through self-attention mechanisms,
each token in the sequence can dynamically attend to any other token, thus understanding
the contextual relationships within the sequence. The output from the encoder is a series of
embedding vectors Z representing the entire input sequence.

The decoder is tasked with generating an output sequence based on the context provided
by the encoder’s output Z. It operates in a token-by-token fashion, beginning with a start-
of-sequence token. The decoder layers employ two types of attention mechanisms: masked
self-attention and encoder-decoder cross-attention. Masked self-attention ensures that
each position can only attend to earlier positions in the output sequence, preserving the
auto-regressive property. This is crucial for preventing the decoder from having access to
future tokens in the output sequence. The encoder-decoder cross-attention mechanism
allows the decoder to focus on relevant parts of the input sequence, utilizing the contextual
embeddings generated by the encoder. This iterative process continues until the decoder
predicts an end-of-sequence token, thereby completing the output sequence generation.

Majority of recent LLMs adopted a decoder-only variant of transformer architecture. This

approach forgoes the traditional encoder-decoder separation, focusing instead on directly
generating the output sequence from the input. The input sequence undergoes a similar

February 2025 16
Foundational Large Language Models & Text Generation

process of embedding and positional encoding before being fed into the decoder. The
decoder then uses masked self-attention to generate predictions for each subsequent
token based on the previously generated tokens. This streamlined approach simplifies the
architecture for specific tasks where encoding and decoding can be effectively merged.

Mixture of Experts (MoE)

A Mixture of Experts (MoE) is a an architecture that combines multiple specialized sub-

models (the “experts”) to improve overall performance, particularly on complex tasks. It’s
a form of ensemble learning, but with a key difference: instead of simply aggregating the
predictions of all experts, it learns to route different parts of the input to different experts.
This allows the model to specialize, with each expert becoming proficient in a specific
sub-domain or aspect of the data. Here’s a more technical breakdown describing the main
components of an MoE:

• Experts: These are the individual sub-models, each designed to handle a specific subset
of the input data or a particular task. They can be any type of model (e.g., neural networks,
decision trees, etc.), but in the context of large language models, they are typically
themselves transformer-based architectures.

• Gating Network (Router): This is a crucial component that learns to route the input to
the appropriate expert(s). It takes the input and produces a probability distribution over
the experts. This distribution determines how much each expert should “contribute” to the
final prediction. The gating network is also typically a neural network.

• Combination Mechanism: This combines the outputs of the experts, weighted by the
probabilities from the gating network, to produce the final prediction. A common approach
is a weighted average.

February 2025 17
Foundational Large Language Models & Text Generation

In practice, A Mixture of Experts (MoE) architecture combines multiple specialized sub-

models, called “experts,” to tackle complex tasks. Instead of simply averaging all expert
predictions, an MoE uses a “gating network” to intelligently route different parts of the
input to the mostt relevant experts. Both the experts and the gating network receive the
input. Each expert processes the input and generates its output. Simultaneously, the
gating network analyzes the input and produces a probability distribution over the experts,
indicating how much each expert should contribute to the final result. These probabilities
then weight the outputs of the experts, and the weighted combination becomes the final
prediction. This allows different experts to specialize in handling specific types of data or
sub-tasks, improving overall performance and, through “sparse activation,” potentially
reducing computational cost by only activating a subset of experts for any given input.

Figure 4. Mixture of experts ensembling70

February 2025 18
Foundational Large Language Models & Text Generation

Large Reasoning Models

Achieving robust reasoning capabilities in Large Models is a complex endeavor that involves
a combination of architectural designs, training methodologies, and prompting strategies.
One crucial aspect is incorporating inductive biases that favor reasoning-conducive patterns.
Transformer architectures, with their self-attention mechanisms, are foundational, allowing
the model to weigh the importance of different parts of the input sequence when generating
an output. However, vanilla Transformers alone are not sufficient for complex reasoning.

Chain-of-Thought prompting explicitly encourages the model to generate intermediate

reasoning steps before arriving at a final answer. By providing examples of step-by-step
reasoning in the prompt, the model learns to decompose complex problems into smaller,
manageable sub-problems. This mimics human reasoning processes and significantly
improves performance on tasks requiring multi-step inference. Tree-of-Thoughts takes this
further, exploring multiple reasoning paths and using a search algorithm to find the most
promising solution. This technique is useful with game trees or combinatorial problems.
Least-to-Most prompting guides the model to solve subproblems, which get progressively
more complex, and the output of one subproblem is used as part of the prompt of the more
complex, subsequent problem.

Fine-tuning on datasets specifically designed for reasoning tasks is also crucial. These
datasets may contain logical puzzles, mathematical problems, or commonsense reasoning
challenges. Instruction tuning, where the model is trained to follow natural language
instructions, further enhances its ability to understand and respond to complex reasoning
prompts. Reinforcement Learning from Human Feedback (RLHF) refines the model’s outputs
based on human preferences, improving the quality and coherence of its reasoning. RLHF
helps in reward models that score reasoning ability as well as “helpfulness.”

February 2025 19
Foundational Large Language Models & Text Generation

Knowledge distillation, transferring knowledge from a larger, more capable “teacher” model
to a smaller, more efficient “student” model, can be used to improve the reasoning abilities
of smaller models while maintaining efficiency. This approach allows the student model to
learn the reasoning patterns of the teacher model without requiring the same computational
resources. During inference, techniques like beam search, which explores multiple candidate
outputs simultaneously, can improve the quality of reasoning by considering different
possibilities. Temperature scaling, adjusting the randomness of the model’s output, can also
influence the exploration-exploitation trade-off in reasoning. Finally, incorporating external
knowledge sources, such as knowledge graphs or structured databases, can provide the
model with additional information to support its reasoning process. This can be achieved
through techniques like retrieval-augmented generation, where the model retrieves relevant
information from an external source before generating an output. These techniques all
combined, across many domains of reasoning, create the best performing reasoning large
language models.

Training the transformer

When talking about machine learning models, it’s important to differentiate between
training and inference. Training typically refers to modifying the parameters of the model,
and involves loss functions and backpropagation. Inference is when model is used only
for the predicted output, without updating the model weights. The model parameters are
fixed during inference. Up until now we learned how transformers generate outputs during
inference. Next, we focus on how to train transformers to perform one or more given tasks.

February 2025 20
Foundational Large Language Models & Text Generation

Data preparation

The first step is data preparation, which involves a few important steps itself. First, clean the
data by applying techniques such as filtering, deduplication, and normalization. The next
step is tokenization where the dataset is converted into tokens using techniques such as
Byte-Pair Encoding8, 9 and Unigram tokenization.8, 10 Tokenization generates a vocabulary,
which is a set of unique tokens used by the LLM. This vocabulary serves as the model’s
’language’ for processing and understanding text. Finally, the data is typically split into a
training dataset for training the model as well as a test dataset which is used to evaluate the
models performance.

Training and loss function

A typical transformer training loop consists of several parts: First, batches of input
sequences are sampled from a training dataset. For each input sequence, there is a
corresponding target sequence. In unsupervised pre-training, the target sequence is
derived from the input sequence itself. The batch of input sequences is then fed into the
transformer. The transformer generates predicted output sequences. The difference
between the predicted and target sequences is measured using a loss function (often cross-
entropy loss)11. Gradients of this loss are calculated, and an optimizer uses them to update
the transformer’s parameters. This process is repeated until the transformer converges to a
certain level of performance or until it has been trained on a pre-specified number of tokens.

There are different approaches to formulating the training task for transformers depending
on the architecture used:

• Decoder-only models are typically pre-trained on the language modeling task (e.g., see
endnote12, 13). The target sequence for the decoder is simply a shifted version of the input
sequence. Given a training sequence like ‘the cat sat on the mat’ various input/target

February 2025 21
Foundational Large Language Models & Text Generation

pairs can be generated for the model. For example the input “the cat sat on” should
predict “the” and subsequently the input “the cat sat on the” should predict target
sequence “mat”.

• Encoder-only models (like BERT)14 are often pre-trained by corrupting the input sequence
in some way and having the model try to reconstruct it. One such approach is masked
language modeling (MLM).14 In our example, the input sequence could be “The [MASK] sat
on the mat” and the sequence target would be the original sentence.

• Encoder-decoder models (like the original transformer) are trained on sequence-to-

sequence supervised tasks such as translation (input sequence “Le chat est assis sur
le tapis” and target “The cat sat on the mat”), question-answering (where the input
sequence is a question and the target sequence is the corresponding answer), and
summarization (where the input sequence is a full article and the target sequence is its
corresponding summary). These models could also be trained in an unsupervised way by
converting other tasks into sequence-to-sequence format. For example, when training
on Wikipedia data, the input sequence might be the first part of an article, and the target
sequence comprises the remainder of the article.

An additional factor to consider during training is the ‘context length’. This refers to the
number of previous tokens the model can ‘remember’ and use to predict the next token in
the sequence. Longer context lengths allow the model to capture more complex relationships
and dependencies within the text, potentially leading to better performance. However, longer
contexts also require more computational resources and memory, which can slow down
training and inference. Choosing an appropriate context length involves balancing these
trade-offs based on the specific task and available resources.

February 2025 22
Foundational Large Language Models & Text Generation

The evolution of transformers

The next sections provide an overview of the various transformer architectures. These
include encoder-only, encoder-decoder, as well as decoder-only transformers. We start with
GPT-1 and BERT and end with Google’s latest family of LLMs called Gemini.

GPT-1

GPT-1 (Generative pre-trained transformer version 1)15 was a decoder-only model developed
by OpenAI in 2018. It was trained on the BooksCorpus dataset (containing approximately
several billion words) and is able to generate text, translate languages, write different kinds
of creative content, and answer questions in an informative way. The main innovations in
GPT-1 were:

• Combining transformers and unsupervised pre-training: Unsupervised pre-training

is a process of training a language model on a large corpus of unlabeled data. Then,
supervised data is used to fine-tune the model for a specific task, such as translation
or sentiment classification. In prior works, most language models were trained using a
supervised learning objective. This means that the model was trained on a dataset of
labeled data, where each example had a corresponding label. This approach has two main
limitations. First, it requires a large amount of labeled data, which can be expensive and
time-consuming to collect. Second, the model can only generalize to tasks that are similar
to the tasks that it was trained on. Semi-supervised sequence learning was one of the first
works that showed that unsupervised pre-training followed by supervised training was
superior than supervised training alone.

Unsupervised pre-training addresses these limitations by training the model on a large

corpus of unlabeled data. This data can be collected more easily and cheaply than labeled
data. Additionally, the model can generalize to tasks that are different from the tasks that

February 2025 23
Foundational Large Language Models & Text Generation

it was trained on. The BooksCorpus dataset is a large (5GB) corpus of unlabeled text that
was used to train the GPT-1 language model. The dataset contains over 7,000 unpublished
books, which provides the model with a large amount of data to learn from. Additionally,
the corpus contains long stretches of contiguous text, which helps the model learn long-
range dependencies. Overall, unsupervised pre-training is a powerful technique that can
be used to train language models that are more accurate and generalizable than models
that are trained using supervised learning alone.

• Task-aware input transformations: There are different kinds of tasks such as textual
entailment and question-answering that require a specific structure. For example,
textual entailment requires a premise and a hypothesis; question-answering requires a
context document; a question and possible answers. One of the contributions of GPT-1
is converting these types of tasks which require structured inputs into an input that the
language model can parse, without requiring task-specific architectures on top of the
pre-trained architecture. For textual entailment, the premise p and the hypothesis h are
concatenated with a delimiter token ($) in between - [p, $, h]. For question answering, the
context document c is concatenated with the question q and a possible answer a with a
delimiter token in between the question and answer - [c,q,$,a].

GPT-1 surpassed previous models on several benchmarks, achieving excellent results. While
GPT-1 was a significant breakthrough in natural language processing (NLP), it had some
limitations. For example, the model was prone to generating repetitive text, especially when
given prompts outside the scope of its training data. It also failed to reason over multiple
turns of dialogue and could not track long-term dependencies in text. Additionally, its
cohesion and fluency were limited to shorter text sequences, and longer passages would
lack cohesion. Despite these limitations, GPT-1 demonstrated the power of unsupervised
pre-training, which laid the foundation for larger and more powerful models based on the
transformer architecture.

February 2025 24
Foundational Large Language Models & Text Generation

BERT

BERT14 which stands for Bidirectional Encoder Representations from Transformers,

distinguishes itself from traditional encoder-decoder transformer models by being an
encoder-only architecture. Instead of translating or producing sequences, BERT focuses
on understanding context deeply by training on a masked language model objective. In
this setup, random words in a sentence are replaced with a [MASK] token, and BERT tries
to predict the original word based on the surrounding context. Another innovative aspect
of BERT’s training regime is the next sentence prediction loss, where it learns to determine
whether a given sentence logically follows a preceding one. By training on these objectives,
BERT captures intricate context dependencies from both the left and right of a word, and
it can discern the relationship between pairs of sentences. Such capabilities make BERT
especially good at tasks that require natural language understanding, such as question-
answering, sentiment analysis, and natural language inference, among others. Since this is an
encoder-only model, BERT cannot generate text.

GPT-2

GPT-2,12 the successor to GPT-1, was released in 2019 by OpenAI. The main innovation of
GPT-2 was a direct scale-up, with a tenfold increase in both its parameter count and the size
of its training dataset:

• Data: GPT-2 was trained on a large (40GB) and diverse dataset called WebText, which
consists of 45 million webpages from Reddit with a Karma rating of at least three. Karma
is a rating metric used on Reddit and a value of three means that all the posts were of a
reasonable level of quality.

February 2025 25
Foundational Large Language Models & Text Generation

• Parameters: GPT-2 had 1.5 billion parameters, which was an order of magnitude larger
than the previous model. More parameters increase the model’s learning capacity. The
authors trained four language models with 117M (the same as GPT-1), 345M, 762M, and 1.5B
(GPT-2) parameters, and found that the model with the most parameters performed better
on every subsequent task.

This scaling up resulted in a model that was able to generate more coherent and realistic text
than GPT-1. Its ability to generate human-like responses made it a valuable tool for various
natural language processing tasks, such as content creation and translation. Specifically,
GPT-2 demonstrated significant improvement in capturing long-range dependencies and
common sense reasoning. While it performed well in some tasks, it did not outperform state-
of-the-art reading comprehension, summarization, and translation. GPT-2’s most significant
achievement was its ability to perform zero-shot learning on a variety of tasks. Zero-shot task
transfer is the ability of a model to generalize to a new task without being trained on it, which
requires the model to understand the task based on the given instruction. For example, for
an English to German translation task, the model might be given an English sentence followed
by the word “German” and a prompt (“:”). The model would then be expected to understand
that this is a translation task and generate the German translation of the English sentence.
GPT-2 was able to perform tasks such as machine translation, text summarization, and
reading comprehension without any explicit supervision.

The study discovered that performance on zero-shot tasks increased in a log-linear manner
as the model’s capacity increased. GPT-2 showed that training on a larger dataset and having
more parameters improved the model’s ability to understand tasks and surpass the state-of-
the-art on many tasks in zero-shot settings.

February 2025 26
Foundational Large Language Models & Text Generation

GPT-3/3.5/4

GPT-3,13 or the third iteration of the Generative Pre-trained Transformer model, represents a
significant evolution from its predecessor, GPT-2, primarily in terms of scale, capabilities, and
flexibility. The most noticeable difference is the sheer size of GPT-3, boasting a whopping
175 billion parameters, compared to GPT-2’s largest model which had 1.5 billion parameters.
This increase in model size allowed GPT-3 to store and recall an even more vast amount of
information, understand nuanced instructions, and generate more coherent and contextually
relevant text over longer passages.

While GPT-2 could be fine-tuned on specific tasks with additional training data, GPT-3 can
understand and execute tasks with just a few examples, or sometimes even without any
explicit examples—simply based on the instruction provided. This highlights GPT-3’s more
dynamic understanding and adaptation abilities, reducing the need for task-specific fine-
tuning which was more prevalent in GPT-2.

Finally, GPT-3’s large model scale and diverse training corpus have led to better
generalization across a broader range of tasks. This means that out-of-the-box, without
any further training, GPT-3 exhibits improved performance on diverse NLP challenges, from
translation to question-answering, compared to GPT-2. It’s also worth noting that the release
approach differed: while OpenAI initially held back GPT-2 due to concerns about misuse,
they chose to make GPT-3 available as a commercial API, reflecting both its utility and the
organization’s evolving stance on deployment.

Instruction tuning was then introduced with InstructGPT17, a version of GPT-3 that was fine-
tuned, using Supervised Fine-Tuning, on a dataset of human demonstrations of desired
model behaviors. Outputs from this model were then ranked and it was then further fine-
tuned using Reinforcement Learning from Human Feedback. This led to improved instruction

February 2025 27
Foundational Large Language Models & Text Generation

following in the model. A 1.3B parameter InstructGPT model had better human evaluations
than the 175B parameter GPT-3 model. It also showed improvements in truthfulness and
reductions in toxicity.

GPT-3.5 models, including GPT-3.5 turbo, improve over GPT-3 as it is capable of

understanding and generating code. It’s been optimized for dialogue. And it’s capable of
receiving context windows of up to 16,385 tokens and can generate outputs of up to 4,096
tokens.

GPT-4 extends GPT-3.5 as a large multimodal model capable of processing image and
text inputs and producing text outputs.19 Specifically, accepting text or images as input
and outputting text. This model has broader general knowledge and advanced reasoning
capabilities. It can receive context windows of up to 128,000 tokens and has a maximum
output of 4,096 tokens. GPT-4 demonstrates remarkable versatility by solving complex tasks
across diverse fields like mathematics, coding, vision, medicine, law, and psychology – all
without specialized instructions. Its performance often matches or even exceeds human
capabilities and significantly outperforms earlier models like GPT-3.5.

LaMDA

Google’s LaMDA,20 which stands for ‘Language Model for Dialogue Applications’ is another
contribution to the arena of large-scale language models, designed primarily to engage in
open-ended conversations. Unlike traditional chatbots which operate in more constrained
and predefined domains, LaMDA is engineered to handle a wide array of topics, delivering
more natural and flowing conversations. LaMDA was trained on dialogue-focused data to
encourage ongoing conversational flow, not just isolated responses, ensuring users can have
more extensive and explorative dialogues.

February 2025 28
Foundational Large Language Models & Text Generation

While GPT models, especially the later iterations like GPT-3, have strived to address a
multitude of tasks simultaneously, from text generation to code writing, LaMDA’s primary
focus is on maintaining and enhancing conversational depth and breadth. GPT models
shine on their ability to produce coherent long-form content and perform various tasks
with minimal prompting, whereas LaMDA emphasizes the flow and progression of dialogue,
striving to mimic the unpredictability and richness of human conversations.

Gopher

Gopher22 is a 280 billion parameter language model based on the decoder-only transformer
architecture, developed by DeepMind in 2021.22 It can generate text, translate languages,
write different kinds of creative content, and answer your questions in an informative way.
Similar to GPT-3, Gopher focused on improving dataset quality and optimization techniques:

• Dataset: The researchers curated a high-quality text dataset called MassiveText, which
contains over 10 terabytes of data and 2.45B documents from web pages, books, news
articles, and code (GitHub). They only trained on 300B tokens, which is 12% of the dataset.
Importantly, they improved the quality of the data by filtering it, such as by removing
duplicate text and deduplicating similar documents. This significantly improved the
model’s performance on downstream tasks.

• Optimization: The researchers used a warmup learning rate for 1,500 steps and then
decayed it using a cosine schedule. They also had an interesting rule that as they
increased the model size, they decreased the learning rate and increased the number of
tokens in each batch. Additionally, they found that clipping gradients to be a maximum of 1
based on the global gradient norm helped stabilize the training.

February 2025 29
Foundational Large Language Models & Text Generation

Gopher was evaluated on a variety of tasks, including mathematics, common sense, logical
reasoning, general knowledge, scientific understanding, ethics, and reading comprehension.
Gopher outperformed previous state-of-the-art models on 81% of the tasks. Specifically,
Gopher performed well on knowledge-intensive tasks but struggled on reasoning-heavy
tasks such as abstract algebra.

The authors also conducted a study on the effect of model size on different types of
tasks. Figure 4 shows the results of this ablation study. Specifically, the authors found that
increasing the number of parameters had a significant impact on logical reasoning and
reading comprehension, but it did not improve performance as much on tasks such as
general knowledge, where performance eventually almost plateaued.

Figure 5. Ablation study22 on the effect of model size on the performance of Gopher on different types
of tasks

February 2025 30
Foundational Large Language Models & Text Generation

GLaM

GLaM (Generalist Language Model)23 was the first sparsely-activated mixture-of-experts

language model. Mixture-of-experts based models are much more computationally efficient
given their parameter count. This is achieved by only activating a subset of their parameters
(i.e. experts) for each input token. GLaM consists of 1.2 trillion parameters but uses only ⅓
of the energy used to train GPT-3 and half of the FLOPs for inference while achieving better
overall performance compared to GPT-3.

Chinchilla

Until 2022, LLMs were primarily scaled by increasing the model size and using datasets that
are relatively small by current standards (up to 300 billion tokens for the largest models).
This approach was informed by the Kaplan et al.24 study, which examined how performance
of a language model, measured by cross-entropy loss, varies with changes in computational
budget, model size, and dataset size. Specifically, given a 100-fold increase in computational
resources (C), Kaplan et al.24 recommended scaling model size by approximately 28.8 times
(Nopt∝ C0.73), while increasing dataset size by only 3.5 times (Dopt∝ C0.27).

The Chinchilla paper,25 revisited the compute optimal scaling laws and used three different
approaches to find that near equal scaling in parameters and data is optimal with increasing
compute. Thus, a 100-fold increase in compute should translate into a tenfold increase in
both data size and model size.

February 2025 31
Foundational Large Language Models & Text Generation

Figure 6. Overlaid predictions from three different approaches from Chinchilla paper,25 along with
projections from Kaplan et al24

To verify the updated scaling law, DeepMind trained a 70B parameter model (called
Chinchilla) using the same compute budget as the previously trained Gopher model.
Chinchilla uniformly and significantly outperformed Gopher (280B),21 GPT-3 (175B),13 and
Megatron-Turing NLG (530B)26 on a large range of downstream evaluation tasks. Due to being
4x smaller than Gopher, both the memory footprint and the inference cost of Chinchilla are
also smaller.

The findings of Chinchilla had significant ramifications for the development of future LLMs.
Focus shifted into finding ways to scale dataset size (while maintaining quality) alongside
increasing parameter count. Extrapolating this trend suggests that training dataset size
may soon be limited by the amount of text data available. This has led to new research by
Muennighoff et al.27 exploring scaling laws in data-constrained regimes.

February 2025 32
Foundational Large Language Models & Text Generation

PaLM

Pathways language model (PaLM)28 is a 540-billion parameter transformer-based large

language model developed by Google AI. It was trained on a massive dataset of text and
code and is capable of performing a wide range of tasks, including common sense reasoning,
arithmetic reasoning, joke explanation, code generation, and translation.

At the time of its release, PaLM was also able to achieve state-of-the-art performance on
many language benchmarks, for example GLUE and SuperGLUE.29

One of the key features of PaLM is its ability to scale efficiently. This is thanks to the
Pathways system, which Google developed to distribute the training of large language
models across two TPU v4 Pods.

PaLM 2

PaLM 230 is a successor to PaLM that was announced in May 2023. Thanks to a number of
architectural and training enhancements, PaLM 2 is even more capable than PaLM, with
fewer total parameters. It excels at advanced reasoning tasks, including code generation,
math, classification, question answering, and translation.

PaLM 2 has also been shown to be more efficient than PaLM and became the basis for a
number of commercial models Google released as part of Google Cloud Generative AI.

February 2025 33
Foundational Large Language Models & Text Generation

Gemini

Figure 7. Gemini can receive multi-modal inputs including text, audio, images, and video data. These are all
tokenized and fed into its transformer model. The transformer generates an output that can contain images
and text.

Gemini31 (Figure 6) is a state-of-the-art multimodal language family of models that can

take interleaved sequences of text, image, audio, and video as input. It’s built on top of
transformer decoders and has architectural improvements for scale as well as optimized
inference on Google’s Tensor Processing Units (TPUs). In its current version, these models
are trained to support contexts of different sizes, up to 2M tokens in the Gemini Pro version
on Vertex AI and employ mechanisms such as multi-query attention for efficiency. Gemini
models also employ a Mixture of Experts architecture to optimize efficiency and capabilities
of the models. Multimodality allows the models to process text, images and video in input,
with more modalities in input and output expected in the future.

The Gemini models are trained on Google’s TPUv5e and TPUv4 processors, depending on
size and configuration. The pre-training data consists of web documents, books, code, and
image, audio, and video data.

February 2025 34
Foundational Large Language Models & Text Generation

Larger models are trained for the compute-optimal number of tokens using the same
approach as in Chinchilla paper,25 while small models are trained on significantly more tokens
than compute optimal to improve performance for a given inference budget.

The Gemini family of models is optimized for different sizes: Gemini Ultra, Gemini Pro, Gemini
Nano and Flash. Gemini Ultra is used for highly complex tasks and achieves state-of-the-
art results in 30 out of 32 benchmark tasks. Gemini Pro enables deployment at scale and
Gemini Nano is designed for on-device applications. The Gemini Nano models leverage
advancements such as distillation to produce state-of-the-art performance for small
language models on tasks such as summarization and reading comprehension. As the Gemini
models are natively multi-modal, it can be seen that training across multiple modalities does
indeed lead to a model that is capable of achieving strong capabilities in each domain.

During the initial part of 2024, Google introduced the latest model of the Gemini family,
Gemini 1.5 Pro,32 a highly compute-efficient multimodal mixture-of-experts model. This
model also dramatically increased the size of the context window to millions of tokens
and is capable of recalling and reasoning over those millions of tokens, including multiple
long documents and hours of video and audio. Gemini 1.5 Pro demonstrates remarkable
capabilities across different domains:

• Code understanding: It can process massive codebases and answer highly specific
code-related questions.

• Language learning: The model can learn new languages never observed at training time
solely based on reference materials provided within its input

• Multimodal reasoning: It understands images and text, allowing it to locate a famous scene
from the novel ‘Les Misérables’ based on a simple sketch.

• Video comprehension: It can analyze entire movies, answering detailed questions and
pinpointing specific timestamps with remarkable accuracy.

February 2025 35
Foundational Large Language Models & Text Generation

Google’s Gemini 1.5 Pro model excels at retrieving information from even very long
documents. In their study,32 it demonstrated 100% recall on documents up to 530,000
tokens, and over 99.7% recall on those up to 1 million tokens. Impressively, it maintains 99.2%
accuracy when finding information in documents up to 10 million tokens.

Moreover, Gemini 1.5 Pro demonstrates a major leap forward in how well LLMs follow complex
instructions. In a rigorous test with 406 multi-step prompts, it significantly outperformed
previous Gemini models. The model accurately followed almost 90% of instructions and fully
completed 66% of the complex tasks.

Gemini Flash is a new addition to the Gemini model family and the fastest Gemini model
served in the API. It’s optimized for high-volume, high-frequency tasks at scale, is more
cost-efficient to serve and features a breakthrough long context window of 1 million tokens.
Although it is a lighter weight model than 1.5 Pro, it is highly capable of multimodal reasoning
across vast amounts of information and delivers impressive quality for its size.

Gemini 2.0 represents a significant leap forward in Google’s multimodal AI models. It builds
upon the foundation of Gemini 1.0 with enhanced capabilities and a focus on efficiency and
new modalities. In more details :

Gemini 2.0 Flash: This version is designed for speed and efficiency, exceeding the
performance of Gemini 1.5 Pro while maintaining the responsiveness developers expect. It
showcases improvements in multimodal understanding, text processing, code generation,
video analysis, and spatial reasoning. Notably, it has enhanced spatial understanding, leading
to more accurate object identification and captioning, especially for small objects in complex
scenes. Gemini 2.0 Flash was introduced in late 2024.

February 2025 36
Foundational Large Language Models & Text Generation

Gemini 2.0 Pro is positioned as a highly capable model for a broad range of tasks. It likely
serves as a workhorse for various applications, balancing performance and efficiency.
It is likely an evolution of the original Gemini Pro model with improvements across
multiple domains.

Gemini 2.0 Nano: As with the previous generation, Nano focuses on on-device deployment.
It is optimized for resource efficiency and speed, enabling AI capabilities directly on devices
like smartphones.

Gemini 2.0 Flash Thinking Experimental is a fast, high-performance reasoning model,

enhanced with explainability through visible “thought processes,” particularly excelling
in complex science and math problems; it accepts text and image inputs, produces text
outputs, supports a 1 million token input context and a 64,000 token output, utilizes code
execution, has a knowledge cutoff of August 2024, is best suited for complex tasks where
latency is not a primary concern, and is available via Google AI Studio, the Gemini API, and
Vertex AI, though currently in an experimental deployment status.

Gemma

Furthermore, recently advanced Gemma is a family of lightweight, state-of-the-art open

models built from the same research and technology used to create the Gemini models.33 The
first model by Gemma boasts a large vocabulary of 256,000 words and has been trained on
a massive 6 trillion token dataset. This makes it a valuable addition to the openly-available
LLM collection. Additionally, the 2B parameter version is intriguing as it can run efficiently on
a single GPU.

February 2025 37
Foundational Large Language Models & Text Generation

Gemma 2,33 developed by Google AI, represents a significant advancement in the field of
open large language models. Designed with a focus on efficiency, the 27-billion parameter
model boasts performance comparable to much larger models like Llama 3 70B33 on standard
benchmarks. This makes Gemma 2 a powerful and accessible tool for a wide range of AI
developers. Its compatibility with diverse tuning toolchains, from cloud-based solutions
to popular community tools, further enhances its versatility. With its strong performance,
efficient architecture, and accessible nature, Gemma 2 plays a vital role in driving innovation
and democratizing AI capabilities.

Gemma 3 represents Google’s latest advancement in its family of open models, built upon
the research and technology that also powers the Gemini models. A key feature of Gemma
3 is its multimodality, enabling it to process both text and image inputs and generate text
outputs. This version significantly expands capabilities with a large, 128K context window, and
broad multilingual support encompassing over 140 languages. To cater to diverse hardware
and performance needs, Gemma 3 is available in various sizes, including 1B, 4B, 12B, and 27B
parameter models. These varieties allow developers to select the most suitable model for
their specific applications, ranging from resource-constrained devices to high-performance
computing environments.

LLaMA

Llama models are transformer-based language models, similar in high-level architecture

to other large language models (LLMs) like GPT. They are primarily based on the decoder-
only architecture, meaning they focus on predicting the next token in a sequence given the
preceding tokens.

February 2025 38
Foundational Large Language Models & Text Generation

Meta has released several key versions of Llama. The original Llama 1 models came in various
sizes, from 7B to 65B parameters, and were notable for their strong performance compared
to other open-source models of similar size. Llama 234 represented a major advancement,
featuring a larger context window extended to 4096 tokens for handling longer texts, and
importantly, it was fine-tuned for chat applications, significantly improving its conversational
abilities. Llama 2 was offered in 7B, 13B, and 70B parameter versions and, unlike Llama 1, was
released with a license allowing commercial use. Llama 3 builds upon these advancements
with enhanced performance across reasoning, coding, and general knowledge, and is
expected to include a wider range of sizes. A key focus of Llama 3 is increased safety, with
efforts to reduce harmful outputs through improved training and alignment techniques. Llama
2 is a family of pretrained and fine-tuned LLMs, ranging from 7B to 70B parameters, with
improvements like a 40% larger training dataset, doubled context length, and grouped-query
attention. The fine-tuned version, Llama 2-Chat, excels in dialogue. The next generation,
Llama 3.2, includes multilingual text-only models and vision LLMs, with quantized versions for
on-device deployment. Llama 3.2 uses grouped-query attention and a 128K token vocabulary.

Mixtral

Developed by Mistral AI35, Mixtral 8x7B is a Sparse Mixture of Experts (SMoE) model. While its
total parameter count is 47B, it utilizes only 13B active parameters per token during inference,
leading to faster inference and higher throughput. This model excels in mathematics, code
generation, and multilingual tasks, often outperforming LLaMA 2 70B in these domains.
Mixtral also supports a 32k token context length, enabling it to handle significantly longer
sequences. Its instruction-tuned version, Mixtral 8x7B- Instruct, surpasses several closed-
source models on human evaluation benchmarks. Mistral makes several of its models
available as open source under the Apache 2.0 license, emphasizing open access to model
weights. In addition, Mistral offers a range of models through its API, providing various sizes
and capabilities to suit diverse requirements.

February 2025 39
Foundational Large Language Models & Text Generation

OpenAI O1

OpenAI’s new “o1” series represents a significant advancement in models, focusing on

complex reasoning abilities honed through reinforcement learning. These models employ an
internal “chain-of-thought” process, engaging in extensive deliberation before generating
a response. This deliberate approach results in exceptional performance on challenging
scientific reasoning tasks. Benchmarks demonstrate their proficiency: o1 models achieve
an 89th percentile ranking on Codeforces programming competitions, score within the top
500 nationally on the AIME (a USA Math Olympiad qualifier), and surpass PhD-level human
accuracy on a comprehensive physics, biology, and chemistry benchmark (GPQA). The API
offers two variants: o1: The flagship model, optimized for tackling difficult problems that
require broad, general world knowledge, and o1-mini: A faster, more cost-effective version,
excelling in domains like coding, mathematics, and scientific tasks where deep specialized
knowledge is more critical than extensive general knowledge.

DeepSeek

DeepSeek has demonstrated that competitive reasoning performance, comparable to

OpenAI’s “o1” series, can be achieved through a novel reinforcement learning approach,
even without relying on extensive labeled data. This is exemplified by their DeepSeek-R1-
Zero model, trained purely with RL. Traditional RL methods for LLMs often depend on a
“critic” model, trained on labeled data, to provide feedback. DeepSeek’s innovation, called
Group Relative Policy Optimization (GRPO), eliminates this critic. Instead, GRPO uses a set
of predefined rules (assessing coherence, completeness, and fluency) to score the model’s
outputs across multiple rounds. The model learns by comparing its performance against the
group average, effectively learning from its own “self-play” without explicit human-provided

February 2025 40
Foundational Large Language Models & Text Generation

labels. This pure-RL approach, while successful in achieving high reasoning scores (matching
“o1” on the AIME 2024 mathematics competition), initially resulted in outputs with poor
readability and language mixing.

To address these shortcomings, DeepSeek developed a multi-stage training process for

their DeepSeek-R1 model. This process begins with supervised fine-tuning (SFT) on a small
“cold start” dataset, providing a basic foundation of language understanding. Next, pure-RL
(using GRPO) is applied to enhance reasoning abilities, similar to the R1-Zero model. Critically,
near the end of the RL phase, rejection sampling (e.g. filtering) is employed. The model
generates multiple outputs, and only the best, according to the GRPO rules, are selected.
This creates a high-quality “synthetic” dataset generated by the model itself. This synthetic
data is then combined with supervised data from the original base model (covering areas
like writing and factual knowledge). A final round of fine-tuning and further RL is performed,
leveraging both the synthetic and supervised data, refining the model’s overall performance
and generalization capabilities. This multi-stage approach leverages the strengths of each
training method: the initial SFT provides a basic linguistic foundation; pure-RL fosters strong
reasoning skills; rejection sampling creates high-quality training data; and the final SFT
and RL steps ensure a polished, well-rounded model. The result is the DeepSeek-R1 model
that matches or exceeds the o1 model in many areas. Chain-of-thought (CoT) reasoning at
inference time is intrinsically linked to this RL-based training. The model learns to generate
intermediate reasoning steps during training, which are essential for its strong performance
on complex tasks at inference. Despite providing model weights, DeepSeek’s models are
effectively closed-source due to the lack of transparency regarding training data, processing
scripts, and data curation methods.

Other open models

The landscape of open LLMs is rapidly evolving, with a growing number of models where
both the code and pre-trained weights are publicly accessible. Below we highlight some of
the known examples:

February 2025 41
Foundational Large Language Models & Text Generation

• Qwen 1.536: This LLM series from Alibaba comes in six sizes: 0.5B, 1.8B, 4B, 7B, 14B, and
72B. Qwen 1.5 models uniformly support a context length of up to 32k tokens and show
strong performance across various benchmarks. Notably, Qwen 1.5-72B outperforms
LLaMA2-70B on all evaluated benchmarks, demonstrating exceptional capabilities in
language understanding, reasoning, and math.

• Yi37: Created by 01.AI, the Yi model family includes 6B and 34B base models pre-trained
on a massive 3.1 trillion token English and Chinese dataset. Yi emphasizes data quality
through rigorous cleaning and filtering processes. The 34B model achieves performance
comparable to GPT-3.5 on many benchmarks and can be efficiently served on consumer-
grade GPUs with 4-bit quantization. Yi also offers extensions like a 200k context model, a
vision-language model (Yi-VL), and a depth-upscaled 9B model.

• Grok 3: Developed by xAI, Grok-3 is released in Grok 3 (Think) and Grok 3 mini (Think).
Both models were trained using reinforcement learning. Grok 3 (Think) learned to refine its
problem-solving strategies, correct errors through backtracking, simplify steps, and utilize
the knowledge it picked up during pretraining. With a context window of 1 million tokens its
8 times larger than previous Grok models.

The pace of innovation with LLMs has been rapid and shows no signs of slowing down. There
have been many contributions to the field in both the academic and commercial settings.
With over 20,000 papers published about LLMs in arxiv.org it is impossible to name all
of the models and teams that have contributed to the development of LLMs. However, an
abbreviated list of open models of interest could include EleutherAI’s GPT-NeoX and GPT-J,
Stanford’s Alpaca, Vicuna from LMSYS, Grok from xAI, Falcon from TII, PHI from Microsoft,
NVLM from Nvidia, DBRX from Databricks, Qwen from Alibaba, Yi from 01.ai, Llama from
Meta mentioned above and many others. Some of notable companies developing commercial
foundation LLM models include Anthropic, Cohere, Character.ai, Reka, AI21, Perplexity, xAI
and many others in addition to Google and OpenAI mentioned in previous sections. It is
important when using a model to confirm that the license is appropriate for your use case as
many models are provided with very specific terms of use.

February 2025 42
Foundational Large Language Models & Text Generation

Comparison

In this section, we observed how transformer-based language models have evolved. They
started as encoder-decoder architectures with hundreds of millions of parameters trained
on hundreds of millions of tokens, and have grown to be massive decoder-only architectures
with billions of parameters and trained on trillions of tokens. Table 1 shows how the
important hyperparameters for all the models discussed in this whitepaper have evolved
over time. The scaling of data and parameters has not only improved the performance of
LLMs on downstream tasks, but has also resulted in emergent behaviors and zero- or few-
shot generalizations to new tasks. However, even the best of these LLMs still have many
limitations. For example, they are not good at engaging in human-like conversations, their
math skills are limited, and they might not be aligned with human ethics (e.g., they might be
biased or generate toxic responses). In the next section, we learn how a lot of these issues
are being addressed.

February 2025 43
Foundational Large Language Models & Text Generation

Model
Attention GPT GPT-2 GPT-3 LaMDA Gopher Chinchilla
(2017) (2018) (2019) (2020) (2021) (2021) (2022)

Optimizer ADAM ADAM ADAM ADAM ADAM ADAM ADAM-W

# Parameters 213M 117M 1.5B 175B 137B 280B 70B
Vocab size ~37K ~40K ~50K ~50K ~32K ~32K ~32K
Embedding
dimension
1024 768 1600 12288 8192 16384 8192

Key dimension 64 64 64 128 128 128 128

# heads (H) 16 12 25 96 128 128 64
# encoder
layers
6 N/A N/A N/A N/A N/A N/A
# decoder
layers
6 12 48 96 64 80 80
Feed forward
dimension
4 * 1024 4 * 768 4 * 1600 4 * 12288 8 * 8192 4 * 16384 4 * 8192
Context Token
Size
N/A 512 1024 2048 N/A 2048 2048
Pre-Training
tokens
~160MA ~1.25BA ~10B ~300B ~168B ~300B ~1.4T

Table 1. Important hyperparameters for transformers-based large language models

A. This number is an estimate based on the reported size of the dataset.

February 2025 44
Foundational Large Language Models & Text Generation

Fine-tuning large language models

Large language models typically undergo multiple training stages. The first stage, often
referred to as pre-training, is the foundational stage where an LLM is trained on large,
diverse, and unlabelled text datasets where it’s tasked to predict the next token given the
previous context. The goal of this stage is to leverage a large, general distribution of data
and to create a model that is good at sampling from this general distribution. After language
model pretraining, the resulting LLM usually demonstrates a reasonable level of language
understanding and language generation skills across a variety of different tasks which
are typically tested through zero-shot or few-shot prompting (augmenting the instruction
with a few examples / demonstrations). Pretraining is the most expensive in terms of time
(from weeks to months depending on the size of the model) and the amount of required
computational resources, (GPU/TPU hours).

After training, the model can be further specialized via fine-tuning, typically called
instruction-tuning or simply supervised fine-tuning (SFT). SFT involves training an LLM on a
set of task-specific demonstration datasets where its performance is also measured across
a set of domain-specific tasks. The following are some examples of behaviors that can be
improved using fine-tuning:

• Instruction-tuning/instruction following: The LLM is provided as input an instruction to

follow which might include summarizing a piece of text, writing a piece of code, or writing
a poem in a certain style.17

• Dialogue-tuning: This is a special case of instruction tuning where the LLM is fine-tuned
on conversational data in the form of questions and responses. This is often called
multi-turn dialogue.39

February 2025 45
Foundational Large Language Models & Text Generation

• Safety tuning: This is crucial for mitigating risks associated with bias, discrimination, and
toxic outputs. It involves a multi-pronged approach encompassing careful data selection,
human-in-the-loop validation, and incorporating safety guardrails. Techniques like
reinforcement learning with human feedback (RLHF)40 enable the LLM to prioritize safe
and ethical responses.

Fine-tuning is considerably less costly and more data efficient compared to pre-training.
Numerous techniques exist to optimize the costs further which are discussed later in
this whitepaper.

Supervised fine-tuning

As mentioned in the previous section, SFT is the process of improving an LLM’s performance
on a specific task or set of tasks by further training it on domain-specific, labeled data. The
dataset is typically significantly smaller than the pre-training datasets, and is usually human-
curated and of high quality.

In this setting, each data point consists of an input (prompt) and a demonstration (target
response). For example, questions (prompt) and answers (target response), translations from
one language (prompt) to another language (target response), a document to summarize
(prompt), and the corresponding summary (target response).

It’s important to note that, while fine-tuning can be used to improve the performance on
particular tasks as mentioned above, it can also serve the purpose of helping the LLM
improve its behavior to be safer, less toxic, more conversational, and better at following
instructions.

February 2025 46
Foundational Large Language Models & Text Generation

Reinforcement learning from human feedback

Typically, after performing SFT, a second stage of fine-tuning occurs which is called
reinforcement learning from human feedback (RLHF). This is a very powerful fine-tuning
technique that enables an LLM to better align with human-preferred responses (i.e. making
its responses more helpful, truthful, safer, etc.).

Figure 8. An example RLHF procedure

In contrast to SFT, where an LLM is only exposed to positive examples (e.g. high-quality
demonstration data), RLHF makes it possible to also leverage negative outputs thus
penalizing an LLM when it generates responses that exhibit undesired properties. Penalizing
negative output makes it less likely to generate unhelpful or unsafe responses.

To leverage RLHF, a reward model (RM) typically needs to be trained with a procedure similar
to that in Figure 8. An RM is usually initialized with a pretrained transformer model, often also
one that is SFT. Then it is tuned on human preference data which is either single sided (with a
prompt, response and a score) or composed of a prompt and a pair of responses along with

February 2025 47
Foundational Large Language Models & Text Generation

a preference label indicating which of the two responses was preferred. For example, given
two summaries, A and B, of the same article, a human rater selects a preferred summary
(relying on the detailed guidance). We refer to the provided preference labels as human
feedback. Preferences can be in the binary form (e.g. ‘good’ or ‘bad’), on the Likert scale42,
rank order when more than 2 candidates are evaluated, or a more detailed assessment of the
summary quality. The preference signal can also incorporate many dimensions that capture
various aspects that define a high quality response, e.g., as safety, helpfulness, fairness, and
truthfulness.

Figure 8 shows a typical RLHF pipeline where a Reward model is initialized and finetuned on
preference pairs. Once an RM has been trained, it’s then used by a Reinforcement Learning
(RL)43 policy gradient algorithm, which further finetunes a previously instruction-tuned LLM to
generate responses that are better aligned with human preferences.

To better scale RLHF, RL from AI Feedback (RLAIF)44 leverages AI feedback instead of human
feedback to generate preference labels. It’s also possible to remove the need for training
RLHF by leveraging approaches such as direct preference optimization (DPO).45 Both RLHF
and RLAIF can be used on Google Cloud.

February 2025 48
Foundational Large Language Models & Text Generation

Parameter Efficient Fine-Tuning

Both SFT and RLHF are still very costly in terms of compute time and accelerators required,
especially when full-fine tuning entire LLMs on the orders of billions of parameters. Luckily,
there are some really useful and effective techniques that can make fine-tuning significantly
cheaper and faster compared to pre-training and full fine-tuning. One such family of
methods is parameter efficient fine-tuning (PEFT) techniques.

At a high-level, PEFT approaches append a significantly smaller set of weights (e.g., on the
order of thousands of parameters) that are used to ‘perturb’ the pre-trained LLM weights.
The perturbation has the effect of fine-tuning the LLM to perform a new task or set of tasks.
This has the benefit of training a significantly smaller set of weights, compared to traditional
fine-tuning of the entire model.

Some common PEFT techniques include the adapter, low-rank adaptation, and
soft prompting:

• Adapter-based fine-tuning46 employs small modules, called adapters, to the pre-

trained model. Only the adapter parameters are trained, resulting in significantly fewer
parameters than traditional SFT.

• Low-Rank Adaptation (LoRA)47 tackles efficiency differently. It uses two smaller matrices
to approximate the original weight matrix update instead of fine-tuning the whole LLM.
This technique freezes the original weights and trains these update matrices, significantly
reducing resource requirements with minimum additional inference latency. Additionally,
LoRA has improved variants such as QLoRA,48 which uses quantized weights for even
greater efficiency. A nice advantage of LoRA modules is that they can be plug-and-play,
meaning you can train a LoRA module that specializes in one task and easily replace it with
another LoRA module trained on a different task. It also makes it easier to transfer the
model since assuming the receiver has the original matrix, only the update matrices need
to be provided.

February 2025 49
Foundational Large Language Models & Text Generation

• Soft prompting49 is a technique for conditioning frozen large language models with
learnable vectors instead of hand-crafted text prompts. These vectors, called soft
prompts, are optimized on the training data and can be as few as five tokens, making them
parameter-efficient and enabling mixed-task inference.

For most tasks, full fine-tuning is still the most performant, followed by LoRA and Soft
prompting, but the order is reversed when it comes to cost. All three approaches are more
memory efficient than traditional fine-tuning and achieve comparable performance.

February 2025 50
Foundational Large Language Models & Text Generation

Python

# Before you start run this command:

# pip install --upgrade --user --quiet google-cloud-aiplatform
# after running pip install make sure you restart your kernel

import vertexai
from vertexai.generative_models import GenerativeModel
from vertexai.preview.tuning import sft

# TODO : Set values as per your requirements

# Project and Storage Constants
PROJECT_ID = ‘<project_id>’
REGION = ‘<region>’

vertexai.init(project=PROJECT_ID, location=REGION)

# define training & eval dataset.

TRAINING_DATASET = ‘gs://cloud-samples-data/vertex-ai/model-evaluation/
peft_train_sample.jsonl’
# set base model and specify a name for the tuned model
BASE_MODEL = ‘gemini-1.5-pro-002’
TUNED_MODEL_DISPLAY_NAME = ‘gemini-fine-tuning-v1’

# start the fine-tuning job

sft_tuning_job = sft.train(
source_model=BASE_MODEL,
train_dataset=TRAINING_DATASET,
# # Optional:
tuned_model_display_name=TUNED_MODEL_DISPLAY_NAME,
)

# Get the tuning job info.

sft_tuning_job.to_dict()

# tuned model endpoint name

tuned_model_endpoint_name = sft_tuning_job.tuned_model_endpoint_name

# use the tuned model

tuned_genai_model = GenerativeModel(tuned_model_endpoint_name)
print(tuned_genai_model.generate_content(contents=’What is a LLM?’))

Snippet 1. SFT fine tuning on Google cloud

February 2025 51
Foundational Large Language Models & Text Generation

Using large language models

Prompt engineering and sampling techniques have a strong influence on the performance of
LLMs. Prompt engineering is the process of designing and refining the text inputs (prompts)
that you feed into an LLM to achieve desired and relevant outputs. Sampling techniques
determine the way in which output tokens are chosen and influence the correctness,
creativity and diversity of the resulting output. We next discuss different variants of prompt
engineering and sampling techniques as well as touch on some important parameters that
can have a significant impact on LLM performance.

Prompt engineering

LLMs are very powerful, but they still need guidance to unleash their full potential. Prompt
engineering is a critical component in guiding an LLM to yield desired outputs. This might
include grounding the model to yield factual responses or unleashing the creativity of the
model to tell a story or write a song. Examples of prompt engineering include providing
clear instructions to the LLM, giving examples, using keywords, and formatting to emphasize
important information, providing additional background details etc.

You will often hear the terms zero-shot, few-shot, and chain-of-thought prompting in the
context of prompt engineering. We define these terms below:

• Few-shot prompting: This is when you provide the LLM with a task description, as well
as a few (e.g. three to five) carefully chosen examples, that will help guide the LLM’s
response. For example, you might provide the model with the name of a few countries
and their capital cities, then ask it to generate the capital for a new country that isn’t in
the examples.

February 2025 52
Foundational Large Language Models & Text Generation

• Zero-shot prompting: This is when you provide the LLM directly with a prompt with
instructions. You usually give the LLM a task description and the LLM relies heavily on its
existing knowledge to output the correct response. This requires no additional data or
examples, hence the name ‘Zero-shot’ but can be less reliable than few-shot prompting.

• Chain-of-thought prompting: This technique aims to improve performance on complex

reasoning tasks. Rather than simply asking the LLM a question, you provide a prompt
that demonstrates how to solve similar problems using step-by-step reasoning. The
LLM then generates its own chain of thought for the new problem, breaking it down into
smaller steps and explaining its reasoning. Finally, it provides an answer based on its
reasoning process.

Prompt engineering is an active area of research.

Sampling Techniques and Parameters

A variety of sampling techniques can be employed to determine how the model chooses
the next token in a sequence. They are essential for controlling the quality, creativity, and
diversity of the LLM’s output. The following is a breakdown of different sampling techniques
and their important parameters:

• Greedy search50: Selects the token with the highest probability at each step. This is the
simplest option but it can lead to repetitive and predictable outputs.

• Random sampling:50 Selects the next token according to the probability distribution, where
each token is sampled proportionally to its predicted probability. This can produce more
surprising and creative text, but also a higher chance of nonsensical output.

• Temperature sampling:50 Adjusts the probability distribution by a temperature parameter.

Higher temperatures promote diversity, lower temperatures favor high-probability tokens.

February 2025 53
Foundational Large Language Models & Text Generation

• Top-K sampling: Randomly samples from the top K most probable tokens. The value of K
controls the degree of randomness.

• Top-P sampling (nucleus sampling):51 Samples from a dynamic subset of tokens whose
cumulative probability adds up to P. This allows the model to adapt the number of potential
candidates depending on its confidence, favoring more diversity when uncertain and
focusing on a smaller set of highly probable words when confident.

• Best-of-N sampling: Generates N separate responses and selects the one deemed best
according to a predetermined metric (e.g., a reward model or a logical consistency check).
This is particularly useful for short snippets or situations where logic and reasoning
are key.

By combining prompt engineering with sampling techniques and correctly calibrated

hyperparameters, you can greatly influence the LLM’s response, making it more relevant,
creative, and consistent for your specific needs.

Until now, we have seen the various types of LLM architectures, their underlying technology,
as well as the approaches used to train, tune, and adapt these models for various tasks. Let’s
now look at some key research about how the decoding process in LLMs can be sped up
considerably to generate faster responses.

Task-based Evaluation

The emergence of LLMs has reduced the obstacles to building AI applications, but moving
from MVP to production introduces challenges such as prompt engineering, model selection,
and performance monitoring. A tailored evaluation framework is essential for navigating
LLM application development by validating functionality and user experience, identifying

February 2025 54
Foundational Large Language Models & Text Generation

potential issues, facilitating communication about capabilities, and establishing a roadmap

for improvement. For building a tailored evaluation framework, application builders need to
provide their own evaluation data, development context, a definition of good performance.

• Evaluation data: Public leaderboards that showcase LLM capabilities often fall short for
application developers who require a more tailored approach. A dedicated evaluation
dataset that mirrors the expected production traffic as closely as possible is needed.
During prototyping this can be a manually curated dataset, one that is continuously
enriched with real user interactions, production logs, and synthetically generated data to
test specific scenarios.

• Development Context: The evaluation should extend beyond just the model’s output
to analyze the entire system, including components like data augmentation (e.g.,
Retrieval Augmented Generation or RAG) and agentic workflows. This approach ensures
understanding of how all components interact and contribute to the application’s
overall performance.

• Definition of “Good”: Traditional metrics that prioritize matching a single “correct”

answer can unfairly penalize unexpected solutions. When working with LLMs, we can
address this by moving beyond similarity to ground truth as a definition of good, but
rather establish dataset level criteria that reflect desired business outcomes or even
rubrics that capture the core elements of the desired outputs depending on the input user
instructions.

Application builders can evaluate LLM performance using three methods:

• Traditional Evaluation Methods: Similar to evaluating predictive models, these methods

use quantitative metrics to compare model outputs to ideal responses, aiming to offer
objective insights. However, they may penalize creative or unexpected outputs, limiting
their effectiveness for generative tasks that have multiple possible solutions.

February 2025 55
Foundational Large Language Models & Text Generation

• Human Evaluation: Considered the gold standard, human judgment provides nuanced
assessment of complex generative outputs.

• LLM-Powered Autoraters: LLM-powered autoraters try to mimic human judgment,

offering scalable and efficient evaluations. Unlike computation-based methods, they
can operate with or without reference data. A basic setup involves providing the task,
criteria, and candidate responses (with optional references), which the autorater uses to
generate and parse LLM output for final evaluation results. In addition to the final outputs,
autoraters can provide rationales to explain a given decision to the user. While generative
models are common autoraters, reward and discriminative models are also used. Crucially,
like any measurement tool, autoraters require calibration. Meta-evaluation, i.e. comparing
autorater outputs to human judgments, ensures the autorater aligns with desired
preferences. This calibration typically involves agreement in terms of model preference
or correlation measures, tailored to the evaluation task. In this meta-evaluation, it is
important to keep in mind potential limitations of autorater models.69

While the basic setup described above focuses on providing autoraters with fixed criteria
that can be used to evaluate an entire dataset of examples, approaches are emerging to
leverage rubrics and multi-step processes in order to obtain interpretable evaluation metrics.
At a high level, an LLM breaks an example down into multiple evaluation subtasks and then
evaluates each subtask to give an interpretable, detailed report for this example. As the
task has been broken down, domain-specialized models can be leveraged for specific tasks
to improve reliability if necessary. Results are then aggregated for a given example to yield
an overall score, or across related subtasks to evaluate how well a model performs along a
particular axis. This setup is especially useful in media generation, where different examples
may require vastly different skills (e.g. object generation vs text generation) and a single
score obfuscates how relative strengths and weaknesses of different models contribute to
the final result.

February 2025 56
Foundational Large Language Models & Text Generation

Accelerating inference
The scaling laws for LLMs which were initially explored by the Kaplan et al.24 study continue
to hold today. Language models have been consistently increasing in size and this has been
a direct contributor to the vast improvement in these models’ quality and accuracy over the
last few years. As increasing the number of parameters has improved the quality of LLMs it
has also increased the computational resources needed to run them. Numerous approaches
have been used to try and improve the efficiency of LLMs for different tasks as developers
are incentivized to reduce cost and latency for model users. Balancing the expense of
serving a model in terms of time, money, energy is known as the cost-performance tradeoff
and often needs adjusting for particular use cases.

Two of the main resources used by LLMs are memory and computation. Techniques for
improving the efficiency or speed of inference focus primarily on these resources. The
speed of the connection between memory and compute is also critical, but usually hardware
constrained. As LLMs have grown in size 1000x from millions to billions of parameters.
Additional parameters increase both the size of memory required to hold the model and
computations needed to produce the model results.

With LLMs being increasingly adopted for large-scale and low-latency use cases, finding
ways to optimize their inference performance has become a priority and an active research
topic with significant advancements. We will explore a number of methods and a few
tradeoffs for accelerating inference.

February 2025 57
Foundational Large Language Models & Text Generation

Trade offs

Many of the high yielding inference optimisation methods mandate trading off a number of
factors, this can be tweaked on a case-by-case basis allowing for tailored approaches to
different inference use cases and requirements. A number of the optimization methods we
will discuss later fall somewhere on the spectrum of these tradeoffs.

Trading off one factor against the other (e.g. latency vs quality or cost) doesn’t mean that
we’re completely sacrificing that factor, it just means that we’re accepting what might be
a marginal degradation in quality, latency or cost for the benefit of substantially improving
another factor.

The Quality vs Latency/Cost Tradeoff

It is possible to improve the speed and cost of inference significantly through accepting
what might be marginal to negligible drops in the model’s accuracy. One example of this
is using a smaller model to perform the task. Another example is quantisation where we
decrease the precision of the model’s parameters thereby leading to faster and less memory
intensive calculations.

One important distinction when approaching this trade-off is between the theoretical
possibility of a quality loss versus the practical capability of the model to perform the desired
task. This is use case specific and exploring it will often lead to significant speedups without
sacrificing quality in a meaningful or noticeable way. For example, if the task we want the
model to perform is simple, then a smaller model or a quantised one will likely be able to
perform this task well. Reduction in parametric capacity or precision does not automatically
mean that the model is less capable at that specific task.

February 2025 58
Foundational Large Language Models & Text Generation

The Latency vs Cost Tradeoff

Another name for this tradeoff is the latency vs throughput tradeoff. Where throughput refers
to the system’s ability at handling multiple requests efficiently. Better throughput on the same
hardware means that our LLM inference cost is reduced, and vice versa.

Much like traditional software systems, there are often multiple opportunities to tradeoff
latency against the cost of LLM inference. This is an important tradeoff since LLM inference
tends to be the slowest and most expensive component in the entire stack; balancing latency
and cost intentionally is key to making sure we tailor LLM performance to the product or use
case it’s being used in. An example would be bulk inference use cases (e.g. offline labeling)
where cost can be a more important factor than the latency of any particular request. On the
other hand, an LLM chatbot product will place much higher importance on request latency.

Now that we’ve covered some of the important tradeoffs to consider when optimizing
inference, let’s examine some of the most effective inference acceleration techniques. As
discussed in the tradeoffs section, some optimization techniques can have an impact on the
model’s output. Therefore we will split the methods into two types: output-approximating
and output-preserving.

As of this writing, Gemini 2.0 Flash Thinking offers an unparalleled balance of quality, as
measured by its ELO score, and affordability, with a cost per million tokens that is ten times
lower than comparable models; its position on a quality-versus-cost graph (where further
up and to the right is superior) demonstrates its transformative development. Moreover, the
picture highlights the rapid advancements in reasoning and thinking capabilities within the AI
field, with a 27-fold improvement observed in the last three months.

February 2025 59
Foundational Large Language Models & Text Generation

Output-approximating methods

Quantization

LLMs are fundamentally composed of multiple numerical matrices (a.k.a the model weights).
During inference, matrix operations are then applied to these model weights to produce
numerical outputs (a.k.a activations). Quantization is the process of decreasing the numerical
precision in which weights and activations are stored, transferred and operated upon. The
default representation of weights and activations is usually 32 bits floating numbers, with
quantization we can drop the precision to 8 or even 4 bit integers.

Quantization has multiple performance benefits, it reduces the memory footprint of

the model, allowing to fit larger models on the same hardware, it also reduces the
communication overhead of weights and activations within one chip and across chips in
a distributed inference setup- therefore speeding up inference as communication is a
major contributor to latency. In addition, decreasing the precision of weights/activations
can enable faster arithmetic operations on these models as some accelerator hardware
(e.g. TPUs/GPUs) natively supports faster matrix multiplication operations for some lower
precision representations.

Quantization’s impact on quality can be very mild to non-existent depending on the use
case and model. Further, in cases where quantisation might introduce a quality regression,
that regression can be small compared to the performance gain, therefore allowing for an
effective Quality vs Latency/Cost Tradeoff. For example, Benoit Jacob et al.55 reported a 2X
speed-up for a 2% drop in accuracy for the FaceDetection task on MobileNet SSD.

February 2025 60
Foundational Large Language Models & Text Generation

Quantization can be either applied as an inference-only operation, or it can be incorporated

into the training (referred to as Quantisation Aware Training QAT). QAT is generally
considered to be a more resilient approach as the model is able to recover some of the
quantisation-related quality losses during training. To make sure we get the best cost/quality
tradeoff, we tweak the quantization strategy (e.g. select different precisions for weights
vs activations) and the granularity in which we apply quantisation to Tensors (e.g. channel
or group-wise58).

Distillation

Using a smaller model to perform a task is one of the most efficient inference optimization
techniques, however, smaller models can demonstrate significant regressions on quality
compared to their larger counterparts.

Distillation is a set of training techniques that targets improving the quality of a smaller model
(the student) using a larger model (the teacher). This method can be effective because larger
models outperform smaller ones even if both are trained on the same data, mainly due to
parametric capacity and training dynamics. The gap in performance continues as the training
dataset grows as illustrated by Figure 9.

It is worth noticing that even at low volumes of training data, large models can already
demonstrate better performance than the correspondingly trained smaller models, this fact
leads us to the first variant of distillation which is referred to as data distillation or model
compression.56 We use a large model which was trained on the data we have to generate
more synthetic data to train the smaller student model, the increase in data volume will help
move the the student further along the quality line compared to only training on the original
data. Synthetic data needs to be approached carefully as it needs to be of high quality and
can lead to negative effects otherwise.

February 2025 61
Foundational Large Language Models & Text Generation

Figure 9. An illustration of the performance of models of various sizes as a function of the training
dataset’s size

Other distillation techniques attempt to bring the student model closer to the teacher
on a more granular level than just synthetic data generation. One prominent technique is
knowledge distillation57, in this approach we attempt to align the output token distribution
of the student model to that of the teacher’s, this can be much more sample efficient than
data distillation. On-policy distillation59 is another technique that leverages feedback from
the teacher model on each sequence generated by the student in a reinforcement learning
setup.

Output-preserving methods

These methods are guaranteed to be quality neutral, they cause no changes to the model
output which often makes them obvious first steps to optimize inference before facing the
more nuanced tradeoffs of the approximating methods

February 2025 62
Foundational Large Language Models & Text Generation

Flash Attention

Scaled Dot-product Attention, which is the predominant attention mechanism in the

transformer architecture, is a quadratic operation on the input length. Optimizing the self-
attention calculation can bring significant latency and cost wins.

Flash Attention, introduced in by Tri Dao et al.62, optimizes the attention calculation by making
the attention algorithm IO Aware, particularly trying to minimize the amount of data we move
between the slow HBM (high bandwidth memory) to the faster memory tier (SRAM/VMEM) in
TPUs and GPUs. When calculating attention, the order of operations is changed and multiple
layers are fused so we can utilize the faster memory tiers as efficiently as possible.

Flash Attention is an exact algorithm, it maintains the numerical output of the attention
computation and can yield significant latency benefits due to reducing the IO overhead, Tri
Dao et al.62 showed 2-4X latency improvements in the attention computation.

Prefix Caching

One of the most compute intensive, and thus slowest, operations in LLM inference is
calculating the attention key and value scores (a.k.a KV) for the input we’re passing to the
LLM, this operation is often referred to as prefill. The final output of prefill is what is termed
KV Cache which includes the attention key and value scores for each layer of the transformer
for the entire input. This cache is vital during the decoding phase which produces the output
tokens, the KV cache allows us to avoid recalculating attention scores for the input on each
autoregressive decode step.

Prefix Caching refers to the process of caching the KV Cache itself between subsequent
inference requests in order to reduce the latency and cost of the prefill operation. The way
the self-attention mechanism works makes reusing KV caches possible because tokens will

February 2025 63
Foundational Large Language Models & Text Generation

only pay attention to tokens that came before them in the sequence. If there’s new input
being appended to input that the model has seen before, then we can potentially avoid
recalculating the prefill for the older input.

Figure 10. An illustration of Prefix Caching in a chat scenario

Figure 10 illustrates how prefix caching works in a multi-turn scenario with a document
upload. On the first user turn, the prefill operation has to process the entire document
therefor taking 500ms, the resulting KV cache is then stored so that on the second user turn,
we can retrieve the cache directly from storage and avoid recomputing it for the long doc,
therefore saving substantial amounts of compute and latency.

February 2025 64
Foundational Large Language Models & Text Generation

Prefix caches can be stored either in memory or on disk and fetched on-demand. One
important consideration is making sure that the input structure/schema remains prefix-
caching friendly, we should avoid changing the prefix in subsequent requests as that will
invalidate the cache for all the tokens that follow For example, putting a fresh timestamp at
the very beginning of each request will invalidate the cache completely as every subsequent
request will have a new prefix.

Many LLM use cases lend themselves naturally to prefix caching. For example, LLM Chatbots
where users will have a multi-turn conversation that can span 10s of 1000s of tokens and
we can avoid recalculating the KV cache for the previous parts of the conversation. Large
document/code uploads is another use case where the artifact the user uploads will remain
unchanged from one request to the next. All that’s changing are the questions the user is
asking, so caching the KV cache for the document (especially for larger artifacts) can result
in significant latency and cost savings.

Prefix caching is available as a service called Context Caching on Google AI studio52 and
Vertex AI on Google Cloud53.

Speculative Decoding

The first phase of LLM inference, known as prefill, is compute bound due large matrix
operations on many tokens occurring in parallel. The second phase, known as decode, is
generally memory bound as tokens are auto-regressively decoded one at a time.

It is not easy to naively use additional parallel compute capacity to speed up decode
given the need to wait for the current token to be produced before we can calculate what
the next token should be (as per the self-attention mechanism), the decode process is
inherently serial.

February 2025 65
Foundational Large Language Models & Text Generation

Speculative decoding (Leviathan at al.63) aims to overcome this limitation in decode by finding
a way to utilize the spare compute capacity to make each decode step faster. The main idea
is to use a much smaller secondary model (often referred to as the drafter) to run ahead of
the main model and predict more tokens. (e.g. 4 tokens ahead). This will happen very quickly
as the drafter is much faster and smaller than the main model. We then use the main model to
verify the hypotheses of the drafter in parallel for each of the 4 steps (i.e. the first token, the
first two tokens, the first 3 tokens and finally all 4 tokens), and we then select the accepted
hypothesis with the maximum number of tokens. For example:

Figure 11. An illustration of speculative decoding over 3 tokens

Note that the 3 main model steps run in parallel. And because we are not compute bound in
decode, we can use the spare capacity to get much better decode latencies. In the example
above, let’s say a single main model step needs 10ms, while the drafter needs 1ms. Without
speculative decoding, we need 3 * 10ms = 30ms to produce the response, with speculative
decoding, there’s only one main model step on the critical path due to parallelization, so we
need 3 * 1ms + 10ms = 13ms. A significant latency improvement. This technique is completely

February 2025 66
Foundational Large Language Models & Text Generation

quality neutral, the main model will reject any tokens that it wouldn’t have predicted itself
in the first place, so the only thing speculative decoding does is run ahead and present
hypotheses that the main model can accept or reject in parallel.

One important condition for speculative decoding to work effectively is that the drafter model
has good levels of alignment with the main model, otherwise we won’t be able to accept any
of the tokens. So investing in the training quality of the drafter model is worthwhile to get
better latencies.

Now that we have seen some methods to make LLM generate their responses faster, let’s
look at some examples of how these models can be applied to various tasks to get an idea
how to use them.

Batching and Parallelization

Most of the optimization techniques we’ve discussed so far are specific to Machine Learning
and Transformer architecture in particular. However, much like any software system, there
are opportunities to improve throughput and latency through a combination of 1) batching
less compute-intensive operations (i.e. we can run multiple requests on the same hardware
simultaneously to better utilize the spare compute) and 2) parallelizing the more compute-
intensive parts of the computations (i.e. we can divide the computation and split it amongst
more hardware instances to get more compute capacity and therefore better latencies

Batching in LLMs is most useful on the decode side - as we explained in the Speculative
Decoding section, decode is not compute-bound and therefore there’s an opportunity
to batch more requests. We need to be careful that we batch computations in a way that
enables utilization of the spare capacity which is possible to do on accelerators (e.g. TPUs
and GPUs). We also need to make sure we remain within the memory limits, as decode is a

February 2025 67
Foundational Large Language Models & Text Generation

memory intensive operations, batching more requests will put more pressure on the free
memory available. Batching has become an important component in most high-throughput
LLM inference setups.

Parallelization is also a widely used technique given the variety of opportunities in

transformers for horizontal scaling across more hardware instances. There are multiple
parallelism techniques across the model input (Sequence parallelism) the model layers
(Pipeline parallelism), and within a single layer (Tensor parallelism). One of the most important
considerations for parallelism is the cost of communication and synchronization between
the different shards that we distribute to other machines. Communication is a significant
overhead and can erode the benefits of adding more computational capacity if we’re not
careful about which parallelization strategy to use. On the other hand, selecting the right
strategy to balance the need for additional compute and the communication cost can yield
significant latency wins.

Now that we have seen some methods to make LLM generate their responses faster, let’s
look at some examples of how these models can be applied to various tasks to get an idea
how to use them.

Applications
Large language models are revolutionizing the way we interact with and process information.
With their unprecedented ability to understand context and generate content, they’re
transforming numerous applications in the worlds of text, code, images, audio and video.
Here we collected a few examples of application areas, but the reader should keep in mind
that this is not a comprehensive list and that many new ideas are emerging continuously
about how to best utilize the capabilities of these new tools. For more information about
optimally building and deploying functioning applications based on the following mentioned
use cases, refer to the subsequent whitepapers.

February 2025 68
Foundational Large Language Models & Text Generation

It is also very simple to generate text-based responses for your use case using either
the Google Cloud Vertex AI SDK or the Developer focused AI studio. Snippet 3 shows
code examples from these SDKs to generate responses to text prompts using the Gemini
model. Note that the multimodal aspects of Gemini are covered in their respective
dedicated whitepapers.

February 2025 69
Foundational Large Language Models & Text Generation

Python

%pip install --upgrade --quiet google-genai

import sys

if “google.colab” in sys.modules:
from google.colab import auth

auth.authenticate_user()

from IPython.display import HTML, Markdown, display

from google import genai
from google.genai.types import (
FunctionDeclaration,
GenerateContentConfig,
GoogleSearch,
HarmBlockThreshold,
HarmCategory,
MediaResolution,
Part,
Retrieval,
SafetySetting,
Tool,
ToolCodeExecution,
VertexAISearch,
)

import os

PROJECT_ID = “[your-project-id]” # @param {type: “string”, placeholder: “[your-project-

id]”, isTemplate: true}
if not PROJECT_ID or PROJECT_ID == “[your-project-id]”:
PROJECT_ID = str(os.environ.get(“GOOGLE_CLOUD_PROJECT”))

LOCATION = os.environ.get(“GOOGLE_CLOUD_REGION”, “us-central1”)

client = genai.Client(vertexai=True, project=PROJECT_ID, location=LOCATION)

MODEL_ID = “gemini-2.0-flash-001” # @param {type: “string”}

response = client.models.generate_content(
model=MODEL_ID, contents=”What’s the largest planet in our solar system?”
)

display(Markdown(response.text))

Snippet 2. Using Vertex AI and Google AI studio SDKs for unimodal text generation

February 2025 70
Foundational Large Language Models & Text Generation

Code and mathematics

Generative models can comprehend and generate code and algorithms to supercharge
developers by assisting them across many application areas. Some of the popular use cases
for code include:

• Code generation: LLMs can be prompted in natural language to generate code in a

specific programming language to perform certain operations. The output can be used as
a draft.

• Code completion: LLMS can proactively suggest useful code as the user types it. This
can save developers time and improve code quality.

• Code refactoring and debugging: LLMs can help reduce technical debt by refactoring
and debugging code to improve quality, efficiency and correctness.

• Code translation: LLMs can significantly help developer time and effort by helping to
convert code from one programming language to another. For example, an LLM might
convert Python code to Java.

• Test case generation: LLMs can be prompted to generate unit tests for a provided
codebase which saves considerable time and reduces errors.

• Code documentation and understanding: LLMs can be used in a conversational manner

to engage in a natural language chat to help you understand a codebase. They can also
generate appropriate comments, understand copyright status, and create release notes.

Recently, a number of exciting advancements have been made in the space of competitive
coding and mathematics. AlphaCode 2,64 combines Gemini’s reasoning capabilities with
search and the use of tools to solve competitive coding problems. It receives as input a
description of a problem to solve, and outputs a code solution that solves the problem. It

February 2025 71
Foundational Large Language Models & Text Generation

now ranks among the top 15% competitive coders on the popular Codeforces competitive
coding platform. FunSearch65 uses an evolutionary procedure which is based on pairing
a pre-trained LLM with a systematic evaluator. It solved the cap set problem66, an open
problem in mathematics, and also discovered more efficient bin-packing algorithms which
are used in many applications such as making data centers more efficient. Another recent
approach called AlphaGeometry tackles the problem of finding proofs for complex geometric
theorems. It comprises a neuro-symbolic system made up of a neural language model and
a symbolic deduction engine. AlphaGeometry managed to solve 25 out of 30 Olympiad
geometry problems, where the average human gold medalist scores on average 25.9. 67

Machine translation

LLMs are capable of generating fluid, high-quality and contextually accurate translations.
This is possible due to the LLM’s deep understanding of linguistic nuances, idioms, and
context. The following are some possible real world use cases:

• Instant messaging apps: In messaging platforms, LLMs can provide on-the-fly

translations that feel natural. Unlike previous algorithms that might translate word-
for-word, LLMs understand slang, colloquialisms, and regional differences, enhancing
cross-language communication.

• E-commerce: On global platforms like AliExpress, product descriptions are automatically

translated. LLMs help with ensuring cultural nuances and idiomatic expressions in product
details are appropriately translated, leading to fewer misunderstandings.

• Travel apps: In apps like Google Translate, travelers get real-time spoken translations.
With LLMs, the translated conversations are smoother, making interactions in foreign
countries more effortless.

February 2025 72
Foundational Large Language Models & Text Generation

Text summarization

Text summarization is a core capability of many of the LLMs mentioned in this whitepaper.
There are a number of natural potential use cases which include:

• News aggregators: LLMs could craft summaries that capture not only the main
events but also the sentiment and tone of the article, providing readers with a more
holistic understanding.

• Research databases: LLMs could help researchers generate abstracts that encapsulate
the core findings and implications of scientific papers.

• Chat management: In platforms like Google Chat, LLM-based systems could generate
thread summaries that capture the urgency and tone, aiding users in prioritizing
their responses.

Question-answering

The older generation of QA systems often worked by keyword matching, frequently missing
out on the contextual depth of user queries. LLMs, however, dive deep into context. They can
infer user intent, traverse vast information banks, and provide answers that are contextually
rich and precise. Some of the examples where this could be used include:

• Virtual assistants: LLMs can offer detailed explanations of a weather forecast

considering the user’s location, time of year, and recent weather trends.

• Customer support: In business platforms, LLM-based bots could provide answers that
take into account the user’s purchase history, past queries, and potential issues, offering
personalized assistance.

February 2025 73
Foundational Large Language Models & Text Generation

• Academic platforms: On academic platforms like Wolfram Alpha, LLMs could cater to
user queries by understanding the depth and context of academic questions, offering
answers that suit everyone from a high school student to a postgraduate researcher.

The quality of the generated answers, as well as the corresponding citations and sources
can be significantly improved by using advanced search systems (such as those based on
Retrieval Augmented Generation (RAG) architectures) to expand the prompt with relevant
information, as well as post-hoc grounding after the response has been generated. Clear
instructions, roles of what should and should not be used to answer the question, and
advanced prompt engineering approaches (such as chain of thought and search/RAG
architectures), combined with a lower temperature value amongst other things can also
help greatly.

Chatbots

Earlier chatbots followed scripted pathways, leading to ‘mechanical’ conversations. LLMs

transform this space by offering dynamic, human-like interactions. They can analyze
sentiment, context, and even humor, making digital conversations feel more authentic. Some
examples of where this can be used include:

• Customer service: A chatbot on retail platforms like Zara could not only answer product-
related queries but also offer fashion advice based on current trends.

• Entertainment: On Media LLM-driven chatbots could engage with users dynamically,

reacting to live events in the stream and moderating chats with contextual understanding.

February 2025 74
Foundational Large Language Models & Text Generation

Content generation

Text generation isn’t new, but what LLMs bring to the table is the unprecedented ability
to generate human-like text that’s contextually relevant and rich in detail. Earlier models
would often lose context or coherence over longer passages. LLMs, with their vast
knowledge and nuanced understanding, can craft text spanning various styles, tones, and
complexities, mixing factuality with creativity (depending on the context) effectively bridging
the gap between machine-generated and human-written content. The following are some
real-world examples:

• Content creation: Platforms could utilize LLMs to help marketers develop advertisements.
Instead of generic content, the LLMs could generate creative, targeted, and
audience-specific messages.

• Scriptwriting: LLMs could potentially assist with producing scripts for movies or TV
shows. Writers could input themes or plot points, and the model can suggest dialogues or
scene descriptions, enhancing the creative process.

Text generation is a wide task encompassing a variety of use cases that might range from
those where correctness of the generated output is more or less important than its creativity/
diversity of the language. The sampling methods and parameters like temperature should be
tuned accordingly. For more information, see the prompt engineering and architecting for
LLM applications whitepapers.

Natural language inference

Natural language inference (NLI) is the task of determining whether a given textual
hypothesis can be logically inferred from a textual premise.

February 2025 75
Foundational Large Language Models & Text Generation

Traditional models struggled with nuanced relationships or those that require a deeper
understanding of context. LLMs, with their intricate grasp of semantics and context, excel
at tasks like these, bringing accuracy levels close to human performance. The following are
some real-world examples:

• Sentiment analysis: Businesses could utilize LLMs to infer customer sentiment from
product reviews. Instead of just basic positive or negative tags, they could extract
nuanced emotions like ‘satisfaction,’ ‘disappointment,’ or ‘elation’.

• Legal document review: Law firms could employ LLMs to infer implications
and intentions in contracts, ensuring there are no contradictions or potentially
problematic clauses.

• Medical diagnoses: By analyzing patient descriptions and histories, LLMs could assist
doctors in inferring potential diagnoses or health risks, ensuring early intervention.

The whitepapers on domain specific LLMs, prompt engineering, and architecting for LLM
applications give further insight into these use cases.

Text classification

Text classification involves categorizing text into predefined groups. While traditional
algorithms were efficient, they often struggled with ambiguous or overlapping categories.
LLMs, given their deep understanding of context, can classify text with higher precision, even
when faced with subtle distinctions. Some examples of this include:

• Spam detection: Email services could utilize LLMs to classify emails as spam or
legitimate. Instead of just keyword-based detection, the models understand the context
and intent, potentially reducing false positives.

February 2025 76
Foundational Large Language Models & Text Generation

• News categorization: News platforms could employ LLMs to categorize articles into
topics like ‘technology,’ ‘politics,’ or ‘sports,’ even when articles blur the boundaries
between categories.

• Customer feedback sorting: Businesses could analyze customer feedback through

LLMs to categorize them into areas like ‘product design,’ ‘customer service,’ or ‘pricing,’
ensuring targeted responses.

• Evaluating LLMs as autorater: LLMs could be used to rate, compare and rank the
generated outputs of other LLMs as well.

Text analysis

LLMs excel at deep text analysis – extracting patterns, understanding themes, and gleaning
insights from vast textual datasets. Where traditional tools would scratch the surface, LLMs
delve deep, offering rich and actionable insights. Some potential real-world examples are:

• Market research: Companies could leverage LLMs to analyze consumer conversations on

social media, extracting trends, preferences, and emerging needs.

• Literary analysis: Academics could employ LLMs to understand themes, motifs, and
character developments in literary works, offering fresh perspectives on classic and
contemporary literature.

February 2025 77
Foundational Large Language Models & Text Generation

Multimodal applications

Multimodal LLMs, capable of processing and generating text, images, audio, and video, have
opened up a new frontier in AI, offering a range of exciting and innovative applications across
various sectors. The following are some examples:

Creative content generation:

• Storytelling: An AI system could watch an image or video and spin a captivating narrative,
integrating details from the visual with its knowledge base.

• Advertising and marketing: Generating targeted and emotionally resonant

advertisements based on product photos or videos.

Education and accessibility:

• Personalized learning: Tailoring educational materials to individual learning styles by

combining text with interactive visual and audio elements.

• Assistive technology: Multimodal LLMs could power tools that describe images, videos,
and audio for visually or hearing-impaired individuals.

Business and industry:

• Document understanding and summarization: Automatically extracting key information

from complex documents, combining text and visuals like invoices and contracts.

• Customer service: Multimodal chatbots can understand and respond to customer

queries combining text and images, offering a richer and more personalized experience.
Science and research:

February 2025 78
Foundational Large Language Models & Text Generation

• Medical diagnosis: Analyzing medical scans and reports together, identifying potential
issues and providing insights for doctors.

• Bioinformatics and drug discovery: Integrating knowledge from diverse data sources
like medical images, protein structures, and research papers to accelerate research.

These examples are just the tip of the iceberg. As research progresses, the applications
of multimodal LLMs are only expected to grow, transforming our daily lives in diverse and
profound ways. Multimodal LLMs also benefit greatly from the existing methodologies of
Unimodal LLMs ( i.e., text based LLMs).

LLMs, thanks to their ability to understand and process language, are reshaping how we
interact with, generate, and analyze text across diverse sectors. As they continue to evolve,
their applications will only grow, boosting the ability for machines and humans to have rich
natural language interactions.

February 2025 79
Foundational Large Language Models & Text Generation

Summary
In this whitepaper we have discussed the basics of transformers, upon which all modern-day
LLMs are based. We detailed the evolution of the various LLM model architectures and their
components. We’ve also seen the various methodologies you can use to train and fine-tune
models efficiently and effectively. We briefly discussed prompt engineering and sampling
techniques that greatly influence the output of an LLM, and also touched on possible
applications of this technology. There are a number of key takeaways to keep in mind:

• The transformer architecture is the basis for all modern-day LLMs. Across the various
architectures mentioned in this whitepaper we see that it’s important not only to add more
parameters to the model, but the composition of the dataset is equally important.

• The order and strategies used for fine-tuning is important and may include multiple steps
such as Instruction Tuning, Safety Tuning, etc. Supervised Fine Tuning (SFT) is important
in capturing the essence of a task. RLHF, and potentially RLAIF, can be used to shift the
distribution from the pretraining distribution to a more desired one through the power of
the reward function, that can reward desirable behaviors and penalize undesirable ones.

• Making inference from neural models efficient is an important problem and an active
field of research. Many methods exist to reduce serving costs and latency with minimal
impact to model performance, and some exact acceleration methods guarantee identical
model outputs.

• Large language models can be used for a variety of tasks including summarization,
translation, question answering, chat, code generation, and many more. You can
create your own tasks using the Vertex and Makersuite text generation services which
leverage Google’s latest language models. After the model has been trained and tuned,
it is important to experiment with engineering prompts. You should use the technique
most appropriate for the task-at-hand because LLMs can be sensitive to prompts k.

February 2025 80
Foundational Large Language Models & Text Generation

Furthermore, it is also possible to enhance task specific performance or creativity and

diversity by tweaking the parameters corresponding to sampling techniques such as
Top-K, Top-P, and Max decoding steps to find the optimum mix of correctness, diversity,
and creativity required for the task at hand.

February 2025 81
Foundational Large Language Models & Text Generation

Endnotes

1. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I., 2017, Attention is
all you need. Advances in Neural Information Processing Systems, 30.

2. Wikipedia, 2024, Word n-gram language model. Available at:

https://en.wikipedia.org/wiki/Word_n-gram_language_model.

3. Sutskever, I., Vinyals, O., & Le, Q. V., 2014, Sequence to sequence learning with neural networks. Advances in
Neural Information Processing Systems, 27.

4. Gu, A., Goel, K., & Ré, C., 2021, Efficiently modeling long sequences with structured state spaces.
arXiv preprint arXiv:2111.00396.

5. Jalammar, J. (n.d.). The illustrated transformer. Available at:

https://jalammar.github.io/illustrated-transformer/.

6. Ba, J. L., Kiros, J. R., & Hinton, G. E., 2016, Layer normalization.
arXiv preprint arXiv:1607.06450.

7. He, K., Zhang, X., Ren, S., & Sun, J., 2016, Deep residual learning for image recognition. Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition.

8. HuggingFace., 2024, Byte Pair Encoding. Available at:

https://huggingface.co/learn/nlp-course/chapter6/5?fw=pt.

9. Kudo, T., & Richardson, J., 2018, Sentencepiece: A simple and language independent subword tokenizer and
detokenizer for neural text processing. arXiv preprint arXiv:1808.06226.

10. HuggingFace, 2024, Unigram tokenization. Available at:

https://huggingface.co/learn/nlp-course/chapter6/7?fw=pt.

11. Goodfellow et. al., 2016, Deep Learning. MIT Press. Available at: http://www.deeplearningbook.org.

12. Radford, Alec et al., 2019, Language models are unsupervised multitask learners.

13. Brown, Tom, et al., 2020, Language models are few-shot learners. Advances in Neural Information
Processing Systems, 33, 1877-1901.

14. Devlin, Jacob, et al., 2018, BERT: Pre-training of deep bidirectional transformers for language understanding.
arXiv preprint arXiv:1810.04805.

February 2025 82
Foundational Large Language Models & Text Generation

15. Radford, A., & Narasimhan, K., 2018, Improving language understanding by generative pre-training.

16. Dai, A., & Le, Q., 2015, Semi-supervised sequence learning. Advances in Neural Information
Processing Systems.

17. Ouyang, Long, et al., 2022, Training language models to follow instructions with human feedback. Advances
in Neural Information Processing Systems, 35, 27730-27744.-27744.

18. OpenAI., 2023, GPT-3.5. Available at: https://platform.openai.com/docs/models/gpt-3-5.

19. OpenAI., 2023, GPT-4 Technical Report. Available at: https://arxiv.org/abs/2303.08774.

20. Thoppilan, Romal, et al., 2022, Lamda: Language models for dialog applications.
arXiv preprint arXiv:2201.08239.

21. Llama 3.2: Revolutionizing edge AI and vision with open, customizable models. Available
at: https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/.

22. Rae, J. W., Borgeaud, S., Cai, T., Millican, K., Hoffmann, J., Song, F., ... & Irving, G., 2021, Scaling language
models: Methods, analysis & insights from training Gopher. Available at: https://arxiv.org/pdf/2112.11446.pdf.

23. Du, N., He, H., Dai, Z., Mccarthy, J., Patwary, M. A., & Zhou, L., 2022, GLAM: Efficient scaling of language
models with mixture-of-experts. In International Conference on Machine Learning (pp. 2790-2800). PMLR.

24. Kaplan, J., McCandlish, S., Henighan, T., Brown, T. B., Chess, B., Child, R., ... & Amodei, D., 2020, Scaling laws
for neural language models. arXiv preprint arXiv:2001.08361.

25. Hoffmann, Jordan, et al., 2022, Training compute-optimal large language models. arXiv
preprint arXiv:2203.15556.

26. Shoeybi, Mohammad, et al., 2019, Megatron-LM: Training multi-billion parameter language models using
model parallelism. arXiv preprint arXiv:1909.08053.

27. Muennighoff, N. et al., 2023, Scaling data-constrained language models. arXiv preprint arXiv:2305.16264.

28. Chowdhery, Aakanksha, et al., 2023, Palm: Scaling language modeling with pathways. Journal of Machine
Learning Research, 24(240), 1-113.

29. Wang, Alex, et al.,2019, SuperGLUE: A stickier benchmark for general-purpose language understanding
systems. Advances in Neural Information Processing Systems, 32.

30. Anil, Rohan, et al., 2023, Palm 2 technical report. arXiv preprint arXiv:2305.10403.

February 2025 83
Foundational Large Language Models & Text Generation

31. DeepMind, 2023, Gemini: A family of highly capable multimodal models. Available at:
https://storage.googleapis.com/deepmind-media/gemini/gemini_1_report.pdf.

32. DeepMind, 2024, Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.
Available at: https://storage.googleapis.com/deepmind-media/gemini/gemini_v1_5_report.pdf.

33. Google Developers, 2024, Introducing PaLi-Gemma, Gemma 2, and an upgraded responsible AI toolkit.
Available at: https://developers.googleblog.com/en/gemma-family-and-toolkit-expansion-io-2024/.

34. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M., Lacroix, T., ... & Jegou, H., 2023, Llama 2: Open
foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.

35. Jiang, A. Q., 2024, Mixtral of experts. arXiv preprint arXiv:2401.04088.

36. Qwen, 2024, Introducing Qwen1.5. Available at: https://qwenlm.github.io/blog/qwen1.5/.

37. Young, A., 2024, Yi: Open foundation models by 01.AI. arXiv preprint arXiv:2403.04652.

38. Grok-1, 2024, Available at: https://github.com/xai-org/grok-1.

39. Duan, Haodong, et al., 2023, BotChat: Evaluating LLMs’ capabilities of having multi-turn dialogues.
arXiv preprint arXiv:2310.13650.

40. Google Cloud, 2024, Tune text models with reinforcement learning from human feedback. Available at:
https://cloud.google.com/vertex-ai/generative-ai/docs/models/tune-text-models-rlhf.

41. Bai, Yuntao, et al., 2022, Constitutional AI: Harmlessness from AI feedback. arXiv preprint arXiv:2212.08073.

42. Wikipedia, 2024, Likert scale. Available at: https://en.wikipedia.org/wiki/Likert_scale.

43. Sutton, R. S., & Barto, A. G., 2018, Reinforcement learning: An introduction. MIT Press.

44. Bai, Yuntao, et al, 2022, Constitutional AI: Harmlessness from AI feedback. arXiv preprint arXiv:2212.08073.

45. Rafailov, Rafael, et al., 2023, Direct preference optimization: Your language model is secretly a reward
model. arXiv preprint arXiv:2305.18290.

46. Houlsby, Neil, et al., 2019, Parameter-efficient transfer learning for NLP. In International Conference on
Machine Learning (pp. 2790-2799). PMLR.

47. Hu, Edward J., et al., 2021, LoRA: Low-rank adaptation of large language models.
arXiv preprint arXiv:2106.09685.

48. Dettmers, Tim, et al., 2023, QLoRA: Efficient finetuning of quantized LLMs. arXiv preprint arXiv:2305.14314.

February 2025 84
Foundational Large Language Models & Text Generation

49. Lester, B., Al-Rfou, R., & Constant, N., 2021, The power of scale for parameter-efficient prompt tuning. arXiv
preprint arXiv:2104.08691.

50. HuggingFace., 2020, How to generate text? Available at: https://huggingface.co/blog/how-to-generate.

51. Google AI Studio Context caching. Available

at: https://ai.google.dev/gemini-api/docs/caching?lang=python.

52. Vertex AI Context caching overview. Available

at: https://cloud.google.com/vertex-ai/generative-ai/docs/context-cache/context-cache-overview.

53. Gu, A., Goel, K., & Ré, C., 2021, Efficiently modeling long sequences with structured state spaces.
Available at: https://arxiv.org/abs/2111.00396.

54. Hubara et al., 2016, Quantized neural networks: Training neural networks with low precision weights and
activations. Available at: https://arxiv.org/abs/1609.07061.

55. Benoit Jacob et al., 2017, Quantization and training of neural networks for efficient integer-arithmetic-only
inference. Available at: https://arxiv.org/abs/1712.05877.

56. Bucila, C., Caruana, R., & Niculescu-Mizil, A., 2006, Model compression. Knowledge Discovery and Data
Mining. Available at: https://www.cs.cornell.edu/~caruana/compression.kdd06.pdf.

57. Hinton, G., Vinyals, O., & Dean, J., 2015, Distilling the knowledge in a neural network.
Available at: https://arxiv.org/abs/1503.02531.

58. Zhang, L., Fei, W., Wu w., He Y., Lou Z., Zhou H., 2023, Dual Grained Quantisation: Efficient Finegrained
Quantisation for LLM. Available at: https://arxiv.org/abs/2310.04836.

59. Agarwal, R., Vieillard, N., Zhou, Y., Stanczyk, P., Ramos, S., Geist, M., Bachem, O., 2024, On-
Policy Distillation of Language Models: Learning from Self-Generated Mistakes. Available
at: https://arxiv.org/abs/2306.13649.

60. Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., & Dean, J., 2017, Outrageously large neural
networks: The sparsely-gated mixture-of-experts layer. Available at: https://arxiv.org/abs/1701.06538.

61. Schuster, T., Fried, D., & Jurafsky, D., 2022, Confident adaptive language modeling. Available at:
https://arxiv.org/abs/2207.07061.

62. Tri Dao et al. “FlashAttention. Available at:

https://arxiv.org/abs/2205.14135.

February 2025 85
Foundational Large Language Models & Text Generation

63. Leviathan, Y., Ram, O., Desbordes, T., & Haussmann, E., 2022, Fast inference from transformers via
speculative decoding. Available at: https://arxiv.org/abs/2211.17192.

64. Li, Y., Humphreys, P., Sun, T., Carr, A., Cass, S., Hawkins, P., ... & Bortolussi, L., 2022, Competition-level code
generation with AlphaCode. Science, 378(1092-1097). DOI: 10.1126/science.abq1158.

65. Romera-Paredes, B., Barekatain, M., Novikov, A., Novikov, A., Rashed, S., & Yang, J., 2023, Mathematical
discoveries from program search with large language models. Nature. DOI: 10.1038/s41586-023-06924-6.

66. Wikipedia., 2024, Cap set. Available at: https://en.wikipedia.org/wiki/Cap_set.

67. Trinh, T. H., Wu, Y., & Le, Q. V. et al., 2024, Solving olympiad geometry without human demonstrations. Nature,
625, 476–482. DOI: 10.1038/s41586-023-06747-5.

68. Mikolov, T., Chen, K., Corrado, G., Dean, J., 2013, Efficient Estimation of Word Representations in Vector
Space. Available at: https://arxiv.org/pdf/1301.3781.

69. Shi, L., Ma, C., Liang, W., Ma, W., Vosoughi, S., 2024, Judging the Judges: A Systematic Study of Position Bias
in LLM-as-a-Judge. Available at: https://arxiv.org/abs/2406.07791

70. Pandit, B., 2024, What Is Mixture of Experts (MoE)? How It Works, Use Cases & More. Available
at: https://www.datacamp.com/blog/mixture-of-experts-moe

February 2025 86