Large Language Models Survey: Complete Guide to GPT-4, LLaMA and Modern LLM Architecture

The LLM Landscape: From GPT-3 to Frontier Models

The field of artificial intelligence has been fundamentally transformed by large language models—massive neural networks trained on vast text corpora that exhibit capabilities far beyond what their predecessors could achieve. A comprehensive survey published in early 2024 mapped this rapidly evolving landscape, documenting over 50 significant models and the techniques that make them work. This analysis of that landmark paper reveals the architectural innovations, training methodologies and alignment strategies that define the current state of the art in language AI.

Large language models are distinguished from their smaller predecessors by more than just parameter count. When models reach sufficient scale—typically tens to hundreds of billions of parameters—they begin exhibiting emergent capabilities that were not explicitly trained. These include in-context learning, where the model adapts its behavior based on examples provided in the prompt, instruction following after targeted fine-tuning, and multi-step reasoning through techniques like chain-of-thought prompting. These emergent abilities represent a qualitative shift in what language technology can accomplish.

The survey identifies two fundamental trends driving LLM progress. The first is the well-documented relationship between model scale, training data volume and emergent abilities—captured formally by neural scaling laws. The second, less obvious trend is that improvements in data quality, training procedures and alignment techniques can produce gains that rival or exceed those from scaling alone. This insight has profound implications for organizations seeking to deploy AI capabilities, as it suggests that brute-force scaling is not the only path to powerful models.

The competitive landscape has split into two ecosystems that feed off each other productively. Closed-source models like GPT-4 push the frontier of raw capabilities, while open-source releases like LLaMA and Mistral enable widespread experimentation, fine-tuning and research. This dynamic has accelerated progress dramatically, creating a virtuous cycle where innovations in one ecosystem quickly influence the other. For enterprises evaluating AI technology assessments, understanding both ecosystems is critical for strategic decision-making.

Architectural Foundations: Decoder, Encoder and Hybrid Models

At the heart of every LLM lies the Transformer architecture, introduced in 2017, which uses self-attention mechanisms to process text in parallel rather than sequentially. The survey categorizes modern LLMs into three architectural families, each optimized for different use cases and with distinct strengths that determine their practical applications.

Decoder-only models represent the dominant paradigm for text generation. These models, exemplified by the GPT family and LLaMA, use causal attention masking to predict each next token based only on preceding tokens. This autoregressive approach makes them natural choices for open-ended generation, conversation and creative tasks. Nearly all state-of-the-art LLMs use this architecture because it scales effectively and produces fluent, coherent text across diverse domains.

Encoder-only models like BERT and its descendants (RoBERTa, DeBERTa, ELECTRA) use bidirectional attention to build rich contextual representations of input text. Originally trained with masked language modeling objectives, these models excel at understanding tasks—classification, question answering, named entity recognition and semantic similarity. While they have been somewhat overshadowed by decoder-only models in the public imagination, encoder-based models remain workhorses in production NLP systems where understanding accuracy matters more than generation fluency.

Encoder-decoder models like T5 and BART combine both approaches, processing input text through a bidirectional encoder and generating output through an autoregressive decoder. This architecture naturally supports sequence-to-sequence tasks including translation, summarization and instruction following. Google’s Flan-T5 demonstrated that instruction tuning an encoder-decoder model on diverse tasks produced powerful zero-shot generalization.

Beyond these three families, the survey documents emerging hybrid architectures that challenge the pure Transformer paradigm. Mixture-of-Experts models like GLaM (1.2 trillion parameters) use sparse routing to activate only a subset of parameters for each input, scaling capacity without proportionally increasing compute costs. Structured state space models and RWKV offer RNN-like inference efficiency while maintaining Transformer-level training parallelism, potentially solving the quadratic attention cost that limits context length in standard Transformers.

The GPT Family: Evolution from GPT-1 to GPT-4

OpenAI’s GPT family represents the most commercially impactful lineage of language models, and the survey traces its evolution in illuminating detail. GPT-3, launched in 2020 with 175 billion parameters trained on approximately 300 billion tokens, demonstrated for the first time that a sufficiently large language model could perform useful tasks purely through in-context learning—without any task-specific training. This capability, while imperfect, suggested that scaling alone could unlock increasingly general intelligence.

The transition from GPT-3 to GPT-3.5 and ChatGPT introduced a paradigm shift in how language models interact with users. Rather than simply scaling the model, OpenAI applied instruction tuning and RLHF to create InstructGPT, producing a model that followed instructions more reliably, generated more helpful responses and exhibited fewer harmful behaviors. ChatGPT, released in November 2022, brought these capabilities to the public and triggered the AI revolution that continues to reshape industries.

GPT-4, the most capable model in the family, is estimated by the survey at approximately 1.76 trillion parameters trained on roughly 13 trillion tokens—though OpenAI has not officially confirmed these specifications. GPT-4 demonstrated substantial improvements across virtually every benchmark, with particular strength in reasoning, code generation and multi-modal understanding. The model’s capabilities on professional exams—passing the bar exam, medical licensing exam and other tests at or above human performance—illustrated that LLM capabilities had reached genuinely practical levels for knowledge work.

The Codex branch of the GPT family, optimized for code generation, demonstrated that domain-specific fine-tuning could produce dramatic improvements in specialized tasks. Codex and its successor GitHub Copilot have become the most commercially successful AI developer tools, establishing code generation as one of the clearest value propositions for enterprise AI deployment.

Open-Source Revolution: LLaMA, Mistral and the Democratization of AI

Meta’s release of LLaMA in February 2023 catalyzed what may be the most significant development in the LLM ecosystem: the democratization of large-scale language AI. LLaMA 1, available in 7B, 13B, 33B and 65B parameter variants trained on 1-1.4 trillion tokens of public data, proved that open-weight models could approach the performance of much larger closed models. This revelation triggered an explosion of community-driven innovation.

Within weeks of LLaMA’s release, the research community produced an astonishing array of fine-tuned derivatives. Alpaca demonstrated that instruction-tuning a 7B model on just 52,000 examples could produce GPT-3.5-like performance at a fraction of the cost. Vicuna showed that fine-tuning on high-quality conversation data from ShareGPT could create a chatbot that impressed human evaluators. Guanaco, Koala, LongLLaMA and dozens of other variants explored different specializations and efficiency techniques, collectively demonstrating that the value in LLMs increasingly lies in data and training methodology rather than raw model size.

LLaMA 2, released in July 2023 with 7B, 13B and 70B parameter versions trained on 2 trillion tokens, strengthened this ecosystem further. Meta released both foundation and chat-optimized versions, providing the community with production-ready models suitable for commercial deployment. The 70B chat model achieved competitive performance on public benchmarks, challenging the notion that only trillion-parameter closed models could deliver high-quality AI interactions.

Mistral emerged as another pivotal player in the open-source ecosystem, with its 7B model achieving remarkable performance-per-parameter efficiency through architectural innovations including sliding window attention and grouped-query attention. These efficiency techniques reduced inference costs while maintaining quality, making Mistral particularly attractive for deployment scenarios where compute cost matters. The broader impact of open-source LLMs extends beyond model availability: they enable academic research, independent safety auditing and custom deployments that closed models cannot support.

Google’s PaLM and the Instruction Tuning Breakthrough

Google’s PaLM (Pathways Language Model) family represents a parallel track of LLM development that produced several seminal insights. PaLM, at 540 billion parameters trained on 780 billion tokens, demonstrated breakthrough performance on reasoning tasks, particularly on chain-of-thought prompting benchmarks. The model’s training on a diverse mixture of web documents, books, code and conversation data contributed to broad capabilities across domains.

The most impactful contribution from the PaLM ecosystem was Flan-PaLM, which demonstrated the power of massive-scale instruction tuning. By fine-tuning PaLM on 1,836 tasks drawn from 473 datasets, Google achieved an average improvement of 9.4% over the base model across benchmarks. This result was remarkable because it showed that the same compute budget invested in instruction tuning produced gains comparable to doubling the model size—a finding with enormous practical implications for organizations that cannot afford to train the largest models.

The medical domain provided a compelling case study for domain-specific adaptation of PaLM. Med-PaLM 2 achieved 86.5% on MedQA, a 19% improvement over its predecessor, approaching expert human physician performance on medical question answering. This demonstrated that combining large-scale pretraining with domain expertise through fine-tuning could create AI systems with genuine professional-grade capabilities in specialized fields. The implications for healthcare AI applications are profound.

PaLM-2, the successor model, introduced improved compute efficiency and multilingual capabilities. Trained with a mixture of objectives rather than pure autoregressive pretraining, PaLM-2 demonstrated that training methodology innovations could compensate for smaller model sizes, achieving competitive results with more efficient resource utilization. This trend toward training efficiency—getting more capability per compute dollar—represents one of the most important ongoing themes in LLM development.

Training Data: Quality, Curation and the Chinchilla Scaling Laws

The survey reveals that training data quality and composition may be the most underappreciated factor in LLM performance. While public attention has focused on model architecture and parameter counts, the research demonstrates repeatedly that data decisions drive outcomes at least as much as architectural choices. The Chinchilla paper, produced by DeepMind, fundamentally reshaped how the field thinks about training efficiency.

Chinchilla, with 70 billion parameters trained on 1.4 trillion tokens, outperformed the much larger Gopher (280B parameters) by simply training a smaller model on more data. This compute-optimal scaling law showed that for any given compute budget, model size and training tokens should be scaled together in approximately equal proportions. The practical implication was revolutionary: many existing models were undertrained relative to their size, and the field had been wasting compute by building models too large for their training data.

Data quality research has produced equally important insights. The RefinedWeb dataset, used to train the Falcon models, demonstrated that aggressive filtering and deduplication of CommonCrawl web data could produce models competitive with those trained on carefully curated mixtures of books, academic papers and code. Specifically, the research showed that refined and deduplicated web-only data, combined with careful filtering, can produce better models than mixing unfiltered sources—challenging the conventional wisdom that diverse data sources are always better.

The survey documents several critical data processing techniques that materially impact model quality. Deduplication removes repeated documents that cause memorization and reduce generalization. Toxicity filtering removes harmful content that the model would otherwise reproduce. Domain filtering balances the mixture of web text, books, code and academic content to produce well-rounded capabilities. The ROOTS corpus used for BLOOM and other curated multilingual datasets demonstrate that these techniques apply equally to multilingual training scenarios.

Alignment Techniques: RLHF, DPO and Making Models Safe

Perhaps the most consequential section of the survey covers alignment—the techniques used to make language models helpful, honest and harmless. Without alignment, even the most capable language model may generate toxic content, follow harmful instructions or produce outputs that conflict with human values. The evolution of alignment techniques from RLHF to newer methods represents one of the fastest-moving areas of LLM research.

Reinforcement Learning from Human Feedback (RLHF) remains the foundational alignment technique, employed by InstructGPT, ChatGPT, WebGPT and other models. The process involves three stages: collecting human preference comparisons between model outputs, training a reward model to predict human preferences, and fine-tuning the language model using reinforcement learning (typically PPO) to maximize the learned reward signal. While effective, RLHF is complex, expensive and somewhat unstable, motivating the search for simpler alternatives.

Direct Preference Optimization (DPO) emerged as a breakthrough simplification of RLHF. Rather than training a separate reward model and applying RL, DPO directly optimizes the language model policy to satisfy human preferences through a classification-like objective. The approach eliminates the reward model and RL sampling stages entirely, producing more stable training that is computationally cheaper while achieving comparable or superior alignment quality. DPO has been rapidly adopted by both research labs and companies fine-tuning open-source models.

KTO (Kahneman-Tversky Optimization) simplifies the data requirements further by using examples labeled simply as desirable or undesirable, rather than requiring paired preference comparisons. This makes data collection substantially easier and cheaper in many practical settings. Drawing inspiration from behavioral economics and prospect theory, KTO accounts for the human tendency to weigh losses more heavily than gains, producing aligned models from simpler feedback signals.

The survey also documents RLAIF (Reinforcement Learning from AI Feedback), where an aligned model serves as a surrogate annotator to scale feedback beyond what human annotation can provide. This technique addresses one of the fundamental bottlenecks in alignment: the cost and speed of human evaluation. While RLAIF introduces questions about circular reasoning and the potential for amplifying biases present in the teacher model, it has proven effective in practice for extending the reach of human-sourced alignment data.

Efficient Fine-Tuning: LoRA, QLoRA and Accessible AI

The survey’s coverage of efficient fine-tuning techniques reveals how the barriers to customizing large language models have fallen dramatically, enabling organizations of all sizes to adapt frontier models to their specific needs. These techniques transform LLMs from monolithic systems requiring massive compute into flexible tools that can be specialized on modest hardware.

LoRA (Low-Rank Adaptation) achieves parameter-efficient fine-tuning by freezing the base model weights and learning small low-rank update matrices that are applied to specific layers. Instead of updating all model parameters during fine-tuning—which for a 70B model would require enormous GPU memory—LoRA learns only the update matrices, typically reducing trainable parameters by 99% or more. The technique produces models that perform comparably to full fine-tuning while being dramatically cheaper to train and store.

QLoRA extends this approach by combining LoRA with 4-bit quantization of the base model, enabling fine-tuning of a 65-billion parameter model on a single 48GB GPU. This capability was transformative because it brought large-model customization within reach of academic researchers, startups and small teams who could not afford multi-GPU clusters. The Guanaco model, trained using QLoRA, achieved performance competitive with commercial models while requiring only a single consumer GPU for fine-tuning.

Beyond LoRA, the survey documents knowledge distillation as a complementary efficiency technique. Distillation trains a smaller “student” model to mimic the outputs of a larger “teacher” model, producing compact models that retain much of the teacher’s capability. API distillation, where the student learns from the teacher’s outputs rather than its internal representations, has been particularly popular because it can be applied even when the teacher model’s weights are not available—as is the case with closed-source models like GPT-4.

Quantization techniques beyond QLoRA further reduce the cost of deploying LLMs in production. By reducing model precision from 16-bit or 32-bit floating point to 8-bit, 4-bit or even more aggressive quantization levels, organizations can serve models faster and cheaper without catastrophic quality degradation. The combination of quantization, LoRA fine-tuning and distillation creates a practical toolkit for enterprises seeking to deploy customized AI capabilities at manageable cost.

Emergent Capabilities and the Future of Language Models

The survey’s documentation of emergent capabilities paints a picture of systems that continue to surprise even their creators. In-context learning—the ability to adapt behavior based on examples in the prompt—was not explicitly trained but appeared naturally in sufficiently large models. Chain-of-thought reasoning, where models solve complex problems by articulating intermediate steps, emerged similarly and has been refined through techniques like self-consistency and tree-of-thought prompting.

Tool use and agentic behavior represent perhaps the most practically significant emergent capabilities. Models like ChatGPT with plugins, WebGPT with browser access, and various implementations of LLM agents demonstrate that language models can coordinate with external tools, databases and APIs to accomplish complex multi-step tasks. This capability transforms LLMs from passive text generators into active agents that can retrieve information, execute calculations, interact with software systems and plan sequences of actions.

The survey identifies retrieval augmentation as a critical technique for enhancing LLM capabilities without increasing model size. RETRO and similar retrieval-enhanced architectures condition generation on relevant chunks retrieved from external databases, enabling models to access far more information than could be stored in their parameters alone. This approach addresses the well-known limitation of LLM knowledge being frozen at training time, allowing models to access current information and domain-specific knowledge bases.

Looking forward, the survey points to several promising research directions. Non-Transformer architectures like structured state space models could solve the quadratic attention cost that currently limits context length. Multimodal models that process text, images, audio and video within a unified architecture are rapidly improving. And the combination of language models with formal reasoning systems—hybrid neuro-symbolic approaches—may eventually produce AI systems that combine the fluency of neural networks with the reliability of symbolic reasoning.

Practical Implications for Enterprise AI Adoption

For organizations evaluating AI adoption strategies, the survey’s findings provide a clear framework for decision-making. The most important practical insight is that the choice between closed-source frontier models and open-source alternatives is not binary—most organizations will benefit from a portfolio approach that leverages each ecosystem’s strengths.

Closed-source models like GPT-4 offer the highest raw capabilities and require no infrastructure investment, making them ideal for use cases where quality matters most and data sensitivity is manageable. Open-source models like LLaMA offer customization, deployment flexibility, data privacy and potentially lower costs at scale, making them attractive for high-volume production use cases and applications involving sensitive data that cannot leave organizational boundaries.

The efficiency techniques documented in the survey—LoRA, QLoRA, distillation and quantization—mean that the barrier to customizing models has fallen dramatically. Organizations no longer need massive compute budgets to create specialized AI systems. A domain expert with a curated dataset and a single GPU can produce a fine-tuned model that outperforms general-purpose frontier models on specific tasks. This democratization creates opportunities for competitive differentiation through AI specialization rather than just AI adoption.

The alignment and safety considerations raised by the survey are equally relevant for enterprise deployment. Organizations deploying LLMs need to understand the alignment techniques applied to their models, implement appropriate guardrails and safety measures, and maintain human oversight of AI-generated outputs. The survey’s documentation of RLHF, DPO and related techniques provides the conceptual framework for evaluating vendor claims about model safety and understanding the inherent limitations of current alignment approaches. As noted by the National Institute of Standards and Technology, robust AI governance frameworks are essential for responsible enterprise deployment.