LARGE LANGUAGE MODELS

Language as the Foundation of General-Purpose Artificial Intelligence

The pursuit of intelligent language processing has occupied researchers since the earliest years of Artificial Intelligence. Human language represents one of the most sophisticated manifestations of cognition, integrating perception, memory, reasoning, abstraction and social understanding within a highly flexible symbolic system. Replicating even a fraction of this capability has therefore long been regarded as one of the defining challenges of computational intelligence. Early research concentrated upon symbolic approaches in which linguistic knowledge was represented explicitly through handcrafted grammatical rules, semantic networks and logical inference systems. Although these methods demonstrated important theoretical insights, they proved difficult to scale because the complexity and ambiguity of natural language continually exceeded the capacity of manually constructed representations.

The emergence of statistical learning during the final decades of the twentieth century introduced an alternative perspective. Rather than attempting to encode linguistic knowledge directly, researchers increasingly allowed computational systems to discover patterns from large collections of textual information. Machine learning algorithms progressively replaced manually engineered rules, enabling systems to estimate probabilistic relationships between words, phrases and grammatical structures. This transition significantly improved practical performance whilst establishing the intellectual foundations upon which modern language modelling would later develop.

Deep learning transformed this trajectory still further. Artificial neural networks capable of learning complex hierarchical representations demonstrated remarkable success across vision, speech and language processing. Recurrent neural networks and later attention-based transformer architectures enabled computational systems to model increasingly sophisticated contextual relationships extending across long sequences of language. These developments culminated in the emergence of Large Language Models trained upon enormous textual corpora comprising books, scientific publications, technical documentation, websites and numerous other forms of human communication. The resulting systems demonstrated unprecedented flexibility, capable of performing diverse intellectual tasks without requiring extensive task-specific engineering.

The emergence of Large Language Models therefore represents more than a technological improvement. It reflects a profound conceptual shift within Artificial Intelligence itself. Rather than constructing specialised systems for individual applications, researchers increasingly develop general-purpose language models capable of adapting to multiple domains through interaction and contextual understanding. Language becomes the universal interface through which knowledge is acquired, represented and applied, allowing Artificial Intelligence to participate in activities extending far beyond conventional natural language processing. Large Language Models consequently occupy a central position within contemporary Artificial Intelligence because they provide the intellectual infrastructure supporting numerous subsequent developments, including reasoning models, action models, multimodal systems and autonomous intelligent agents.

Defining Distributed Knowledge and Probabilistic Language Generation

Large Language Models are Artificial Intelligence systems designed to understand, generate and manipulate human language through statistical learning applied to extensive collections of textual information. Their defining characteristic is the capacity to construct high-dimensional internal representations of linguistic relationships that capture semantic meaning, grammatical structure and contextual dependency across an enormous diversity of written material. Rather than storing explicit definitions or manually programmed linguistic rules, these models acquire distributed representations that enable them to generalise flexibly across unfamiliar situations whilst maintaining coherent communication.

The designation "large" refers not solely to computational scale but also to the breadth of knowledge encoded within the model. Contemporary Large Language Models frequently contain billions of parameters whose values are adjusted during training to represent subtle statistical regularities present throughout human language. These parameters do not correspond directly to individual facts or concepts but collectively encode extraordinarily rich representations enabling the model to predict appropriate linguistic continuations across diverse contexts. Knowledge therefore emerges as a distributed property of the model rather than existing as isolated symbolic entries.

Language prediction remains the fundamental computational objective underlying these systems. During training the model repeatedly receives sequences of words and learns to estimate the most probable continuation according to patterns observed throughout the training corpus. Although apparently simple, this objective encourages the development of increasingly sophisticated internal representations because accurate prediction requires understanding syntax, semantics, discourse structure, factual relationships and numerous forms of implicit human knowledge. Consequently, prediction becomes the mechanism through which broader linguistic competence emerges.

Importantly, Large Language Models do not retrieve fixed answers from predefined databases in the manner of conventional information systems. Instead, each response is generated dynamically according to the interaction between the user's prompt, the model's learned representations and the probabilistic relationships acquired during training. Every output therefore reflects the continual synthesis of contextual information rather than simple retrieval. This characteristic explains both the remarkable flexibility of Large Language Models and the uncertainty that occasionally accompanies their responses.

From Symbolic Linguistics to Transformer Models

The historical development of Large Language Models reflects several decades of research into computational linguistics and machine learning. Early natural language processing systems relied heavily upon symbolic approaches in which linguistic knowledge was represented through handcrafted grammars, dictionaries and logical rules. Although theoretically elegant, these methods required continual manual maintenance and struggled to accommodate the ambiguity, variability and contextual richness characteristic of human language.

Statistical language modelling introduced a more adaptive alternative by estimating probabilistic relationships between words from observed textual data. N-gram models represented an important early achievement, demonstrating that surprisingly effective language processing could emerge from relatively simple statistical methods. Nevertheless, these approaches remained constrained because they considered only limited contextual information and therefore struggled to capture long-range linguistic dependencies.

Neural language models overcame many of these limitations by representing words within continuous mathematical spaces known as embeddings. Rather than treating each word as an isolated symbolic unit, embedding techniques positioned semantically related terms close together within high-dimensional vector spaces, enabling computational systems to capture subtle conceptual relationships unavailable to traditional statistical methods. Subsequent advances in recurrent neural networks extended contextual understanding further by maintaining internal memory across sequences of language.

The decisive breakthrough occurred with the introduction of transformer architectures, whose attention mechanisms enabled efficient modelling of long-range contextual relationships whilst supporting parallel computation across entire sequences of text. Unlike recurrent architectures that processed language sequentially, transformers evaluated relationships between every component of an input simultaneously, dramatically improving both computational efficiency and representational capability. This innovation enabled researchers to train substantially larger models upon vastly expanded textual corpora, giving rise to the first generation of Large Language Models possessing genuinely general-purpose linguistic competence.

The progression from symbolic systems to transformer architectures illustrates a broader intellectual transition within Artificial Intelligence. Knowledge increasingly became an emergent property of statistical learning rather than an explicit product of manual engineering. Language processing consequently evolved from narrow task-specific systems towards general computational frameworks capable of supporting numerous intellectual activities through unified representations of language itself.

Representation Learning, Semantics, Generalisation and Scale

The remarkable capabilities exhibited by Large Language Models arise from the interaction of several complementary theoretical principles rather than from any single technological innovation. At the heart of these systems lies the concept of representation learning, whereby Artificial Intelligence constructs internal mathematical representations directly from observed data rather than relying upon manually specified features. These representations capture increasingly abstract linguistic patterns across successive computational layers, enabling the model to distinguish subtle semantic relationships, contextual dependencies and conceptual structures.

Distributional semantics provides an important theoretical foundation for this process. Linguistic meaning is inferred from patterns of usage rather than explicit symbolic definitions. Words appearing within similar contexts gradually acquire related internal representations because the statistical structure of language reflects broader conceptual relationships present within human communication. Consequently, the model develops semantic understanding through observation of language itself rather than through direct instruction concerning meaning.

Generalisation represents another fundamental principle. Effective intelligence requires the ability to apply previously acquired knowledge to unfamiliar situations rather than merely reproducing observed examples. Large Language Models achieve this through distributed representations that encode abstract regularities extending beyond individual documents or sentences. During interaction the model continually combines these representations with contextual information provided by the user's prompt, enabling responses to novel questions for which no identical training example exists.

Scale also contributes significantly to model capability. Increasing parameter counts, larger training datasets and greater computational resources permit increasingly sophisticated representations to emerge throughout training. However, scale alone does not explain the success of Large Language Models. Architectural innovation, optimisation techniques, training methodology and data quality each contribute substantially to overall performance. Contemporary research therefore increasingly recognises that capability depends upon the interaction of multiple design principles rather than computational magnitude alone.

These intellectual foundations collectively transform language into a universal representational medium through which knowledge may be encoded, organised and applied. Large Language Models therefore function not merely as sophisticated text generators but as probabilistic knowledge representation systems capable of supporting reasoning, communication and problem solving across an exceptionally broad range of domains.

Attention, Hierarchical Representation and Parallel Computation

The transformer architecture constitutes the technological foundation upon which virtually all contemporary Large Language Models are constructed. Introduced to overcome limitations associated with sequential neural networks, transformers employ attention mechanisms that enable every element within a sequence of language to interact directly with every other element regardless of positional distance. This capability fundamentally changes the manner in which contextual information is represented because long-range dependencies may be modelled with far greater efficiency than previously possible.

Attention mechanisms operate by assigning varying degrees of importance to different components of an input sequence according to their relevance within the current context. Rather than processing language through rigid sequential progression, the model dynamically determines which words, phrases or concepts contribute most significantly to interpreting each element of the text. Multiple attention heads examine these relationships simultaneously from different analytical perspectives, enabling rich contextual representations to emerge throughout successive computational layers.

The hierarchical organisation of transformer architectures allows increasingly abstract representations to develop as information progresses through the network. Initial layers identify relatively simple linguistic relationships, while deeper layers capture progressively more sophisticated semantic, conceptual and contextual structures. This hierarchical abstraction contributes substantially to the versatility of Large Language Models because knowledge acquired within one domain may be applied flexibly across many others.

The efficiency of transformer computation has also proved decisive. Parallel processing enables models containing billions of parameters to be trained upon enormous textual corpora within practical timescales. Without this computational efficiency the emergence of contemporary Large Language Models would almost certainly have remained impossible. Transformer architectures therefore represent not merely an incremental improvement but the foundational technological innovation underpinning the present generation of Artificial Intelligence language systems.

Large-Scale Training, Tokenisation and Distributed Representation

The remarkable capabilities exhibited by Large Language Models are the consequence of extensive training processes through which Artificial Intelligence progressively acquires increasingly sophisticated representations of human language. Training involves exposing the model to vast collections of textual information drawn from books, academic literature, technical documentation, software repositories, newspapers and numerous other forms of written communication. Rather than memorising these documents directly, the model repeatedly predicts missing or subsequent linguistic elements, gradually refining billions of internal parameters until they encode highly complex statistical relationships governing language itself. The resulting representations capture syntax, semantics, contextual dependency and many aspects of general knowledge distributed throughout the training corpus.

Tokenisation represents the first stage of this learning process. Human language cannot be processed directly by computational systems and must therefore be transformed into numerical representations. Instead of treating complete words as indivisible units, contemporary Large Language Models frequently divide language into smaller components known as tokens. Individual tokens may correspond to complete words, prefixes, suffixes or common sequences of characters depending upon their frequency within the training corpus. This flexible representation enables the model to process unfamiliar vocabulary efficiently whilst maintaining manageable computational complexity across numerous languages and specialised domains.

Embeddings and Hierarchical Representation Learning

Once tokenised, language is converted into high-dimensional mathematical representations through embedding layers. Each token acquires a numerical position within a multidimensional vector space where linguistic relationships become encoded geometrically. Words possessing similar meanings or occurring within comparable contexts gradually occupy neighbouring regions of this space, enabling the model to capture conceptual similarity without relying upon manually defined semantic rules. These distributed representations evolve continually throughout training, becoming increasingly refined as the model encounters additional examples of language usage.

Representation learning constitutes the intellectual foundation of this process. Earlier approaches to natural language processing depended heavily upon manually engineered linguistic features designed by human experts. Large Language Models instead discover relevant representations automatically through optimisation. Successive computational layers identify progressively more abstract linguistic structures, beginning with local grammatical relationships before advancing towards discourse organisation, conceptual association and contextual reasoning. Knowledge therefore emerges through hierarchical abstraction rather than explicit symbolic programming.

Training requires extraordinary computational resources because every parameter within the model is repeatedly adjusted according to prediction error across billions of examples. Sophisticated optimisation algorithms gradually minimise these errors, enabling the model to construct increasingly accurate internal representations of language. Importantly, this process does not create explicit databases of factual information but rather probabilistic networks of conceptual relationships capable of supporting flexible reasoning and generation. The resulting model therefore possesses the capacity to interpret unfamiliar inputs by combining previously acquired representations in novel ways rather than retrieving predefined responses.

Emergence, Generalisation, In-Context Learning and Abstraction

One of the most remarkable characteristics of Large Language Models is the emergence of capabilities that were not explicitly programmed during development. As model scale, training data and computational resources increase, new forms of behaviour frequently appear that cannot be explained solely through incremental improvements in language prediction. These emergent capabilities include sophisticated reasoning, summarisation, translation, question answering, software development, mathematical problem solving and conceptual abstraction across multiple domains. Their appearance suggests that sufficiently rich statistical representations may give rise to unexpectedly broad intellectual competence.

Generalisation plays a central role in these capabilities. Effective Artificial Intelligence must apply previously acquired knowledge to unfamiliar circumstances rather than merely reproducing examples encountered during training. Large Language Models achieve this through distributed internal representations that encode abstract conceptual relationships extending across numerous linguistic contexts. Consequently, when presented with novel problems, the model synthesises relevant patterns acquired from many different sources rather than searching for identical examples within its training data.

In-context learning represents another important emergent property. Unlike earlier machine learning systems that generally required retraining before adapting to new tasks, Large Language Models frequently modify their behaviour solely through information contained within the user's prompt. Examples, instructions and contextual information supplied during interaction temporarily shape subsequent responses without altering the underlying parameters of the model itself. This remarkable flexibility enables a single Artificial Intelligence system to perform numerous intellectual activities without requiring extensive task-specific modification.

Abstraction further contributes to generalisation. Human reasoning frequently depends upon recognising structural similarities between apparently unrelated problems, allowing principles acquired in one context to inform decisions made within another. Large Language Models increasingly demonstrate analogous behaviour by identifying conceptual relationships extending beyond surface linguistic similarity. Such abstraction underpins their growing effectiveness within education, scientific analysis, engineering design and numerous other knowledge-intensive disciplines.

Nevertheless, emergent capability should not be confused with unrestricted intelligence. Large Language Models remain probabilistic systems whose behaviour reflects statistical regularities acquired during training rather than conscious understanding. Their generalisation may therefore occasionally produce plausible but incorrect conclusions when confronted with unfamiliar or highly specialised problems. Appreciating both the strengths and limitations of emergence remains essential for responsible deployment within professional environments.

Prompt Design and Efficient Model Adaptation

The practical effectiveness of Large Language Models depends not solely upon their internal architecture but also upon the methods through which humans interact with them. Prompt engineering has therefore emerged as an important discipline concerned with designing instructions that enable Artificial Intelligence to interpret objectives accurately whilst producing reliable and contextually appropriate responses. Carefully constructed prompts provide structure, clarify expectations and establish operational constraints that significantly influence model behaviour.

Prompt engineering reflects an important principle of human communication. Effective collaboration frequently depends less upon the knowledge possessed by participants than upon the clarity with which objectives are expressed. Large Language Models exhibit similar characteristics because their responses are conditioned by the contextual information supplied during interaction. Explicit instructions concerning style, reasoning, evidence or domain expertise frequently improve both accuracy and coherence by reducing ambiguity regarding the intended task.

In-context learning complements prompt engineering by enabling models to adapt temporarily through examples provided within the conversation itself. Rather than modifying the underlying neural architecture, the model interprets demonstrations contained within the prompt and applies analogous reasoning throughout subsequent responses. This capability allows organisations to establish consistent analytical styles, reporting formats or procedural methodologies without undertaking computationally expensive retraining.

Fine-tuning represents a more permanent form of adaptation. Following general pre-training, Large Language Models may be refined using carefully curated domain-specific information aligned with particular organisational objectives. Healthcare institutions, engineering companies, legal practices, financial organisations and scientific laboratories increasingly employ fine-tuning to develop specialised Artificial Intelligence systems possessing expertise closely matched to their operational environments. Such adaptation strengthens performance whilst preserving the broad linguistic competence acquired during initial training.

More recently, parameter-efficient fine-tuning techniques have significantly reduced the computational resources required for customisation. Rather than modifying every parameter within extremely large models, smaller adaptable components may be trained whilst the majority of the network remains unchanged. These methods improve accessibility by enabling organisations to develop highly specialised Large Language Models without the substantial infrastructure previously associated with large-scale model training.

Applications Across Research, Industry and Public Services

The widespread adoption of Large Language Models reflects their extraordinary versatility across diverse organisational environments. Their ability to interpret natural language, synthesise extensive information and generate coherent analytical outputs enables them to support numerous professional activities previously requiring substantial human effort. Consequently, Large Language Models have become foundational technologies throughout contemporary enterprise computing.

Within scientific research, these models assist investigators by reviewing literature, summarising research findings, identifying conceptual relationships and supporting the preparation of technical documentation. Researchers retain responsibility for scientific judgement and experimental design, yet Artificial Intelligence substantially accelerates information synthesis across rapidly expanding bodies of knowledge. Similar benefits extend to engineering, where technical standards, design documentation and operational procedures may be analysed with unprecedented efficiency.

Financial institutions employ Large Language Models for document analysis, regulatory interpretation, customer communication and risk reporting. Legal organisations utilise them to examine legislation, summarise case law and assist contractual analysis. Healthcare providers increasingly integrate language models into administrative workflows, clinical documentation and knowledge management whilst maintaining appropriate safeguards concerning patient confidentiality and professional responsibility.

Education likewise represents an important area of application. Large Language Models support personalised learning by explaining complex concepts, generating educational material and providing adaptive tutoring responsive to individual learning requirements. Universities employ these technologies to strengthen academic support, research administration and knowledge dissemination whilst encouraging responsible use consistent with scholarly integrity.

Public administration, manufacturing, software development and customer service each illustrate further domains in which Large Language Models improve organisational efficiency by transforming unstructured textual information into actionable knowledge. Their value lies not simply in automation but in strengthening the capacity of institutions to manage increasing informational complexity whilst maintaining high standards of communication and analytical consistency.

Hallucination, Knowledge Limits, Bias and Reliability

Despite their remarkable capabilities, Large Language Models possess important limitations arising from their probabilistic foundations. Because responses are generated through statistical prediction rather than explicit factual verification, models may occasionally produce information that appears convincing despite lacking objective accuracy. Such behaviour, commonly described as hallucination, represents one of the principal challenges confronting the responsible deployment of contemporary Artificial Intelligence.

Hallucinations arise because the model seeks to generate coherent continuations according to learned statistical relationships rather than consulting authoritative knowledge sources directly. When confronted with incomplete, ambiguous or unfamiliar information, probabilistic prediction may produce plausible yet unsupported assertions. Although continuing advances in training methodology and retrieval-augmented generation have reduced the frequency of such behaviour, it remains an important consideration whenever Artificial Intelligence contributes to professional decision-making.

Current Knowledge, Retrieval Grounding and Bias

Knowledge limitations also arise from the static nature of training. Once training concludes, the internal representations contained within the model reflect the information available during that process. Subsequent developments in science, technology, legislation or public policy may therefore remain absent unless incorporated through additional adaptation or external retrieval systems. Organisations increasingly address this limitation by integrating Large Language Models with continually updated knowledge repositories capable of supplying authoritative information during inference.

Bias constitutes another significant challenge. Training corpora inevitably reflect cultural, historical and linguistic characteristics present within human communication. Consequently, statistical representations may inadvertently reproduce undesirable patterns or assumptions embedded within the underlying data. Responsible deployment therefore requires continual evaluation, monitoring and mitigation to ensure that Artificial Intelligence supports fairness, inclusivity and equitable decision-making.

Reliability ultimately depends upon recognising that Large Language Models function most effectively as intelligent assistants rather than infallible authorities. Human expertise, critical evaluation and appropriate governance remain indispensable whenever outputs influence significant professional, legal or ethical decisions.

Transparent, Private and Responsible Artificial Intelligence Governance

The increasing integration of Large Language Models into organisational decision-making has elevated governance from a peripheral concern to a central requirement of Artificial Intelligence strategy. Institutions deploying these technologies must ensure that operational benefits are accompanied by appropriate safeguards concerning transparency, accountability, privacy and legal compliance. Responsible Artificial Intelligence therefore encompasses not only technical performance but also the institutional frameworks through which these systems are developed, evaluated and supervised.

Transparency remains an important objective because users must understand the capabilities and limitations of probabilistic reasoning. Artificial Intelligence should communicate uncertainty appropriately, distinguish evidence from inference and avoid presenting unsupported conclusions as established fact. Such transparency strengthens trust whilst enabling informed human judgement regarding the appropriate use of model outputs.

Privacy assumes particular importance where Large Language Models process commercially sensitive, personal or confidential information. Many organisations therefore deploy models within secure environments or combine them with retrieval systems that maintain strict control over proprietary knowledge. These approaches support compliance with data protection legislation whilst preserving organisational sovereignty over valuable information assets.

Ethical considerations extend beyond technical implementation towards broader societal questions concerning employment, intellectual property, misinformation and equitable access to advanced technologies. Large Language Models should enhance rather than diminish human capability, supporting professional expertise whilst preserving meaningful human responsibility for consequential decisions. Effective governance therefore requires collaboration between technologists, policymakers, organisational leaders and academic researchers to ensure that Artificial Intelligence develops consistently with broader societal interests.

Multimodal, Efficient and Integrated Future Models

The future evolution of Large Language Models is unlikely to depend solely upon increasing computational scale. Although larger models continue to demonstrate important improvements, contemporary research increasingly emphasises architectural innovation, reasoning capability, efficiency and multimodal integration. Future systems are expected to combine textual understanding with visual, auditory and structured information, enabling more comprehensive representations of complex real-world environments.

Convergence with Reasoning and Action Architectures

Reasoning capability will continue to strengthen through closer integration between language models and specialised reasoning architectures capable of sustained analytical deliberation. Similarly, Large Action Models will increasingly combine linguistic understanding with autonomous planning and execution, allowing Artificial Intelligence to coordinate practical activities across digital and physical environments. Large Language Models will therefore remain foundational components within broader intelligent ecosystems rather than existing as isolated conversational systems.

Efficiency will become increasingly significant as organisations seek sustainable Artificial Intelligence capable of operating across local infrastructure, edge computing environments and resource-constrained devices. Advances in compression, quantisation and parameter-efficient adaptation will broaden accessibility whilst reducing operational cost and environmental impact.

Ultimately, Large Language Models will continue to evolve from sophisticated language processors into comprehensive knowledge infrastructures supporting reasoning, collaboration and intelligent action. Their future significance will depend not solely upon computational sophistication but upon the extent to which they remain trustworthy, explainable and aligned with human values throughout increasingly widespread deployment.

Large Language Models as Knowledge Infrastructure for Future Artificial Intelligence

Large Language Models represent a defining achievement in the history of Artificial Intelligence because they establish language as the principal medium through which computational systems acquire, represent and apply knowledge. Their emergence has transformed natural language processing from a collection of specialised techniques into a general framework capable of supporting communication, reasoning, knowledge synthesis and problem solving across an extraordinary diversity of intellectual domains. The combination of transformer architectures, large-scale representation learning and probabilistic inference has produced Artificial Intelligence systems whose versatility fundamentally alters expectations regarding computational capability.

Their influence extends far beyond language generation itself. Large Language Models now underpin contemporary developments in reasoning models, action models, multimodal Artificial Intelligence and intelligent agents, providing the foundational representations upon which increasingly sophisticated cognitive architectures are constructed. They have become essential components of scientific research, engineering, finance, healthcare, education and public administration, enabling organisations to transform rapidly expanding information resources into actionable knowledge.

Nevertheless, their continued success depends upon recognising both their capabilities and their limitations. Responsible deployment requires robust governance, continual evaluation, transparent communication and sustained human oversight to ensure that probabilistic reasoning supports rather than replaces informed professional judgement. As Artificial Intelligence continues to evolve, Large Language Models will remain central to that evolution, not because they represent the final stage of intelligent computation, but because they provide the linguistic and conceptual foundation from which future generations of reasoning, action and collaborative intelligence will continue to emerge.

Bibliography

• Bommasani, R. and others, On the Opportunities and Risks of Foundation Models (Stanford: Stanford University, 2021).
• Brown, T. B. and others, ‘Language Models are Few-Shot Learners’, Advances in Neural Information Processing Systems, 33 (2020), 1877-1901.
• Devlin, J., Ming-Wei Chang, Kenton Lee and Kristina Toutanova, ‘BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding’, Proceedings of NAACL-HLT (2019).
• Hoffmann, J. and others, ‘Training Compute-Optimal Large Language Models’, Advances in Neural Information Processing Systems, 35 (2022).
• Kaplan, J. and others, ‘Scaling Laws for Neural Language Models’, arXiv (2020).
• Liu, P. and others, ‘Pre-train, Prompt and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing’, ACM Computing Surveys, 55.9 (2023).
• Radford, A. and others, ‘Improving Language Understanding by Generative Pre-Training’ (OpenAI, 2018).
• Raffel, C. and others, ‘Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer’, Journal of Machine Learning Research, 21 (2020), 1-67.
• Touvron, H. and others, ‘Llama 2: Open Foundation and Fine-Tuned Chat Models’, arXiv (2023).
• Vaswani, A. and others, ‘Attention Is All You Need’, Advances in Neural Information Processing Systems, 30 (2017).
• Wei, J. and others, ‘Emergent Abilities of Large Language Models’, Transactions on Machine Learning Research(2022).
• Wolf, T. and others, ‘Transformers: State-of-the-Art Natural Language Processing’, Proceedings of EMNLP(2020).
• Zhao, W. X. and others, ‘A Survey of Large Language Models’, ACM Computing Surveys, 57.3 (2025), 1-38.