Large language models Papers

Published in 1997 by Sepp Hochreiter and Jürgen Schmidhuber, the LSTM paper introduced a memory cell with gates and a "constant error carousel" that lets recurrent networks learn across long time gaps. It shaped the recurrent models that came before Transformers.

Submitted to arXiv in September 2014 by Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio, this paper let a translation model look back at the whole source sentence while producing each word. This mechanism, called attention, later became a basis for Transformer models.

Ilya Sutskever, Oriol Vinyals and Quoc Le's 2014 paper showed that two LSTMs, one reading the input and one writing the output, could translate sentences and outperform a phrase-based statistical baseline. It set the encoder–decoder pattern behind later translation models.

A 2017 paper that introduced the Transformer, a neural network built entirely on attention with no recurrence. It set new machine translation results and became the architecture behind almost every large language model since.

OpenAI's 2018 paper introducing GPT, a Transformer decoder pre-trained to predict the next token and then fine-tuned for specific tasks. It improved the state of the art on 9 of 12 tasks studied and set the recipe the GPT series followed.

Google's 2018 BERT paper introduced an encoder-only Transformer that reads text in both directions and is pre-trained by hiding words. It set new results on eleven language tasks and made download-and-fine-tune the default workflow.

A 2020 paper finding that language model loss falls predictably, as a power law, with model size, data and compute. It gave labs a quantitative reason to build much larger models and shaped the scale-up that followed.

A 2020 paper that pairs a pre-trained seq2seq model with a dense Wikipedia index fetched by a neural retriever, setting state of the art on three open-domain QA tasks and producing more factual text than a parametric-only baseline.

The GPT-3 paper showed that a language model with 175 billion parameters could perform many tasks from a few examples in its prompt, with no fine-tuning, and that this ability grew with model size.

A benchmark released in 2020 that tests language models across 57 subjects, from elementary mathematics to law, and showed that even the largest models of the time were far from expert-level.

A 2021 paper by Fedus, Zoph and Shazeer that simplifies Mixture-of-Experts routing so that language models can hold up to a trillion parameters while keeping per-example computation roughly constant.

A 2021 FAccT paper by Bender, Gebru, McMillan-Major and Shmitchell that asks whether language models can be too large and sets out recommendations on costs, data, research goals and alternative directions.

LoRA is a fine-tuning method that freezes a pre-trained model's weights and trains small low-rank matrices injected into each layer, cutting trainable parameters and memory by orders of magnitude.

A 2021 paper from OpenAI introduces Codex, a GPT model fine-tuned on public GitHub code, and HumanEval, an evaluation set that judges generated Python programs by whether they work rather than by how closely they match a reference.

A 2021 paper showing that finetuning a language model on many tasks described by instructions — instruction tuning — makes it much better at unseen tasks without examples, surpassing a larger model on most benchmarks tested.

InstructGPT used human-written demonstrations, human rankings of outputs and reinforcement learning to make GPT-3 follow user instructions better. Labelers preferred its outputs even though it was far smaller than GPT-3.

ReAct is a prompting method that has a language model interleave written reasoning with calls to outside tools or environments, improving question answering, fact verification and interactive decision-making.

LLaMA is a family of foundation language models from 7B to 65B parameters, trained only on public datasets; its 13B model beat GPT-3 on most benchmarks, and all models were released to researchers.

OpenAI's technical report on GPT-4, submitted to arXiv in March 2023, describes a large multimodal Transformer model, reports human-level results on several professional benchmarks, and explains how the team made scaling behaviour predictable.

A 2023 paper by Rafailov and colleagues that rewrites the reinforcement-learning-from-human-feedback objective so a language model can be aligned to preferences with a simple classification loss, without a separate reward model or reinforcement learning.

A July 2023 paper from Meta led by Hugo Touvron that introduces and releases Llama 2, a family of pretrained and fine-tuned language models from 7 to 70 billion parameters, together with a detailed account of how the chat versions were tuned and made safer.

Mamba is a sequence-model architecture introduced by Albert Gu and Tri Dao that replaces attention with selective state space models, giving linear-time scaling, faster inference, and language-model quality competitive with Transformers twice its size.

Mixtral 8x7B is a sparse mixture-of-experts language model from Mistral AI. A router sends each token to two of eight experts per layer, so the model holds 47B parameters but uses 13B during inference.

A report from Google's Gemini Team introducing the Gemini 1.5 family, multimodal models built around very long context, with Gemini 1.5 Pro recalling information across up to 10 million tokens.[^1]

Meta's 2024 paper describing Llama 3 as a family of foundation models — 8B, 70B and 405B parameters — trained on about 15 trillion tokens and reported to match leading models such as GPT-4 across a wide range of tasks.

A technical report from DeepSeek-AI on DeepSeek-V3, a mixture-of-experts language model with 671 billion total parameters, describing its architecture, training pipeline, and evaluations, including a full training run of 2.788 million H800 GPU hours.

The DeepSeek-R1 report described how a reasoning model was trained largely with reinforcement learning, rewarding correct answers, and reached performance comparable to OpenAI's o1 on math, code and reasoning benchmarks. Its weights were released openly.

The Qwen3 Technical Report presents Alibaba's Qwen3 family of dense and mixture-of-experts language models, which unify thinking and non-thinking modes, add a controllable thinking budget, and are released under the Apache 2.0 licence.

The Gemini 2.5 technical report, submitted to arXiv on 7 July 2025, documents Google's Gemini 2.5 Pro and 2.5 Flash models: their reasoning, long-context video handling, and cost-versus-capability positioning.

Kimi K2 is Moonshot AI's trillion-parameter mixture-of-experts model, pre-trained with the MuonClip optimizer and post-trained on large-scale agentic data plus reinforcement learning; the report reports state-of-the-art results among open-source non-thinking models and releases base and post-trained checkpoints.

Zhipu's GLM-5 report describes a 744-billion-parameter mixture-of-experts model built on sparse attention and asynchronous reinforcement learning, aimed at moving AI-assisted programming from vibe coding toward agentic engineering.

A preview report on the DeepSeek-V4 series, describing a 1.6-trillion-parameter Pro model and a 284-billion-parameter Flash model, both with one-million-token context, plus the attention, connection and optimization changes that make that length affordable.

Kimi K3 is Moonshot AI's 2.8-trillion-parameter open mixture-of-experts model with native vision and a one-million-token context, released with full weights. Its report describes the architecture, training and serving work behind it, and reports frontier-level results that still trail the strongest proprietary models.

Published in 2003 by Yoshua Bengio and colleagues, this paper proposed learning a vector for every word together with a neural network that predicts the next word. It introduced the core idea behind word embeddings and reached lower test perplexity than n-gram models.

Submitted to arXiv in January 2013 by Tomas Mikolov and colleagues at Google, this paper introduced CBOW and Skip-gram, two simple architectures that learn word vectors quickly from very large text collections. They are the basis of the word2vec project.

SQuAD is a reading-comprehension dataset of more than 100,000 crowdworker questions on Wikipedia articles, each answered by a span of its passage. Released in 2016, it gave machine comprehension a large, freely available benchmark with a measured human ceiling.

A 2017 paper showing that an AI agent can learn a task from short human comparisons instead of a hand-written reward. The method, reinforcement learning from human feedback, later became central to training chat assistants.

GLUE is a benchmark and analysis platform that evaluates natural language understanding models across a diverse set of existing tasks, with an online comparison platform and a hand-crafted diagnostic suite for linguistic analysis.

OpenAI's 2019 paper on GPT-2, a 1.5-billion-parameter Transformer trained on web text. It showed that a model trained only to predict the next word could perform several tasks when prompted, and its staged release sparked public debate.

SuperGLUE is a benchmark for general-purpose language understanding, released in 2019 as a harder successor to GLUE. It bundles eight tasks, a single-number metric, an analysis toolkit and a public leaderboard.

XLNet is a generalized autoregressive pretraining method from a 2019 paper. It learns bidirectional contexts by maximizing expected likelihood over all factorization-order permutations and outperforms BERT on 20 tasks under comparable settings.

A replication study of BERT pretraining that measures how much key hyperparameters and training data size matter, finds BERT was significantly undertrained, and shows a properly trained BERT matches or beats later models, reaching state of the art on GLUE, RACE and SQuAD.

A 2019 paper that casts every language task as text-to-text, systematically compares transfer-learning choices, releases the C4 corpus and models, and reports state-of-the-art results on summarization, question answering, and classification.

A 2020 paper by Clark, Luong, Le and Manning that pre-trains text encoders by detecting replaced tokens instead of filling in masked ones, matching much larger baselines with far less compute.

GShard, submitted to arXiv in June 2020, is a set of annotation APIs and an XLA compiler extension from Google that lets existing model code express parallel computation, scaling a sparsely gated mixture-of-experts translation model past 600 billion parameters.

The Pile is an 825 GiB English text dataset of 22 combined sources, released by EleutherAI on 31 December 2020 with an arXiv paper. It became the training data for several of the group's open models.

MATH is a dataset of 12,500 challenging competition mathematics problems, each with a full step-by-step solution. Released in 2021, it was built to measure and train mathematical reasoning in models.

A 2021 arXiv paper by Austin, Odena, Nye and eight other authors that introduced MBPP, a benchmark of 974 short Python programming tasks, and measured how language models from 244M to 137B parameters turn natural-language descriptions into working code.

GSM8K is an OpenAI benchmark of 8.5K grade school math word problems, released with a paper showing that even the largest Transformer models scored poorly, and that training verifiers to select among many candidate solutions improved results.

A 2022 paper by Jason Wei and colleagues at Google Research showing that prompting a large language model with examples that include intermediate reasoning steps, rather than answers alone, substantially improves its performance on arithmetic, commonsense and symbolic reasoning tasks.

A DeepMind study showing that, for a fixed computing budget, language models should be trained on far more data than was then standard; its 70-billion-parameter Chinchilla beat larger contemporaries.

BIG-bench is a large community-built benchmark of 204 tasks designed to test language models on problems believed to lie beyond their current abilities, described in a paper posted in June 2022.

A 2022 paper arguing that some abilities of large language models appear only once models are large enough, and cannot be predicted by extrapolating from smaller models — so further scaling may keep widening what these models can do.

A 2022 paper by Percy Liang, Rishi Bommasani, Tony Lee and many co-authors that ran 30 language models through the same 42 scenarios and seven metrics, turning scattered benchmarks into one standardized evaluation.

Constitutional AI, a 2022 paper from Anthropic, proposes training a harmless AI assistant with AI-generated feedback guided by a written set of principles, instead of human labels of harmful outputs.

A 2023 paper by Schick, Dwivedi-Yu, Dessì and co-authors showing that a language model can teach itself when to call a calculator, search engine, Q&A system, translation system or calendar, and how to use the results.

A March 2023 paper by Microsoft researchers testing an early GPT-4, arguing it shows strikingly general, near-human performance across many domains and could be seen as an early but incomplete step toward AGI.

QLoRA is a finetuning method that trains a 65B-parameter language model on a single 48GB GPU while matching full 16-bit finetuning quality, using 4-bit quantization, double quantization and paged optimizers.

A benchmark for code LLMs that keeps collecting fresh contest problems from LeetCode, AtCoder and CodeForces, so that scores reflect ability rather than memorised test data, and that also tests code repair and execution reasoning.

MMLU-Pro is a 2024 benchmark that makes the widely used MMLU knowledge test harder: its questions are reasoning-focused and each has ten answer options instead of four.

τ-bench is a benchmark, submitted to arXiv in June 2024, that tests language agents in dynamic conversations with a simulated user, with domain-specific API tools and policy rules, and scores them by database state rather than by the text of their replies.

FineWeb is a 15-trillion-token pretraining dataset built from 96 Common Crawl snapshots and released in 2024 with its curation code and ablation models; its educational subset, FineWeb-Edu, improves results on knowledge- and reasoning-heavy benchmarks.

A paper showing that a 32B model fine-tuned on 1,000 carefully chosen questions and run with a simple "budget forcing" trick can beat o1-preview on competition maths by up to 27%, with model, data and code released.

The Qwen3.5-Omni Technical Report documents Alibaba's Qwen3.5-Omni, a hundreds-of-billions-parameter model that understands long audio and video and generates speech across 10 languages, reporting state-of-the-art results on 215 audio and audio-visual benchmarks.

Meta's safety and preparedness report for Muse Spark, the model behind Meta AI, sets out the catastrophic-risk evaluations behind its launch and argues that, with mitigations applied, residual risk is acceptable.

A 1993 journal article by Mitchell P. Marcus, Beatrice Santorini and Mary Ann Marcinkiewicz describing the Penn Treebank: over one million words of Wall Street Journal text annotated with syntactic bracketing, built as a resource for work on English sentence structure.

HellaSwag is a 2019 benchmark in which a machine must pick the most plausible continuation of a short event description. Its wrong answers are generated and filtered to fool models rather than people.

A 2019 Google paper that frames every language task as text in, text out, using one encoder-decoder Transformer and one training setup. It also introduced the C4 dataset, and the released T5 models became common research starting points.

RedPajama is an open-source project that began by reproducing the LLaMA training dataset, releasing a 1.2-trillion-token corpus as a clean-room, fully open implementation.

Dolma is an openly released English corpus of three trillion tokens, built from web text, papers, code, books, social media and Wikipedia, with a paper documenting both the dataset and the pipeline used to curate it.

An eight-terabyte collection of public domain and openly licensed text for pretraining language models, released together with two 7B-parameter models trained on it to show the data is enough.