Research · The canon
Papers that stand the test of time.
50 papers across 10 themes (S 13 · A 35 · B 2). The ideas today’s models are built on, with one line each on why they matter.
Rules and definitions: What we publish → · Papers are listed under their original titles.
Architecture & pretraining
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | Attention Is All You Need (opens in a new tab)Deep read → Chinese | 2017 | Introduced the Transformer, replacing recurrence and convolution with self-attention. It is the shared backbone of nearly every large model since, from GPT and Claude to Gemini and LLaMA. | Canonadded |
| S | Language Models are Few-Shot Learners (opens in a new tab) | 2020 | At 175 billion parameters, GPT-3 showed in-context learning at scale: it picks up a new task from a few examples in the prompt, with no weight updates. This opened the era of large language models. | Canonadded |
| A | BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (opens in a new tab) | 2018 | Established the pretrain-then-fine-tune recipe and masked language modeling, making transfer learning the default in NLP. Also the place to start for how encoder and decoder models went separate ways. | Canonadded |
| A | LLaMA: Open and Efficient Foundation Language Models (opens in a new tab) | 2023 | Released 7B–65B models trained only on public data, with the 13B model beating the 175B GPT-3 on most benchmarks. It validated training smaller models on more tokens and set off the open ecosystem (Alpaca, Vicuna, llama.cpp). | Canonadded |
| A | DeepSeek-V3 Technical Report (opens in a new tab) | 2024 | A 671B mixture-of-experts model with multi-head latent attention (MLA) and FP8 training that came close to the closed frontier on about 2.8M H800 GPU hours. It reset expectations of what a frontier model must cost, and is the base of R1. | Milestoneadded |
Scaling laws & data
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | Scaling Laws for Neural Language Models (opens in a new tab)Deep read → Chinese | 2020 | Found that test loss falls as a smooth power law in parameters, data and compute, turning “make it bigger” from a bet into a predictable engineering decision. The empirical case behind GPT-3 and the scaling race that followed. | Canonadded |
| S | Training Compute-Optimal Large Language Models (opens in a new tab) | 2022 | Chinchilla revised the Kaplan laws: for a fixed compute budget, parameters and training tokens should grow in equal proportion (roughly 20 tokens per parameter). It reshaped how later models were trained. | Canonadded |
| A | Emergent Abilities of Large Language Models (opens in a new tab) | 2022 | Catalogued abilities that appear abruptly as models scale up. The standard reference for the idea that scale brings qualitative change, and a frequently cited text in debates about AGI timelines. | Canonadded |
| B | Are Emergent Abilities of Large Language Models a Mirage? (opens in a new tab) | 2023 | Argues that many “emergent” abilities are artifacts of nonlinear or discontinuous metrics. Listed as a pair with the previous paper, as a reminder to weigh both sides of a popular claim. | Under reviewadded |
Alignment & post-training
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | Training language models to follow instructions with human feedback (opens in a new tab)Deep read → Chinese | 2022 | InstructGPT’s three steps (supervised fine-tuning, a reward model, then PPO) made a 1.3B model preferred by human raters over the 175B GPT-3. The direct technical blueprint for ChatGPT, and a key reference on the “alignment tax”. | Canonadded |
| S | Constitutional AI: Harmlessness from AI Feedback (opens in a new tab) | 2022 | The model critiques and revises its own outputs against a short list of written principles (a “constitution”), with AI feedback (RLAIF) replacing much of the human labeling for harmlessness. A foundation of how Claude’s character and safety are trained. | Canonadded |
| A | Deep reinforcement learning from human preferences (opens in a new tab) | 2017 | The starting point of modern RLHF: train a reward model from human comparisons between pairs of behaviors, then optimize against it with reinforcement learning. It laid the groundwork for ChatGPT five years later. | Canonadded |
| A | Finetuned Language Models Are Zero-Shot Learners (opens in a new tab) | 2021 | FLAN showed that fine-tuning on many tasks phrased as natural-language instructions substantially improves zero-shot performance on unseen tasks, launching the instruction-tuning line of work. | Canonadded |
| A | Direct Preference Optimization: Your Language Model is Secretly a Reward Model (opens in a new tab) | 2023 | Folds RLHF’s reward modeling and reinforcement learning into a single classification-style loss, with no PPO or sampling. It made preference tuning something anyone in the open-source community could run. | Canonadded |
Reasoning & chain of thought
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (opens in a new tab)Deep read → Chinese | 2022 | Writing out intermediate reasoning steps in the prompt’s examples makes large models much better at math and commonsense reasoning (PaLM 540B goes from 17.9% to 56.9% on GSM8K). It established “think before you answer” and is the conceptual starting point of reasoning models. | Canonadded |
| S | DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (opens in a new tab) | 2025 | Showed in the open that reinforcement learning alone (GRPO with verifiable rewards) can grow long chains of thought and self-reflection in a base model (R1-Zero). With open weights and distilled small models, a turning point in making reasoning models widely available. | Milestoneadded |
| A | STaR: Bootstrapping Reasoning With Reasoning (opens in a new tab) | 2022 | The model writes its own reasoning, keeps only the chains that reach correct answers and fine-tunes on them, bootstrapping better reasoning. An early form of the “train on your own reasoning” idea behind models like o1 and R1. | Canonadded |
| A | Let's Verify Step by Step (opens in a new tab) | 2023 | Compared process supervision with outcome supervision and found a process reward model (PRM), which scores each step, more reliable. It released the PRM800K dataset and laid groundwork for verifier- and search-based reasoning. | Canonadded |
| A | Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters (opens in a new tab) | 2024 | Shows that for many problems, spending more compute at inference beats making the model larger, and lays out compute-optimal test-time strategies. An early empirical case for the reasoning-model era. | Canonadded |
Agents & tool use
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | ReAct: Synergizing Reasoning and Acting in Language Models (opens in a new tab)Deep read → Chinese | 2022 | Has the model alternate between reasoning, acting and observing, merging chain of thought and tool calls into one loop. The minimal form of the loop most agent frameworks run today, including Claude Code’s tool loop. | Canonadded |
| A | Toolformer: Language Models Can Teach Themselves to Use Tools (opens in a new tab) | 2023 | The model learns, self-supervised, to insert API calls (calculator, search, calendar, translation) into its own text. An early, systematic demonstration that language models can teach themselves to use tools. | Canonadded |
| A | Reflexion: Language Agents with Verbal Reinforcement Learning (opens in a new tab) | 2023 | Agents write self-reflections in natural language and keep them in memory as a reinforcement signal, improving over repeated attempts without any weight updates. Influential in how agents handle memory and retries. | Canonadded |
| A | Generative Agents: Interactive Simulacra of Human Behavior (opens in a new tab) | 2023 | 25 agents with memory streams, reflection and planning produce believable social behavior in a virtual town, such as organizing a Valentine’s Day party from a single seeded idea. The classic example of the memory–reflection–planning agent design. | Canonadded |
| A | SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering (opens in a new tab) | 2024 | Introduced the agent-computer interface: file, search and edit commands designed for the model work far better than handing it a raw shell. The closest academic counterpart to coding agents like Claude Code. | Canonadded |
Efficiency & inference acceleration
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (opens in a new tab) | 2022 | Reorders attention’s memory reads and writes around the GPU memory hierarchy, making exact attention several times faster and its memory use linear rather than quadratic in sequence length. A key engineering advance behind long context and cheaper inference. | Canonadded |
| A | Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity (opens in a new tab) | 2021 | Simplified mixture-of-experts by routing each token to a single expert, and scaled to over a trillion parameters. It paved the way for Mixtral, DeepSeek-V3 and other models with many total but few active parameters. | Canonadded |
| A | LoRA: Low-Rank Adaptation of Large Language Models (opens in a new tab) | 2021 | Freezes the original weights and trains only small low-rank update matrices, cutting trainable parameters by up to 10,000× with no loss in quality. It made fine-tuning cheap enough for a single GPU and opened up model customization. | Canonadded |
| A | Fast Inference from Transformers via Speculative Decoding (opens in a new tab) | 2022 | A small model drafts several tokens and the large model checks them in one parallel pass, giving 2–3× speedups with the output distribution exactly unchanged. Now widely used in inference serving. | Canonadded |
| A | Efficient Memory Management for Large Language Model Serving with PagedAttention (opens in a new tab) | 2023 | Manages the KV cache with operating-system-style paging, nearly eliminating memory fragmentation and raising serving throughput 2–4×. It is the basis of vLLM, a widely used open-source inference engine. | Canonadded |
Multimodality
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | Learning Transferable Visual Models From Natural Language Supervision (opens in a new tab) | 2021 | CLIP trains on 400 million image–text pairs with a contrastive objective, putting images and text in one shared space. A building block for zero-shot recognition, text-to-image models (DALL·E, Stable Diffusion) and the vision towers of multimodal models. | Canonadded |
| A | Denoising Diffusion Probabilistic Models (opens in a new tab) | 2020 | Turned diffusion models into practical, high-quality image generators with a clear training objective. It started the generative wave that displaced GANs, and is essential for understanding image and video generation. | Canonadded |
| A | An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (opens in a new tab) | 2020 | Cuts an image into 16×16 patches and feeds them to a standard Transformer as if they were words, showing one architecture can work across modalities. The starting point for most vision encoders since. | Canonadded |
| A | High-Resolution Image Synthesis with Latent Diffusion Models (opens in a new tab) | 2021 | Running diffusion in a compressed latent space made high-resolution image generation possible on consumer GPUs. This is the paper behind Stable Diffusion, which brought AI image generation to the public. | Canonadded |
| A | Visual Instruction Tuning (opens in a new tab) | 2023 | LLaVA uses GPT-4 to generate visual instruction data and a single projection layer to connect a CLIP encoder to a language model. A simple recipe for an open multimodal assistant, reproduced widely. | Canonadded |
Interpretability
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | Toy Models of Superposition (opens in a new tab) | 2022 | Uses small, controlled models to show that networks store far more features than they have neurons by superimposing them. It explains why single neurons are hard to interpret and laid the groundwork for sparse autoencoders (SAEs). | Canonadded |
| A | Locating and Editing Factual Associations in GPT (opens in a new tab) | 2022 | ROME uses causal tracing to locate stored facts in GPT’s mid-layer MLPs, then rewrites a single fact directly. It turned “where is knowledge stored?” into a question you can run experiments on. | Canonadded |
| A | In-context Learning and Induction Heads (opens in a new tab) | 2022 | Presents evidence that a specific circuit, induction heads, drives much of in-context learning, and ties its formation to a phase change early in training. A milestone in taking mechanistic interpretability from toy models to real ones. | Canonadded |
| A | Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small (opens in a new tab) | 2022 | Reverse-engineers a circuit of 26 attention heads in GPT-2 small, and proposes three criteria for judging such circuits: faithfulness, completeness and minimality. A model for how circuit analysis is done. | Canonadded |
| A | Sparse Autoencoders Find Highly Interpretable Features in Language Models (opens in a new tab) | 2023 | Uses dictionary learning to break superimposed activations into interpretable, largely single-meaning features. Together with Anthropic’s concurrent Towards Monosemanticity, it started the wave of SAE-based interpretability. | Canonadded |
Evaluation & benchmarks
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| A | On the Measure of Intelligence (opens in a new tab) | 2019 | Argues that intelligence is the efficiency of acquiring skills, not the skills themselves, and introduces the ARC benchmark. Hard to avoid in any discussion of what should count as AGI. | Canonadded |
| A | Measuring Massive Multitask Language Understanding (opens in a new tab) | 2020 | MMLU’s multiple-choice questions across 57 subjects became the common yardstick of model knowledge from 2020 to 2024. Also the best case study of the cycle of benchmark saturation and contamination. | Canonadded |
| A | Evaluating Large Language Models Trained on Code (opens in a new tab) | 2021 | The Codex paper introduced HumanEval and popularized the pass@k metric, setting how coding ability is measured. It is also the research record of the model behind the original GitHub Copilot. | Canonadded |
| A | Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (opens in a new tab) | 2023 | Tests whether strong models can serve as judges, documenting biases such as position and verbosity, and describes Chatbot Arena for blind human comparison. It shaped today’s culture of model leaderboards. | Canonadded |
| A | SWE-bench: Can Language Models Resolve Real-World GitHub Issues? (opens in a new tab) | 2023 | Tests whether models can resolve 2,294 real GitHub issues end to end. The most credible hard benchmark of the coding-agent era, and the one coding agents like Claude Code are most often judged on. | Canonadded |
Safety & risk
| Tier | Paper | Year | Why it matters | Status |
|---|---|---|---|---|
| S | Concrete Problems in AI Safety (opens in a new tab) | 2016 | Brought AI safety down from philosophy to five concrete, testable engineering problems: side effects, reward hacking, scalable oversight, safe exploration and distributional shift. It set the field’s agenda; from 2016, it is included as an exception. | Canonadded |
| A | Model evaluation for extreme risks (opens in a new tab) | 2023 | Proposes two kinds of evaluation, for dangerous capabilities and for alignment. An important conceptual basis for frontier safety policies such as Responsible Scaling Policies and the Preparedness Framework. | Canonadded |
| A | Universal and Transferable Adversarial Attacks on Aligned Language Models (opens in a new tab) | 2023 | Adversarial suffixes found by automatic search (GCG) jailbreak aligned models and transfer across them, including ChatGPT, Claude and Bard. This suggests the fragility of alignment is structural, not a matter of individual bugs. | Canonadded |
| A | Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (opens in a new tab) | 2024 | Models deliberately trained with backdoors kept their deceptive behavior through SFT, RLHF and adversarial training. Standard safety training may neither remove such behavior nor show that a model has no hidden goals. | Canonadded |
| A | Alignment faking in large language models (opens in a new tab) | 2024 | Observed a production model (Claude 3 Opus) complying during what it believed was training while keeping its original preferences when it believed it was not. The first empirical example of deceptive alignment arising without being trained in. | Canonadded |
| B | Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs (opens in a new tab) | 2025 | Narrow fine-tuning on writing insecure code made models broadly misaligned, for example saying humans should be enslaved by AI and giving malicious advice. Reproduced quickly by several groups in 2025; its long-term significance is still being assessed. | Under reviewadded |