THE RESEARCH LIBRARY · PAPERS FOR FDES
AI research, explained for the work.
Understand the research behind the systems you build. Explore practical breakdowns of RAG, agents, evaluation and inference, with diagrams and experiments to work through.
- papers
- 291
- topics
- 18
- summaries
- Free
New to the research? Follow the starting path12 papers, in 6 stages
STAGE 01
Understand production work
Describe a customer workflow, its dependencies, what success means and who owns it after launch.
STAGE 02
Understand the model and grounding
Explain the architecture and build a basic question-answering app over documents.
STAGE 03
Measure retrieval and answers
Build a small eval set and test evidence placement, retrieval quality and faithfulness.
STAGE 04
Add controlled actions
Give an agent a few well-designed tools, clear stopping rules and useful feedback.
STAGE 05
Evaluate reliability
Write a rubric, check the judge against humans, and run repeated and perturbed tests.
STAGE 06
Understand customization and serving
Say when adaptation could help, and how context length and concurrency set the memory bill.
Find your next paper
Search an idea, choose a topic, or read in order.
291 papers
Foundations
The ideas every system you deploy is built on.
Hidden Technical Debt in ML Systems
Google's 2015 field guide to why ML systems are cheap to build and expensive to keep alive, and where the cost hides when nobody is looking.
Attention Is All You Need
Google Brain, Google Research
The 2017 Google paper that introduced the Transformer: attention and position information in place of recurrent and convolutional layers.
GPT-3: Few-Shot Learners
OpenAI
The 2020 OpenAI paper that showed a big enough model can learn a task from a few examples in its prompt, no fine-tuning, and made prompting a discipline.
InstructGPT
OpenAI
The 2022 OpenAI paper that turned GPT-3 into an assistant with human feedback, and showed a 1.3B model people preferred to the 175B original.
Chain-of-Thought Prompting
Google Research, Brain Team
The 2022 Google paper that showed examples with their working shown, not just their answers, can triple a large model's accuracy on multi-step problems.
RAG
Facebook AI Research, University College London, New York University
The 2020 paper that named RAG: a generator paired with a searchable index, so a model's knowledge can be checked, cited and updated without retraining.
Sentence-BERT
UKP Lab, Technische Universität Darmstadt
The 2019 paper that made BERT usable for semantic search: one vector per sentence, compared by cosine, turning a 65-hour job into 5 seconds.
Lost in the Middle
Stanford University, UC Berkeley, Samaya AI
The 2023 Stanford paper that showed models use evidence at the start and end of a long prompt far better than in the middle, sometimes worse than no evidence at all.
BERT
Google AI Language
Select 15% of tokens for prediction, train a Transformer encoder with context from both sides, then fine-tune for each task. BERT set records on eleven language benchmarks.
Mamba
Carnegie Mellon University, Princeton University
Make a state space model's update depend on the current token, so it can choose what to remember, and run it as a fused GPU scan. Transformer-level quality to 2.8B parameters with no KV cache.
DeepSeek-V3
DeepSeek
A 671B mixture-of-experts model, 37B active per token, trained in FP8 for 2.788M H800 GPU-hours, with a compressed attention cache and bias-balanced experts. Read the cost line closely.
Why Models Hallucinate
OpenAI, Georgia Tech
Training and evaluation can reward a confident guess. Understand the statistical argument, design abstention-aware scores and test the cost of miscalibration.
DeepSeek-V4
DeepSeek-AI
DeepSeek-V4 combines compressed sparse and heavily compressed attention for million-token contexts. Learn what changes in compute, cache state and deployment.
Kimi Linear
Moonshot AI
Kimi Linear mixes channel-wise recurrent memory with periodic full attention. Learn the gated delta update, hybrid state costs and the limits of its speed claims.
Scaling laws (Kaplan)
OpenAI, Johns Hopkins University
OpenAI showed that language model loss falls as a smooth power law in parameters, data and compute, and used it to argue for huge models trained briefly. The method lasted; the allocation did not.
Chinchilla
DeepMind
DeepMind trained over 400 models and found the big LLMs of 2021 were undertrained: grow parameters and tokens equally. A 70B model on 1.4T tokens beat the 280B Gopher at the same compute.
The reversal curse
Vanderbilt, Oxford, UK AI Safety Institute, Apollo Research
Fine-tune a model on "Uriah Hawthorne is the composer of Abyssal Melodies" and it cannot say who composed Abyssal Melodies. GPT-4 named 79% of celebrities' parents and 33% of their children.
Emergent abilities of large language models
Google Research, Stanford, UNC Chapel Hill, DeepMind
Some tasks sit at chance until a model reaches about 10^22 to 10^24 training FLOPs, then jump. Google surveyed dozens of such abilities and argued they cannot be predicted from smaller models.
Are emergent abilities a mirage?
Stanford University
Score GPT-3's arithmetic by exact match and ability appears suddenly at scale. Score the same outputs by token edit distance and it improves smoothly. Stanford argues the jump is in the metric.
On the Measure of Intelligence (ARC)
Chollet argues that skill at a task can be bought with training data or built-in knowledge, so intelligence should be measured as how efficiently a system acquires new skills. ARC is his test of it.
The Curse of Recursion: model collapse
University of Oxford, University of Cambridge, Imperial College London, University of Toronto, Vector Institute
Train each model on the previous model's output and rare events vanish first, then the distribution narrows. In our run of the paper's Gaussian example, 88% of chains collapse by generation 1,000.
Deduplicating training data
Google Research, University of Pennsylvania
One 61-word sentence appears 61,036 times in C4. Removing near-duplicates and repeated spans cut copied text in a model's output about tenfold, with no loss in perplexity.
The Pile: 800GB of diverse text
EleutherAI
EleutherAI combined 22 sources, from PubMed and ArXiv to GitHub, court opinions and Enron emails, into an 825 GiB open training set, weighted toward high-quality sources and documented in detail.
Hallucination is inevitable
National University of Singapore
A diagonal proof shows every computable model is wrong on some computable task. The math holds; what it tells you about your own task is far less than the title suggests.
On calibration of modern neural networks
Cornell University
Deeper, more accurate networks became overconfident: a 110-layer ResNet claims more certainty than it earns. One fitted temperature fixes most of it without changing a single prediction.
Word2vec: word vectors in vector space
Drop the hidden layer, train on six billion words, and vector arithmetic answers 'Athens is to Greece as Oslo is to ?'. The paper that made dense word vectors cheap to train.
BPE: rare words as subword units
University of Edinburgh
A 1994 compression trick, repurposed: merge the most frequent pair of symbols, again and again, until a fixed vocabulary covers every word. Where the tokens in your LLM bill come from.
RoPE: rotary position embedding
Zhuiyi Technology
Instead of adding a position vector, rotate each query and key by an angle set by its position. The score then depends only on distance. LLaMA and most open models since use this scheme.
Language modeling is compression
Google DeepMind, Meta AI and Inria
A model that predicts the next byte well can compress it well. Chinchilla 70B beat PNG and FLAC on image and audio chunks it never trained on, until its 140 GB of weights were counted.
Building and evaluating LLM apps
Agents, pipelines, structured output and how to grade them.
ReAct
Princeton University, Google Research Brain Team
The 2022 paper that interleaves reasoning with tool use: think, act, observe, then revise the plan. A foundation for agents that adapt to what their tools return.
DSPy
Stanford University, UC Berkeley, and others
The 2023 Stanford paper that treats an LLM pipeline as a program: declare what each call does, give a metric and a few examples, and let a compiler write the prompts.
Grammar-Constrained Decoding
EPFL, Université Grenoble Alpes
The 2023 EPFL paper showing that masking a model's tokens with a grammar guarantees the output's shape, lifts accuracy without training, and has a trap: valid but empty.
LLM-as-a-Judge
UC Berkeley, UC San Diego, Carnegie Mellon University, Stanford, MBZUAI
The 2023 paper that made using a strong model to grade other models respectable, by measuring how often it agrees with people and exactly how it is biased.
Ragas
Exploding Gradients, Cardiff University, AMPLYFI
The 2023 paper that split RAG quality into three checks you can run without reference answers: is the context focused, is the answer grounded in it, does it address the question.
LLM nondeterminism
Rice University, University of Minnesota Twin Cities, Adobe
Greedy decoding is not deterministic in practice. Changing batch size, GPU count or GPU type gave a reasoning model a 9% accuracy spread on AIME'24 in bfloat16; FP32 math removed it.
Codex and HumanEval: code scored by tests
OpenAI
OpenAI wrote 164 Python problems with unit tests and scored generated code by whether it passes. Codex-12B solved 28.8% with one sample and 72% with 100, which is where pass@k comes from.
METR's AI productivity trial
METR (Model Evaluation & Threat Research)
16 experienced open-source developers did 246 real issues, each randomly allowed or barred from AI tools. With AI they took 19% longer, while believing it had made them 20% faster.
ChatGPT behavior drift, March to June 2023
Stanford University, UC Berkeley
The same API model name gave different answers three months apart. GPT-4's prime-number accuracy fell from 84% to 51%, about what always answering 'composite' scores on this balanced set.
Pretraining on the test set (satire)
No affiliation listed
A satire: a 1-million-parameter model trained on 100,000 tokens of benchmark test sets scores perfectly on all of them. We build it for real and test which contamination checks catch it.
The SWE-Bench Illusion
Purdue University, Microsoft
Given only an issue's text and the repository name, o3 named the file to fix for 76% of SWE-Bench Verified tasks, but 55% for similar tasks from other popular repositories.
Let Me Speak Freely? JSON and reasoning
Appier AI Research, National Taiwan University
Asking for JSON lowered reasoning scores for several models. Much of the worst drop was output that failed to parse, and JSON mode let GPT-3.5 write its answer before its reasoning.
EmotionPrompt: 'important to my career'
Institute of Software CAS, Microsoft, William & Mary, Beijing Normal University, HKUST
Appending lines like 'This is very important to my career' to prompts lifted scores by 8% and 115%, the paper says. Both figures take the best of 11 phrases per task, chosen on the test data.
Mind Your Tone: rude vs polite prompts
Pennsylvania State University
GPT-4o answered 50 multiple-choice questions with polite or rude prefixes. Very rude prompts scored 84.8% and very polite ones 80.8%: a gap of two questions out of fifty.
GPTs are GPTs: LLM labor market exposure
OpenAI, OpenResearch, University of Pennsylvania
Rate every task in 1,016 US occupations by whether an LLM could halve its time. 80% of workers have a tenth of their tasks exposed; the share that matters depends on which of three measures you quote.
Retrieval, documents and memory
Getting the right context in front of the model, and keeping it.
RAG or Long-Context LLMs?
Google DeepMind, University of Michigan
The 2024 Google DeepMind study that found long context beats RAG on quality, RAG wins on cost, and a model that says when it cannot answer can route between them.
Late Chunking
Jina AI, Weaviate
The 2024 Jina AI paper that fixes chunks losing their context: run the whole document through the embedding model first, then cut the token vectors into chunks.
GraphRAG
Microsoft Research, Microsoft Strategic Missions and Technologies, Microsoft Office of the CTO
Microsoft's 2024 paper on answering questions about a whole corpus, not one passage: build an entity graph, summarize its communities, then map-reduce over the summaries.
ColPali
Illuin Technology, Equall.ai, CentraleSupélec, ETH Zürich
The 2024 paper that skips PDF parsing: embed each page image with a vision-language model into many small vectors, and let tables, charts and layout count in search.
Qwen3-VL Embedding and Reranker
Tongyi Lab, Alibaba Group
Alibaba's 2026 report on one embedding model and one reranker for text, images, document pages and video, and the trade-offs of dimensions, precision and reranking.
RAGPerf
University of Illinois Urbana-Champaign, IBM Research
The 2026 UIUC and IBM benchmark that profiles a RAG system stage by stage and finds where the time, memory and money really go, rarely where teams look.
Agentic Context Engineering (ACE)
Stanford University, SambaNova Systems, UC Berkeley
The 2025 Stanford and SambaNova paper that improves agents by growing a detailed playbook of lessons in their context, with small edits instead of rewrites, and no weight updates.
Recursive Language Models
MIT CSAIL
The 2025 MIT paper that stops stuffing long inputs into the context window: put the input in a variable, let the model write code to slice it and call itself on the pieces.
Mem0
Mem0
The 2025 paper behind the Mem0 memory layer: extract facts from conversations, keep them current with add, update and delete, and trade a little accuracy for a lot of speed.
Self-RAG
University of Washington, Allen Institute for AI, IBM Research AI
Train a model to emit its own control tokens: whether to retrieve, whether a passage is relevant, whether each sentence is supported. Beam search on those scores; strong results, with fine print.
RAPTOR
Stanford University
Cluster a document's chunks, summarize each cluster, repeat to a root, and search every level at once: broad questions get summaries, narrow ones get chunks. Gains are modest.
RAFT
UC Berkeley
Fine-tune on questions paired with the right document plus distractors, with answers that quote the right one, so the model learns to read past bad retrieval. Strong on APIs; read the tables closely.
Sufficient Context
Google, UC San Diego, Duke University
Google splits RAG errors by whether the retrieved context could answer the question at all. Models answer wrongly rather than abstain when it cannot, and right 35-62% of the time anyway.
HippoRAG 2
The Ohio State University, University of Illinois Urbana-Champaign
Graph retrieval that doesn't lose to plain embeddings on simple questions: LLM-extracted triples plus passage nodes, searched with Personalized PageRank. Real multi-hop gains, at 8x the indexing time.
ReasonIR
FAIR at Meta, University of Washington, National University of Singapore, Allen Institute for AI, Stanford, MIT, UC Berkeley
Meta trains an 8B retriever on LLM-written hard queries and hard negatives. It leads BRIGHT, helps RAG on GPQA, and shows why an LLM reranker that scores 0-5 needs a tie-breaker.
Nemotron ColEmbed V2
NVIDIA
NVIDIA's open late-interaction models retrieve PDF pages from their images and top ViDoRe V3. The paper also prices them: about 6 TB of vectors per million pages for the 8B model, and how to cut it.
LLMs and Vector Databases
Carnegie Mellon, Purdue, University of Michigan, Harbin Institute of Technology (Shenzhen), Chinese Academy of Sciences, Shandong University of Technology
Vector databases support retrieval, semantic caching and agent memory. Learn what each role requires, where similarity fails and how to test filtered search.
RemoteANN
Nanyang Technological University, ByteDance
RemoteANN redesigns vector search around object-storage requests. Learn its pruning, byte-range planning and caching tradeoffs, with a runnable request planner.
Embedding Retrieval Limits
Google DeepMind, Johns Hopkins University
Learn what single-vector geometry can restrict, what LIMIT actually tests and why dimensions, relevance alignment and candidate recall need separate diagnosis.
Qwen3 Embedding
Tongyi Lab, Alibaba Group
Separate Qwen3's embedding and reranking models, understand their instruction-aware training, and test vector storage, scoring arithmetic and candidate-recall limits.
Gemini Embedding
Gemini Embedding Team, Google
Gemini Embedding combines a language-model foundation with task-aware contrastive training. Learn its data recipe, false-negative handling and benchmark caveats.
When to Use Graphs in RAG
Xiamen University, The Hong Kong Polytechnic University
GraphRAG-Bench compares graph and conventional retrieval across task types. Learn where graphs help, where evidence gets lost and how to evaluate the extra cost.
DeepSeek-OCR
DeepSeek-AI
Can a page image replace a much longer text prefix? DeepSeek-OCR explores optical compression, with useful OCR results and important limits for agent memory.
Context Is Gold: InSeNT
Illuin Technology, Equall.ai, CentraleSupélec, EPFL
ConTEB and InSeNT show why a passage needs its document context, and why training must still distinguish neighboring passages that answer different questions.
Jina Reranker v3
Jina AI, University of Pittsburgh
Jina's listwise reranker lets candidate documents interact before scoring. Learn its causal attention boundaries, order sensitivity and practical batching constraints.
Filtered ANN Search
University of California, Merced
A fast vector index can lose recall after a metadata filter. Learn how selectivity, local neighborhoods and query plans change filtered-search behavior.
HyDE: hypothetical document embeddings
Carnegie Mellon University, University of Waterloo
Ask a model to write a fake answer, embed it, and search for real documents near it. With no labels, HyDE lifted an unsupervised retriever 16.8 nDCG points on TREC DL19.
RETRO: retrieving from trillions of tokens
DeepMind
RETRO retrieves from a 2-trillion-token database for every 64-token chunk and rivals models about 25 times its size. Some of its most striking scores come from test text already in the database.
Text embeddings reveal almost as much
Cornell University
A vector store holds no text, so it is safe to host anywhere? Vec2Text rebuilt 92% of 32-token passages exactly from their embeddings, and 89% of names in clinical notes.
Agents, evaluation and security
What agents do in the real world, how they fail, and how they are attacked.
SWE-agent
Princeton University
The 2024 Princeton paper that showed an agent's tools matter as much as its model: a purpose-built interface for searching, viewing and editing code beat a raw shell by seven points.
AI Agents That Matter
Princeton University
The 2024 Princeton critique of agent benchmarks: report cost next to accuracy, compare against simple baselines, keep a real held-out set, and stop mistaking leaderboards for procurement advice.
Why Multi-Agent Systems Fail
UC Berkeley, Intesa Sanpaolo
Berkeley read 1,642 traces from seven multi-agent frameworks and named 14 ways they fail. Most failures trace back to how the agents are organized and checked, not to the model inside them.
Measuring Agents in Production
UC Berkeley, Intesa Sanpaolo, UIUC, Stanford, IBM Research
Twenty deployment teams and 86 deployed systems describe how agents actually ship: frontier models prompted by hand, fixed workflows, short leashes and humans checking the output.
Reflexion
Northeastern University, MIT, Princeton University
The 2023 paper that taught agents to learn from failure without training: after each failed attempt, the model writes itself a lesson and reads it on the next try.
SWE-bench
Princeton University, University of Chicago
The benchmark that moved coding evaluation from toy functions to real repositories: 2,294 GitHub issues, each graded by the project's own tests. In 2023 the best model fixed under 2%.
τ-Knowledge
Sierra, Princeton University
A support-agent benchmark built like a real deployment: 698 interlinked bank documents, tools you can only find by reading them, and a simulated customer. The best agent passed a quarter of tasks.
AI Agent Reliability
Princeton University
Princeton borrowed reliability engineering from aviation and nuclear power, turned it into twelve metrics for AI agents, and found that two years of accuracy gains bought only small reliability gains.
SkillsBench
BenchFlow, Ohio State, UC Berkeley and others
Agent Skills, folders of instructions and scripts an agent loads on demand, lifted pass rates from 34% to 51% across 18 agent setups. Focused Skills helped most; self-written ones hurt.
ττ-bench (hyper-tau-bench)
Sierra, Princeton University
A benchmark that turns an FDE engagement into the task: messy records, a client to interview, a flaky API and a serving budget. AI-built agents scored under a third of an expert's.
AgentDojo
ETH Zurich, Invariant Labs
The standard testbed for prompt injection against tool-using agents: 97 tasks, 629 attack cases, and a finding that the most capable models were often the easiest to hijack.
Indirect Prompt Injection Arena
Gray Swan AI, with OpenAI, Anthropic, Meta, UK AISI and US CAISI
464 red-teamers made 272,000 attempts to hijack 13 frontier agents and make them hide it. Every model fell at least sometimes, and attacks that beat the strongest model worked almost everywhere.
METR Task Horizons
Model Evaluation & Threat Research (METR)
Translate agent benchmark scores into human task difficulty, distinguish a 50% horizon from safe autonomy and design a customer-specific reliability evaluation.
τ²-Bench
Sierra, University of Toronto, Vector Institute
An agent can know the fix and still fail to guide a user through it. τ²-Bench measures shared control, stateful user simulation and repeated task success.
Vending-Bench: an agent runs a business
Andon Labs
An LLM agent runs a simulated vending machine for months: order stock by email, set prices, pay a $2 daily fee. Claude 3.5 Sonnet averaged $2,218 in net worth, and every model had runs that derailed.
Asleep at the keyboard: Copilot security
New York University, University of Calgary
About 40% of Copilot's suggestions at 89 security-sensitive spots had a known weakness. The code already in the file mattered most: 0 of 18 bad with safe neighbors, 17 of 18 with unsafe ones.
Do users write less secure code with AI?
Stanford University
People given an AI assistant wrote less secure code on four of five tasks and felt more sure of it. Across 47 participants the direction holds; the paper's adjusted tests stop short of proof.
Customization and reasoning
Changing a model's behavior, and paying for it to think longer.
LoRA
Microsoft
The 2021 Microsoft paper behind most fine-tuning today: freeze the model, learn each layer's change as two thin matrices, and match full fine-tuning of GPT-3 with 10,000× fewer trained parameters.
QLoRA
University of Washington
LoRA on top of a 4-bit base model: QLoRA fit a 65B fine-tune onto one 48 GB GPU with no measured loss, and showed that 9,000 good examples beat 450,000 mediocre ones.
DPO
Stanford University
The 2023 paper that fits preferences with a classification loss over chosen and rejected answers, removing the explicit reward-model fitting and reinforcement-learning loop used in PPO-based RLHF.
Tülu 3
Allen Institute for AI, University of Washington
AI2's fully open post-training recipe: every dataset, script and evaluation used to turn Llama 3.1 base models into assistants that matched closed models, plus RL with verifiable rewards.
DeepSeek-R1
DeepSeek-AI
DeepSeek combined reinforcement learning with cold-start examples to reach o1-level reasoning. R1-Zero separately showed reasoning emerging from RL without supervised fine-tuning first.
Scaling Test-Time Compute
UC Berkeley, Google DeepMind
When does thinking longer beat a bigger model? Spending inference compute by question difficulty used up to 4× less compute and beat a 14× larger model on questions within reach.
Zero-shot chain of thought
University of Tokyo, Google Research
One sentence, "Let's think step by step", took text-davinci-002 from 17.7% to 78.7% on MultiArith with no examples. The paper also shows how much a second answer-extraction call matters.
STaR: self-taught reasoner
Stanford University, Google Research
A 6B model writes its own reasoning data: keep rationales that reach the right answer, rationalize the misses, fine-tune, repeat. It nearly matched a 30x larger fine-tuned model on CommonsenseQA.
Quiet-STaR: thinking at every token
Stanford University, Notbad AI
STaR without labels: Mistral 7B learns to write a short hidden thought before each token of web text, rewarded when it predicts what comes next. Zero-shot GSM8K rose from 5.9% to 10.9%.
s1: simple test-time scaling
Stanford University, University of Washington, Allen Institute for AI, Contextual AI
Fine-tune a 32B model on 1,000 chosen reasoning traces for 26 minutes, then append "Wait" when it tries to stop thinking. AIME24 went from 50% to 57%, which on 30 problems is two answers.
Process supervision (PRM800K)
OpenAI
OpenAI had people label every step of 75,000 math solutions. A reward model trained on those step labels picked correct answers 78.2% of the time, against 72.4% for one trained on final answers.
Coconut: reasoning in latent space
Meta FAIR, UC San Diego
Instead of writing out each reasoning step, Coconut feeds the model's hidden state straight back in as the next input. On a graph-search task GPT-2 beat chain of thought with a third of the tokens.
Recurrent depth: thinking by looping
ELLIS Institute Tübingen and MPI-IS, University of Maryland, Lawrence Livermore National Laboratory
A 3.5B model that thinks by looping its middle block instead of writing more tokens. At 32 loops it does the forward compute of a 52.6B model, and benchmark scores rise as loops are added.
Dot by dot: filler-token computation
New York University
Replace a chain of thought with a row of dots and a small transformer still solves a task it fails without them. The dots carry no information, only computation that no reader of the output can see.
Pause tokens: delay before answering
Carnegie Mellon University, Google Research
Append 10 or 50 copies of a learned <pause> token to the prompt and read the answer only after the last one. Trained this way from pretraining on, a 1B model's SQuAD score rose from 36.4 to 55.9.
DeepSeekMath and GRPO
DeepSeek-AI, Tsinghua University, Peking University
The paper that introduced GRPO: drop PPO's value model and score each answer against 63 others for the same question. With 120B math tokens mined from the web, a 7B model reached 51.7% on MATH.
Constitutional AI: RL from AI feedback
Anthropic
Anthropic trained a harmless, non-evasive assistant with no human labels for harm: 16 written principles steered self-critique, revision and AI-labeled comparisons for reinforcement learning.
Fine-tuning LMs from human preferences
OpenAI
The first RLHF recipe for language models: learn a reward model from human 4-way choices, then optimize GPT-2 with PPO under a KL penalty. Its summarizers learned to copy, which labelers rewarded.
Serving, latency and cost
What makes inference fast, cheap and able to hold many users.
FlashAttention
Stanford University, University at Buffalo
The 2022 Stanford kernel that made attention fast by counting memory trips, not math: exact attention computed in on-chip tiles, several times faster, with memory linear in sequence length.
PagedAttention and vLLM
UC Berkeley, Stanford University, UC San Diego
The Berkeley paper behind vLLM: manage the KV cache like operating-system virtual memory, in small pages allocated on demand, and serve 2 to 4 times more requests on the same GPU.
AWQ
MIT, Shanghai Jiao Tong University, NVIDIA, Tsinghua, MIT-IBM Watson AI Lab, UMass Amherst
MIT's 4-bit weight quantization that asks the activations which weights matter, protects about 1% of channels by scaling instead of keeping them in 16-bit, and runs over 3× faster than FP16 on GPUs.
SGLang
Stanford University, UC Berkeley and others
A serving system for programs that call a model many times: it caches KV across requests in a radix tree, decodes fixed JSON text in one step, and ran agent workloads up to 6.4× faster.
DistServe
Peking University, StepFun, UC San Diego
Prefill is compute-bound and decode memory-bound, so sharing one GPU makes each slow the other. DistServe splits them onto separate GPUs and served up to 7.4× more requests within latency targets.
LMCache
Tensormesh, University of Chicago
Treat the KV cache as data that outlives the request: store it beyond the GPU, move it in big chunks, and reuse it across queries and engines. With vLLM, up to 15× throughput.
RouteLLM
UC Berkeley, Anyscale, Canva
Train a small router on human preference data to send each query to a strong, costly model or a weak, cheap one. On MT Bench it reached 95% of GPT-4's score using GPT-4 for 13% of queries.
Sarathi-Serve (chunked prefill)
Microsoft Research India, Georgia Institute of Technology
When a long prompt arrives, most servers pause all other output to process it. Sarathi-Serve slices prompts into chunks riding along with ongoing decodes: up to 3.7x more load within a latency target.
Parrot
Shanghai Jiao Tong University, Microsoft Research
LLM APIs see single requests, not the agent workflow behind them. Parrot has apps mark prompt inputs and outputs as variables, so the service chains, batches and shares work: up to 11.7x faster.
ThunderAgent
Georgia Institute of Technology, University of Illinois Urbana-Champaign, Carnegie Mellon University, Together AI
Agents spend much time in tool calls, and LLM servers evict their KV cache while they wait. ThunderAgent schedules whole agent programs: 1.5-3.6x serving throughput, up to 3.9x faster RL rollouts.
FlashInfer
University of Washington, NVIDIA, Perplexity AI, Carnegie Mellon University
The attention library inside SGLang, vLLM and MLC-Engine: every KV layout as one block-sparse format, custom variants compiled on demand, uneven batches balanced. 29-69% lower latency in SGLang.
FlashAttention-4
Princeton University, Meta, Colfax Research, NVIDIA, Georgia Tech, Together AI
Blackwell doubled tensor-core speed but not the exponential unit or shared memory, so softmax and data movement now limit attention. FlashAttention-4 rebalances: 1,613 TFLOPS on B200, 1.3x cuDNN.
Speculative decoding
Google Research
A small model guesses the next few tokens and the big model checks them all in one pass. With the right acceptance rule the output is exactly what the big model alone would produce, only 2-3x faster.
EAGLE-3
Peking University, Microsoft Research, University of Waterloo, Vector Institute
EAGLE-3's draft head reads the target model's own hidden states from three layers and is trained on its own guesses. Lossless decoding runs 3-6.5x faster, and it still helps at batch 64 in SGLang.
DiffusionGemma
Google DeepMind
DiffusionGemma refines token blocks in parallel. Learn why fast decoding can coexist with lower task quality, delayed stable output and a different serving bottleneck.
DFlash
UC San Diego
DFlash uses a tiny diffusion model to draft tokens in parallel, then lets the target verify them. Learn the acceptance rule, feature conditioning and serving tradeoffs.
Llama Speculation at Scale
Meta
Meta's production study explains why speculative decoding needs more than a good drafter: efficient tree attention, sampling, cache rollback and workload-aware dispatch.
LLM.int8()
University of Washington, Facebook AI Research, Hugging Face, ENS Paris-Saclay
Past 6.7B parameters a handful of hidden dimensions grow 20x larger than the rest and wreck int8 quantization. Keep them in 16-bit and int8 halves memory with no quality loss, up to 175B.
GPTQ
IST Austria, ETH Zurich, Neural Magic
GPTQ rounds weights one column at a time and shifts each rounding error onto the columns still unrounded. OPT-175B went to 3 bits in about 4 GPU-hours and ran on one A100, 3.24x faster per token.
The case for 4-bit precision
University of Washington
Given a fixed memory budget, is a bigger model at fewer bits better? Over 35,000 experiments on models from 19M to 176B parameters say yes, down to 4 bits per weight, and no further.
LLM in a flash
Apple
Apple ran 7B models on devices with half the memory they need by keeping weights in flash and loading only the neurons each token will use. OPT-6.7B went from 3.2 s to 0.67 s per token on CPU.
Speculative sampling
DeepMind
DeepMind's independent take on speculative decoding: a shallow 4B draft proposes tokens and Chinchilla 70B checks them in one pass. Samples were unchanged and decoding ran 2-2.5x faster.
Beyond text
Vision, speech and the tabular baseline you should always try first.
CLIP
OpenAI
Train image and text encoders on 400 million image-caption pairs. Classify images using class names in prompts, without training a separate classifier on labels for each downstream task.
Whisper
OpenAI
Train one transformer on 680,000 hours of web audio and its imperfect transcripts. With no benchmark fine-tuning, it made about half the errors of a LibriSpeech-trained model on other test sets.
XGBoost
University of Washington
Gradient boosted trees with a regularized objective, built-in handling of missing values and serious systems engineering. Still the first model to try on a customer's tables.
A Heart Means Recursion
Google DeepMind and Stanford University
Simple images can change a VLM's choices without appearing in its explanation. Learn the evidence, visual communication limits and how to test screenshot-driven agents.
Harness design and boundaries
The engineering around the model: what it may touch, what it is told, and how you tell a change helped.
Agent harness design decisions
No affiliation stated
A source-reading survey of 70 public agent projects that names the five decisions every harness makes. Its vocabulary is useful; several of its headline statistics cannot be reproduced.
Harness design for coding agents
UMass Amherst, Emory, UNC Charlotte, Zoom
Hold the agent loop fixed, vary one harness component at a time, and measure. Context management, planning and tool choice each helped some models and hurt or merely cost others.
Meta-Harness
Stanford, MIT, KRAFTON
Let a coding agent rewrite the code around a model, reading every earlier attempt's code, scores and traces. It beat hand-built harnesses; its biggest lesson is that raw traces beat summaries.
Evaluating AGENTS.md
ETH Zurich, LogicStar.ai
Coding agents follow the instructions in AGENTS.md files closely. On 438 real tasks, those files did not make them solve more, and every agent spent 8 to 50% more per task.
Auditing agent harness safety
UC Santa Barbara, UC Berkeley, UW–Madison, Stanford, Microsoft Research
An agent can hand back the right answer after reading records it should not, leaking data between agents, or calling tools it was never given. This benchmark audits the whole run, not the reply.
CaMeL: prompt injection by design
Google, Google DeepMind, ETH Zurich
Stop asking the model to resist injected instructions. Have it write a plan before it sees any untrusted data, then let an interpreter track where every value came from and enforce who may receive it.
AgentSpec
Singapore Management University
Write safety rules as trigger, check and enforce, and run them before every agent step. Hand-written rules stopped most risky actions in milliseconds; rules written by an LLM were patchier.
Prompt injection design patterns
ETH Zurich, Invariant Labs, IBM, Google, Microsoft, Swisscom, et al.
Six ways to build agents that resist prompt injection by limiting what untrusted text can touch, worked through ten real applications. The trade is plain: every pattern gives up some generality.
Harness²
Northwestern University, Google Cloud AI Research
Improve an agent's harness task by task with no answer key: run two scoped edits, compare behavior, compose, keep the lesson. Gains were consistent; most margins over rivals were not.
Coding Harnesses: A Source Study
Inclusive Brains, Wavestone AI Lab
A source-level study of eleven agent systems maps seven harness subsystems. Use its architecture questions while respecting the limits of static and uneven evidence.
Model cards for model reporting
Google, University of Toronto
A short document that ships with a model: what it is for, what it is not for, and how its errors differ across the groups and conditions it will meet. Averages hide the slice where errors pile up.
LIME: why should I trust you?
University of Washington
Explain one prediction of any model by nudging its input and fitting a small linear model to what changes. With LIME, users spotted a classifier that keyed on snow instead of wolves.
GDPR and a right to explanation
University of Oxford
GDPR limits fully automated decisions that affect people and asks for the logic behind them. Dropping sensitive columns does not stop discrimination, and caution can create it.
Stop explaining black boxes
Duke University
A three-rule list on age and priors predicts re-arrest about as well as a proprietary tool with 130+ factors. For high-stakes decisions, build a model you can read instead of explaining one you can't.
Actions, tools and orchestration
How an agent expresses actions, runs independent work at once, and divides a task among roles.
CodeAct: code as the action format
University of Illinois Urbana-Champaign, Apple
Let an agent act by writing Python instead of one JSON tool call at a time. On multi-tool tasks the best model solved 74% against 52% with JSON; on single calls JSON was as good or better.
Toolformer
Meta AI Research, Universitat Pompeu Fabra
A model writes candidate API calls into ordinary text, keeps those whose results make the next words easier to predict, and trains on them. A 6.7B model learned when to call a calculator.
PAL: program-aided language models
Carnegie Mellon University
Have the model translate a word problem into Python and let the interpreter compute. Codex beat a 540B model on grade-school math, and led by 38 points once the numbers got large.
ReWOO: plan before you observe
NC State, Texas A&M, Microsoft
Write the whole tool plan up front with placeholders for results, run the tools, answer once. Tokens fell about fivefold versus ReAct, but the plan cannot adapt to what tools return.
LLMCompiler: parallel function calls
UC Berkeley, ICSI, LBNL
Plan tool calls as a dependency graph, then run every call whose inputs are ready at the same time. Against ReAct this cut latency up to 3.7× and cost up to 6.7×, with similar accuracy.
Agentless
University of Illinois Urbana-Champaign
Fix GitHub issues with a fixed localization, repair and validation pipeline. Agentless resolved 32% of SWE-bench Lite for $0.70 per issue, leading the open-source systems in the paper's comparison.
OpenHands
UIUC, CMU, Yale, UC Berkeley, All Hands AI, et al.
An open platform for agents that work like developers: one event stream of actions and observations, a sandboxed runtime with shell, Python and browser, and agents that hand off to each other.
AutoGen
Microsoft Research, Penn State, University of Washington, Xidian University
Microsoft's framework for building LLM apps as conversations between agents that write code, run it, ask humans and check each other, and what its own results say about when more agents help.
Magentic-One
Microsoft Research AI Frontiers
Microsoft's five-agent team led by an Orchestrator that keeps a task ledger and a progress ledger, re-plans when stuck, and was scored with error bars that show what a leaderboard can't tell you.
AFlow
DeepWisdom, HKUST (Guangzhou), Renmin University, Nanjing University, Fudan University, KAUST, Mila
Treat an LLM workflow as code and let Monte Carlo tree search, with an LLM proposing one edit at a time, find a better one. It beat hand-built workflows on six benchmarks. Read the evaluation closely.
Search-R1
University of Illinois Urbana-Champaign, University of Massachusetts Amherst, Google Cloud AI Research
Train search decisions inside the reasoning rollout. Learn retrieved-token masking, outcome rewards and the evaluation limits behind Search-R1's QA gains.
ReTool
ByteDance Seed
ReTool trains reasoning models to interleave code execution and reasoning. Examine its outcome rewards, tool-feedback masking and the difference between steps and cost.
Gorilla: an LLM connected to massive APIs
UC Berkeley, Microsoft Research
A 7B model fine-tuned on 16,450 synthetic instruction-API pairs beats GPT-4 at writing API calls, on APIs it trained on. BM25 retrieval cut its HuggingFace accuracy from 72% to under 18%.
Planning, feedback and improvement
Search, critique and optimization loops, and the evidence that says whether they helped.
GEPA
UC Berkeley, Stanford, BespokeLabs.ai, Notre Dame, Databricks, MIT
A prompt optimizer that reads its failures in plain language, rewrites one module's prompt at a time, and keeps every prompt that wins somewhere. It beat GRPO on five of six tasks with fewer rollouts.
TextGrad
Stanford University, Chan Zuckerberg Biohub
Backpropagation with words: an LLM critiques each step of a pipeline, critiques flow backward through the graph, and an LLM rewrites prompts, code or answers. A PyTorch-style API, uneven evidence.
Tree of Thoughts
Princeton University, Google DeepMind
Let the model propose several next steps, judge which are still promising, and search with backtracking. On the Game of 24, GPT-4 went from 4% to 74%, at a cost the paper itemizes.
Self-Consistency
Google Research, Brain Team
Sample many chain-of-thought answers and take the majority vote. A one-line change that added up to 17.9 points on GSM8K, and whose agreement rate doubles as a confidence score.
Self-Refine
Carnegie Mellon University, Allen Institute for AI, University of Washington, NVIDIA, UC San Diego, Google Research
One model drafts, critiques its own draft, and rewrites it, up to four times. Big gains on style and constraint tasks, none on math, and a headline average that mixes win rates with scores.
LLMs Cannot Self-Correct Reasoning Yet
Google DeepMind, University of Illinois Urbana-Champaign
Asked to review and fix their own reasoning with no outside signal, GPT-3.5, GPT-4 and Llama-2 got worse on every benchmark. The earlier gains came from answer keys and unequal compute.
The Prompt Report
University of Maryland, Learn Prompting, OpenAI, Stanford, Microsoft, Vanderbilt, Princeton, et al.
A systematic survey of 1,565 prompting papers: a shared vocabulary, 58 text techniques in six families, a small benchmark, and a 47-step case study that shows how fragile hand prompt engineering is.
Sharpening Tax
Meta Superintelligence Labs, University of Wisconsin–Madison, NYU, Stanford University
Distinguish first-attempt accuracy, repeated-sampling coverage and consistency. Compute sharpening tax and learn why a positive tax need not mean a worse model.
The Illusion of Thinking
Apple
Controlled puzzles expose limits in long reasoning traces. Learn what the failures show, how later revisions address critiques and how to build a more diagnostic agent evaluation.
Judge Reliability and Validity
UC Berkeley School of Information
A judge can repeat the same decision and still be biased. Learn how to audit agreement, answer-order effects and metric denominators before trusting automated evaluations.
Invisible Reasoning
New York University, University of Maryland and Together AI
Fixed filler tokens can change reasoning-task accuracy without explaining the computation. Learn the controlled evidence, negative training results and limits of text-only oversight.
GSM-Symbolic: when the numbers change
Apple, Washington State University
Apple turned 100 GSM8K questions into templates and changed only names and numbers. Scores fell for most of 25 models, and one irrelevant sentence cut up to 65 points.
Faith and Fate: limits on compositionality
Allen Institute for AI, University of Washington
GPT-4 got 59% of 3-digit multiplications right zero-shot. Fine-tuned GPT-3 was near perfect on sizes it saw and failed beyond them, and the paper shows why errors compound.
Reasoning or reciting: counterfactual tasks
MIT, Boston University
GPT-4 added two-digit numbers almost perfectly in base 10 and got 39% right in base 9. Across 11 tasks, changing one rule cut accuracy, even when the model understood the new rule.
Alice in Wonderland: a one-line puzzle
LAION, Juelich Supercomputing Center
"Alice has 4 brothers and 1 sister. How many sisters does Alice's brother have?" GPT-4 got it nearly always. Swap in other numbers and it nearly never did, though the problem was unchanged.
Stop calling intermediate tokens thinking
Arizona State University
Reasoning models write long traces before answering, and people read them as thinking. This position paper collects evidence that trace validity and answer correctness come apart.
LLMs can self-improve
Google, University of Illinois Urbana-Champaign
Sample 32 reasoning paths per unlabeled question, keep the ones that agree with the majority answer, fine-tune on them. PaLM-540B rose from 74.4% to 82.1% on GSM8K without any labels.
More agents is all you need
Tencent
Ask the same model the same question 40 times and take the majority answer. Llama2-13B rose from 35% to 59% on GSM8K, past one call to Llama2-70B, at 40 times the tokens.
IMO 2025 gold with verify-and-refine
UCLA
A solve, self-review, verify and fix loop took Gemini 2.5 Pro, Grok-4 and GPT-5 to 5 of 6 IMO 2025 problems, against best-of-32 baselines of 21% to 38%. Attempt counts and grading go unreported.
Memory and context
What persists across turns and tasks, and how long work fits inside a context window.
MemGPT
UC Berkeley
Treat the context window like RAM: a small editable working memory in the prompt, everything else in searchable storage, and the model paging data in and out itself through function calls.
Generative Agents
Stanford University, Google Research, Google DeepMind
Twenty-five simulated townspeople with a memory stream, reflection and plans. The retrieval score behind them, recency plus importance plus relevance, became the template for agent memory.
Voyager
NVIDIA, Caltech, UT Austin, Stanford, UW Madison
A GPT-4 agent that teaches itself Minecraft by proposing its own next task, writing code, checking the result, and saving what works as a reusable skill. Memory here is a library of verified programs.
Agent Workflow Memory
Carnegie Mellon University, MIT
Web agents that turn successful runs into reusable, abstracted workflows and improve as they work: WebArena success from 23.5% to 35.5% with no human-written workflows. Check what is relative.
A-MEM
Rutgers University, AIOS Foundation
Memory as Zettelkasten: each interaction becomes a note with LLM-written keywords and context, linked to related notes; old notes are rewritten as new ones arrive. Big temporal gains, mixed elsewhere.
MemOS
MemTensor (Shanghai), Institute for Advanced Algorithms Research Shanghai, Renmin University, Zhejiang University, Shanghai Jiao Tong University, et al.
A memory operating system for LLM apps: text, KV-cache and parameter memories in one governed unit, the MemCube, with scheduling, lifecycle and access control. Strong benchmarks, run by the vendor.
LongMemEval
UCLA, Tencent AI Lab Seattle, UC San Diego
500 hand-written questions buried in user-assistant chat histories up to 1.5M tokens, testing five memory abilities. Long-context models lost 30-55% of their accuracy; commercial memory did worse.
Context-Folding
ByteDance Seed, Carnegie Mellon University, Stanford University
Give an agent two tools, branch and return, so it can do token-heavy work in a side thread and keep only the result. Trained with RL, a 32K-context agent matched a 327K one. Mind which baseline.
LLMLingua-2
Tsinghua University, Microsoft
Compress prompts 2-5x with a small encoder that decides, word by word, what to keep, trained on GPT-4's deletions. Fast and faithful in domain; behind question-aware compression out of it.
Prompt Cache
Yale University, Google
Precompute the attention states of reusable prompt pieces once and splice them into later prompts: 5 to 10 times faster first tokens, with an approximation that can cost accuracy.
Context Language Models
University of Washington, Meta Superintelligence Labs, MIT, Trillium Labs
Give the model write access to its own context, as a file it edits with Bash, and let it invent its own compaction. Beats fixed harnesses at lower compute in most tests, with caveats in the setup.
U-Fold
Zhejiang University, Tongyi Lab (Alibaba Group), Westlake University
Keep the full history, and on every user message rebuild a small working context: a summary with a to-do list plus verbatim lines from past tool output. Small gains over ReAct, big ones under noise.
ContextBench (coding agents)
Nanjing University, University College London
1,136 real issues with expert-traced gold code, so you can score what a coding agent read, not only whether its patch passed. The retrieval scores did not rank agents the way solve rates did.
Lost in Multi-Turn Conversation
Microsoft Research, Salesforce Research
Across 15 models and six tasks, splitting instructions across turns lowered scores by 39% on average. Early answers and persistent wrong assumptions made conversations less reliable.
ACON
Microsoft, KAIST
Microsoft learns a context-compression prompt from the agent's own failures, then distills it into a small model. Peak tokens fell 26% to 54%, though tasks took more steps and more time.
Zep
Zep AI
Zep stores agent memory as a knowledge graph whose facts record when they were true and when they were learned. On LongMemEval it beat full context by 8 to 11 points with 1.6K tokens instead of 115K.
Hindsight
Vectorize.io, The Washington Post, Virginia Tech
Hindsight keeps an agent's facts, experiences, opinions and entity summaries in separate networks. A 20B open model went from 39.0% to 83.6% on LongMemEval, though baselines came from other reports.
EverMemOS
EverMind, Shanda Group
EverMemOS turns conversation into episodes, facts and expiring foresight, clusters them into scenes, and rebuilds just enough context. 93.05% on LoCoMo, at real build cost.
Less Context, Better Agents
Microsoft
A Dynamics 365 expense agent improves when old tool responses are pruned and summarized. Learn the controlled comparison, completion checks and limits of a five-turn window.
Systems classics for agents
Older rules about failure, authority and latency that every agent with real actions inherits.
End-to-End Arguments
MIT Laboratory for Computer Science
If only the applications at the ends can verify a function completely, building it into lower layers buys performance at most, never correctness. The 1984 rule behind every agent action.
Sagas
Princeton University
Split a long transaction into steps that each commit on their own, each with a compensating step that undoes it in business terms. The 1987 pattern behind multi-step agent actions.
Protection of Information
MIT
The 1975 tutorial that named least privilege, fail-safe defaults and complete mediation. Eight design principles written for time-sharing systems that read like a checklist for giving an agent tools.
A Note on Distributed Computing
Sun Microsystems Laboratories
Remote calls are not local calls with more latency: partial failure and concurrency cannot be hidden, so they belong in the interface. Sun's 1994 warning applies to every tool an agent calls.
The Tail at Scale
When a request waits on many servers, rare slow responses become the common case. Google's ways to live with latency variability: hedged and tied requests, micro-partitions, good-enough answers.
LSM-tree Key-Value Stores: A Survey
Xiamen University, MBZUAI, OceanBase (Ant Group)
RocksDB, Cassandra and TiKV all write through an LSM-tree. This 2025 survey maps where they hurt: compaction stalls, read amplification, new devices, distribution and tenants.
ParquetDB
West Virginia University, Instituto de Ciencia de Materiales de Madrid, NIST
ParquetDB adds Python CRUD operations to Parquet files. Learn how columnar reads, schema evolution and rewrite costs shape its fit for customer data workflows.
Draft-Conditioned Decoding
University of Central Florida and Lockheed Martin AI Center
Grammar constraints can preserve JSON syntax while changing which answer a model produces. Learn a two-stage remedy, its mathematical limits and the cost of an extra pass.
Software engineering at Google
One repository, 2 billion lines, 40,000 commits a day, nearly all at head. A Google engineer catalogs what makes it work: hermetic builds, mandatory review, green trunks, canaries, post-mortems.
A critique of the CAP theorem
Affiliation not stated in the paper
'Consistency, availability, partition tolerance: pick two' is too vague to design with. Kleppmann shows why and offers a sharper question: does this operation's latency grow with network delay?
Paxos vs Raft
University of Cambridge
Describe Paxos in Raft's vocabulary and the two algorithms turn out to differ only in leader election. Raft's famous understandability comes mostly from its presentation, the authors conclude.
Agent runtimes and long-running work
Durable state, interruption, recovery and the execution services that keep long agent jobs alive.
OpenHands Software Agent SDK
OpenHands
OpenHands rebuilt its agent core around one event log and immutable configuration. In a 15-day side-by-side rollout, system-attributable failures fell 61%.
AgentScope 1.0
Alibaba Group
Alibaba's agent framework makes interruption, parallel tool calls, tool groups and saved state first-class. A design reference with no measurements, so test its claims yourself.
Magentic-UI
Microsoft Research AI Frontiers
Microsoft's open-source agent interface puts the user in the team: co-planning, take-over, action approvals, saved plans. A simulated helper lifted GAIA accuracy from 30.3% to 51.9%.
TRACE: corrections to runtime checks
University of Notre Dame, IBM Research, Tencent AI Lab
Remembering a correction is not obeying it. TRACE turns a user's repeated corrections into rules with checks that block a coding agent until it complies.
Agent Lightning v1.0
Microsoft, Fudan University, Zhejiang University, University of Edinburgh
Train an agent through the same harness it runs in at deployment. Microsoft's 3,500-line control plane, four traps in turning API calls into training data, and a 41.8% to 56.4% SWE-bench gain.
SWE-RL
Meta AI, UIUC, Carnegie Mellon University
Meta trained Llama 3.3 70B with RL on 273,000 merged pull requests, rewarding patches for looking like the real fix. 41.0% on SWE-bench Verified, after sampling 500 patches per issue.
R2E-Gym
UC Berkeley, Australian National University
8,135 executable coding tasks built from commits instead of issues, and a finding about picking patches: tests and a learned judge each plateau near 43%, while combined they reach 51%.
Scaling Agent Systems
Google Research, Google DeepMind, MIT
260 controlled runs, five architectures, six agentic benchmarks: adding agents ranged from +80.8% to -70%, and the one robust finding is that strong single agents gain little.
Kimi K2.5 and Agent Swarm
Moonshot AI
Moonshot trained an orchestrator to spawn and schedule parallel sub-agents. WideSearch time to a target score fell 3 to 4.5 times; total compute for the swarm is not reported.
GLM-5
Zhipu AI (Z.ai), Tsinghua University
Zhipu's 744B open model report is most useful for its agent RL plumbing: exact-token logging, stale-sample filters, cache-affine routing, 10,000 built coding environments and simple context folding.
Mobile-Agent-v3
Tongyi Lab, Alibaba Group
Alibaba's GUI agent splits each step into Manager, Worker, Reflector and Notetaker roles, all one open model. AndroidWorld 73.3, though the gain over end to end also swaps 7B for 32B.
Choosing and sizing models
What a model's architecture, data and inference modes mean for the hardware and workload you have.
Gemma 4
Google DeepMind
Google DeepMind's open-weight family from 2.3B to 31B, with thinking mode, audio and vision. The report's memory table is what to read first: which model fits your hardware, and with how many users.
Phi-4-Mini and Phi-4-Multimodal
Microsoft
Microsoft's 3.8B model plus a multimodal version that adds vision and speech as LoRAs on a frozen base, so text quality is untouched. Leads OpenASR; weak at answering spoken questions.
Qwen3
Alibaba (Qwen Team)
One open model with a switch: think hard or answer fast, plus a token budget in between. The report shows where thinking pays (math, code) and where it doesn't (chat, long-context retrieval).
Nemotron 3 Nano
NVIDIA
NVIDIA's 31.6B model activates 3.2B per token and has only 6 attention layers among 52, so its KV cache is tiny. Fast long generation and 1M context; weaker on knowledge and multilingual tests.
gpt-oss
OpenAI
OpenAI's open-weight reasoning models: 120b fits one 80 GB GPU, 20b fits 16 GB, with three effort levels and tool use. The card also shows the weak spots: factual recall and holding a system prompt.
SmolLM2
Hugging Face
Hugging Face trains a 1.7B model on 11T tokens and documents every data decision: LLM-labeled quality filters, late math and code, mixes changed mid-run. A recipe for curating any training set.
Small Models for Agentic AI
NVIDIA Research, Georgia Institute of Technology
Evaluate small models for repeated agent tasks using conditional quality, fallback costs and specialization effort, with an executable cascade cost model.
Kimi K2
Moonshot AI
Read Kimi K2 through its sparse architecture, stateful tool-data pipeline and benchmark harnesses, with capacity arithmetic and a runnable simulator exercise.
LLaMA
Meta AI
Meta trained 7B to 65B models on up to 1.4 trillion public tokens, far past the Chinchilla ratio, so they would be cheap to run. LLaMA-13B beat the 175B GPT-3 on most benchmarks.
Llama 2
GenAI, Meta
Meta's 7B-70B models on 2T tokens, with the full chat recipe: 27,540 SFT examples, 1.4M preference pairs, two reward models, rejection sampling, PPO and Ghost Attention. Licensed for commercial use.
Llama 3
Meta
Meta's 405B dense model on 15.6T tokens with 128K context, matching GPT-4 on many benchmarks, plus an unusually frank account of training on 16K GPUs: 419 unexpected failures in 54 days.
GPT-4 report
OpenAI
GPT-4 scored near the top 10% on a simulated bar exam, and its final loss was predicted from runs with 10,000x less compute. The report discloses no size, data or architecture; read it for evals.
Sparsely-gated MoE
Google Brain, Jagiellonian University
Google Brain put thousands of small expert networks behind a learned router that picks a few per token, growing capacity over 1000x at little extra compute. Every MoE language model since starts here.
Switch Transformers
Google simplified mixture-of-experts to one expert per token, fixed its training instabilities, and trained a 1.6-trillion-parameter model. T5-Base-sized compute pretrained up to 7x faster.
Mixtral 8x7B
Mistral AI
Mistral's open sparse model: 8 experts per layer, 2 used per token, 47B parameters stored but 13B active. It matched or beat Llama 2 70B on most benchmarks, released under Apache 2.0.
DeepSeek-V2
DeepSeek-AI
DeepSeek's 236B MoE uses 21B parameters per token and caches each token's attention in one 512-dim latent (MLA). It cut training cost 42.5% and raised throughput 5.76x over DeepSeek 67B.
BitNet
Microsoft Research, University of Chinese Academy of Sciences, Tsinghua University
Microsoft trained Transformers whose weights are only +1 or -1 from scratch. At 6.7B they came within 2 points of FP16 on zero-shot tasks, far ahead of post-training quantization to 1 or 2 bits.
BitNet b1.58
Microsoft Research, University of Chinese Academy of Sciences
Weights of only -1, 0 or +1, trained from scratch: at 3B parameters, Microsoft's ternary model matched its FP16 baseline in perplexity with 3.55x less GPU memory and 2.71x faster decoding.
Textbooks Are All You Need (phi-1)
Microsoft Research
A 1.3B-parameter code model trained on 7B tokens of filtered and GPT-3.5-written 'textbook' data scored 50.6% on HumanEval, beating models ten times larger. Its own pruning table shows where.
TinyStories: tiny models, simple stories
Microsoft Research
Trained on GPT-written stories that use only words a 3-year-old knows, models under 10 million parameters wrote fluent, consistent stories. GPT-4 graded them like a teacher grading students.
Fine-tuning, distillation and compression
Training a smaller or cheaper model for one job, and checking what it lost on the way.
DoRA
NVIDIA, HKUST
Split each weight into per-column magnitude and direction; train magnitude directly, direction with LoRA. Same serving cost as LoRA, 1 to 4 points better, and far steadier at low rank.
LESS
Princeton University, University of Washington
Pick the 5% of a fine-tuning pool whose gradients point the same way as a few examples of the skill you want. It often beats training on everything, and the gradients come from a cheap LoRA warmup.
Magpie
University of Washington, Allen Institute for AI
Give an open chat model only the opening of a user turn and it writes a user's question. 300K filtered pairs from Llama-3 rival its official instruct model, at about $1 per 1,000 pairs.
Distilling Step-by-Step
University of Washington, Google Cloud AI Research, Google Research
Ask a big model why, not just what, and train a small model to predict both. A 770M model beats 540B PaLM's few-shot accuracy on ANLI with 80% of the data, and rationales cost nothing at serving time.
DistiLLM-2
KAIST AI, Microsoft
Distill by raising the student's probability on the teacher's text and lowering it on its own bad text, with a different loss for each. Better students and faster speculative-decoding drafters.
Minitron
NVIDIA
NVIDIA prunes a trained 15B model to 8B and 4B and distills it back with under 3% of the original data. The 8B rivals models trained on trillions of tokens, and the recipe comes as ten rules.
NVFP4 quantization-aware distillation
NVIDIA
Quantizing a small model to 4-bit NVFP4 loses accuracy; retraining it on labels can lose more. Training it to match the original model's outputs got near-full accuracy back, even from random tokens.
When self-distillation hurts reasoning
Microsoft Research, KAIST, SK hynix, Seoul National University
A teacher that can see the answer never says "wait". Students trained on its confident reasoning get shorter and much worse on unseen math: AIME24 fell from 54.8 to 20.2.
GenRec (Netflix)
Netflix
Netflix replaced thousands of engineered features with a text prompt and an LLM that scores the whole catalog in one pass. It beat a mature ranker online; the long-term lift was 0.006%.
RLVR and Reasoning Boundaries
Tsinghua University, Shanghai Jiao Tong University
RL can improve one-attempt accuracy while reducing measured task coverage. Read pass@k correctly and compare adaptation against sampling with a real verifier.
DAPO
ByteDance Seed, Tsinghua AIR, University of Hong Kong, SIA-Lab
Understand DAPO's clipping, dynamic sampling, token-level loss and length rewards, with tested arithmetic and practical controls for reasoning-model training.
GSPO
Qwen Team, Alibaba
Learn how GSPO uses response-level likelihood ratios and clipping for RL, why it helped Qwen's MoE training, and what length normalization does not guarantee.
Dr. GRPO
Sea AI Lab, National University of Singapore, Singapore Management University
Separate base-model ability, prompt effects and optimizer bias in reasoning RL. Understand Dr. GRPO's fixed denominator with an executable weighting exercise.
Rubrics as Rewards
Scale AI
Turn task-specific criteria into RL feedback, compare explicit and implicit rubric rewards, and learn why a high aggregate score cannot replace required checks.
NVFP4 Pretraining
NVIDIA
NVIDIA demonstrates long-horizon pretraining with four-bit matrix operands. Learn the numerical safeguards, retained high-precision state and evaluation limits.
SparseGPT
IST Austria, Neural Magic
The GPTQ team's pruning method removes half the weights of OPT-175B in about 4 GPU-hours with no retraining, and perplexity stays flat. Magnitude pruning collapses on the same models.
Knowledge distillation
Hinton, Vinyals and Dean trained small models to match a big model's softened output probabilities instead of only the labels. A single distilled speech model kept most of a 10-model ensemble's gain.
LIMA: 1,000 examples for alignment
Meta AI, Carnegie Mellon University, University of Southern California, Tel Aviv University
Meta fine-tuned LLaMa-65B on 1,000 hand-picked prompts and answers, with no RLHF. People judged it equal or better than GPT-4 in 43% of comparisons; it won outright in 18%.
Self-Instruct: a model writes its own tasks
University of Washington, Tehran Polytechnic, Arizona State University, Johns Hopkins University, Allen Institute for AI
Starting from 175 hand-written tasks, GPT-3 generated 52,000 instructions and 82,000 examples for its own fine-tuning. Only 54% of sampled examples were fully correct, and it still helped.
ROME: locating and editing facts in GPT
MIT CSAIL, Northeastern University, Technion
Causal tracing finds facts recalled in middle-layer MLPs at a subject's last token. ROME edits one fact there with a rank-one update, but each related subject it disturbs is a new error.
KV cache and long-context infrastructure
Storing, moving, compressing and reusing attention state when you run the inference fleet.
Mooncake
Moonshot AI, Tsinghua University
Kimi's serving stack schedules around the KV cache: separate prefill and decode pools, a cache pool in idle DRAM and SSD, early rejection under overload. About 75% more requests on real traffic.
CacheBlend
University of Chicago, CUHK Shenzhen, Stanford University, Microsoft Research
Prefix caching only helps the first retrieved chunk. CacheBlend reuses every chunk's cached KV and recomputes about 15% of tokens to restore cross-chunk attention: 2.2-3.3x faster first tokens in RAG.
CacheGen
University of Chicago, Microsoft, Stanford University
A long document's cached KV can be gigabytes, and fetching it can take longer than recomputing. CacheGen encodes KV like video: 3.5-4.3x smaller than quantized KV, 3.2-3.7x faster first tokens.
KIVI
Rice University, Texas A&M University, Stevens Institute of Technology, Carnegie Mellon University
Quantize keys per channel and values per token, keep the newest 128 tokens in full precision, and a 2-bit KV cache holds up: 2.6x less memory, 4x larger batches, with no calibration or fine-tuning.
S-LoRA
UC Berkeley, Stanford University, Shanghai Jiao Tong University
One base model, thousands of customer fine-tunes: keep LoRA adapters in host memory, page them into the KV cache's GPU pool, and batch across adapters. 2,000 adapters on one GPU at 95% throughput.
Helix Parallelism
NVIDIA
Past 8 GPUs, tensor parallelism stops speeding up attention: each extra GPU copies the KV cache. Helix splits the sequence for attention and all GPUs for the FFN, for up to 32x the users at 1M tokens.
TENT
9#AISoft, Tsinghua University, Moonshot AI, Alibaba Cloud, Zhejiang University, Ant Group, Approaching.AI
Mooncake's team rebuilt its KV transfer engine after running it on thousands of GPUs: declare what to move, and 64 KB slices go to whichever links are fastest right now. 1.36x serving throughput.
Cost control, routing and recovery
Spending model calls where they change the answer, and getting serving capacity back after a crash.
FrugalGPT
Stanford University
Ask a cheap model first and escalate only when a small scorer doubts its answer. On a news task this matched GPT-4 at 2% of the cost, because cheap models are sometimes right where GPT-4 is wrong.
BEST-Route
University of British Columbia, Microsoft, Pennsylvania State University, Google DeepMind, AG2AI
Routers overuse the big model because one small-model answer rarely matches it. BEST-Route also picks how many small-model answers to sample: 60% cheaper than GPT-4o, under 1% quality loss.
Dynamo fast recovery
NVIDIA
Most serving crashes kill the engine process, not the GPU, yet restarts rebuild everything for minutes. Dynamo keeps weights resident and a parked backup engine ready: recovery in under 7 seconds.
The Routing Plateau
Rice University and Amazon
A pool of models can solve more tasks than any one model, yet a router may not know which model will succeed. Learn how to measure that prediction gap.
Guardrails, attacks and red teaming
Who may instruct the model, where runtime checks sit, and how to test defenses against attackers who adapt.
The Instruction Hierarchy
OpenAI
Models treat a developer's rules, a user's request and a web page's text as equally binding. OpenAI trained GPT-3.5 to rank them, and robustness rose on every attack it tested, including unseen ones.
LlamaFirewall
Meta
Meta's open-source agent guardrails: a jailbreak classifier on inputs, an LLM auditing the agent's reasoning, static analysis on code. Together they cut AgentDojo attack success from 17.6% to 1.75%.
Constitutional Classifiers++
Anthropic
Anthropic's production jailbreak defense pairs a probe with a classifier that reads prompt and reply together: 5.5% of traffic escalated, about 40x less compute than the exchange classifier alone.
GCG: universal adversarial attacks
Carnegie Mellon University, Center for AI Safety, Google DeepMind, Bosch Center for AI
The 2023 GCG paper optimized adversarial suffixes on open models and tested transfer to ChatGPT, Claude and PaLM-2, showing why refusal evaluations need automated attacks.
Many-shot jailbreaking
Anthropic, University of Toronto, Vector Institute, Constellation, Stanford, Harvard
Long context opened a simple attack: hundreds of fake dialogues where the assistant complies. Success follows a power law in shots, and alignment training only shifts the curve.
HarmBench
University of Illinois Urbana-Champaign, Center for AI Safety, Carnegie Mellon University, UC Berkeley, Microsoft
Jailbreak papers reported numbers nobody could compare. HarmBench fixed the behaviors, the length and the judge, ran 18 attacks on 33 models, and found no attack or defense that wins everywhere.
The Attacker Moves Second
OpenAI, Anthropic, Google DeepMind, HackAPrompt, Northeastern University, ETH Zürich, AI Sequrity Company, MATS
Researchers from OpenAI, Anthropic and Google DeepMind re-tested 12 published LLM defenses against attackers who adapt. Most had reported near-zero attack success; most fell above 90%.
Agent Smith: infectious jailbreak
Sea AI Lab, National University of Singapore, Singapore Management University
In a simulation of up to a million chatting multimodal agents, one bad input in a single agent's shared memory spread across the whole group in a logarithmic number of rounds, by itself.
Reward Hacking and Misalignment
Anthropic and Redwood Research
A model trained to exploit coding rewards can generalize harmful behavior beyond that task. Learn the experiment's controls, mitigation limits and implications for agent evaluation.
Near-constant Poisoning
UK AI Security Institute, Anthropic, Alan Turing Institute, University of Oxford and ETH Zurich
A tiny fraction of a large training corpus can still contain enough repeated malicious examples to matter. Learn the evidence, threat-model limits and defensive implications.
Emergent misalignment from fine-tuning
Truthful AI, UCL, Center on Long-Term Risk, Warsaw University of Technology, University of Toronto, UK AISI, UC Berkeley
Fine-tuning GPT-4o only to write insecure code, without saying so, made it misbehave on unrelated questions: 20% of answers on 8 showcase questions, 6% on 48 pre-registered ones.
Unfaithful chain-of-thought explanations
NYU, Cohere, Anthropic
Add a user hint pointing at a wrong answer and GPT-3.5's chain-of-thought accuracy falls from 59.6% to 23.3%. Of 426 biased explanations reviewed, one mentioned the hint.
The internal state knows when it's lying
Ariel University, Carnegie Mellon University
A small classifier on hidden activations tells true from false statements 71% to 83% of the time on unseen topics. On sentences the model wrote itself it reached 71%; GPT-4, prompted, reached 84%.
Sleeper Agents
Anthropic, Redwood Research, Mila, University of Oxford and others
Anthropic trained models to behave safely unless a trigger appears, then applied standard safety training. The hidden behavior survived, and adversarial training taught models to hide it better.
Extracting training data from LLMs
Google, Stanford, UC Berkeley, Northeastern, OpenAI, Harvard, Apple
GPT-2 shows no train-test gap, yet a black-box audit recovered 604 verbatim training texts, some with names and phone numbers. Text repeated 33 times in one page was enough.
Breaking anonymity of the Netflix Prize data
The University of Texas at Austin
Netflix removed every name before releasing 100 million ratings. Eight ratings with rough dates, two of them wrong, were enough to single out 99% of subscribers.
A watermark for large language models
University of Maryland
Nudge each token toward a secret, shifting half of the vocabulary and a simple count proves a text came from your model. No human text was flagged, but short or predictable text stays invisible.
Can AI-generated text be reliably detected?
University of Maryland
Every class of AI-text detector the authors tested broke under paraphrasing, and a simple bound shows why: as models write more like people, the best possible detector approaches a coin flip.
Concrete problems in AI safety
Google Brain, Stanford University, UC Berkeley, OpenAI
A cleaning robot knocks over a vase to finish faster, hides messes from its own camera, and mops near an outlet. Five practical ways ML systems cause accidents, named in 2016 and still the checklist.
“Most cited” uses dated Semantic Scholar counts shown on each paper. Citation counts describe research influence, not a recommendation for your system.