# Papers: notable data and AI research papers, by lab

> 40 notable data and AI research papers from 15 labs and companies (OpenAI, Anthropic, Google, Google DeepMind, Meta, Microsoft, NVIDIA, Databricks, Snowflake, DeepSeek, Alibaba, Ai2 and universities), each with one plain-language line on what it shows, by fru.dev.

For each paper: title, lab (with its companies.fru.dev registry slug), first arXiv date, topic, a one-line summary, the arXiv link, citations and influential citations from Semantic Scholar, and upvotes from Hugging Face papers. Seed papers were summarised by hand from their abstracts; new papers are found weekly on Hugging Face daily papers (organisation is a tracked lab, 25+ upvotes) and listed as Unverified until checked. Counts refresh weekly (Monday 12:15 UTC) and every change is kept in an append-only history. Last successful refresh: 2026-09-24 10:27:04 UTC.

Disclaimer: One-line summaries for orientation, not a substitute for the paper. Citation and upvote counts from Semantic Scholar and Hugging Face. Logos via logo.dev; trademarks belong to their owners. Corrections: use Suggest a correction on any page.

## Pages

- [Home: papers by lab on a timeline, and the latest papers](https://papers.fru.dev/)
- [All papers](https://papers.fru.dev/papers)
- [Labs and companies](https://papers.fru.dev/labs)
- [Topics](https://papers.fru.dev/topics)
- [Changes: citation, upvote and review history](https://papers.fru.dev/changes)
- [Sources and method](https://papers.fru.dev/method)
- [OpenAI: 5 papers](https://papers.fru.dev/labs/openai)
- [Google: 4 papers](https://papers.fru.dev/labs/google)
- [Meta: 4 papers](https://papers.fru.dev/labs/meta)
- [Microsoft: 4 papers](https://papers.fru.dev/labs/microsoft)
- [Anthropic: 4 papers](https://papers.fru.dev/labs/anthropic)
- [Google DeepMind: 3 papers](https://papers.fru.dev/labs/deepmind)
- [NVIDIA: 3 papers](https://papers.fru.dev/labs/nvidia)
- [Snowflake: 3 papers](https://papers.fru.dev/labs/snowflake)
- [Stanford University: 2 papers](https://papers.fru.dev/labs/stanford)
- [UC Berkeley: 2 papers](https://papers.fru.dev/labs/uc-berkeley)
- [DeepSeek: 2 papers](https://papers.fru.dev/labs/deepseek)
- [Carnegie Mellon University: 1 papers](https://papers.fru.dev/labs/carnegie-mellon)
- [Databricks: 1 papers](https://papers.fru.dev/labs/databricks)
- [Ai2: 1 papers](https://papers.fru.dev/labs/ai2)
- [Alibaba (Qwen): 1 papers](https://papers.fru.dev/labs/alibaba)
- [Agents and evaluation](https://papers.fru.dev/topics/agents-evals)
- [Alignment and safety](https://papers.fru.dev/topics/alignment-safety)
- [Reasoning](https://papers.fru.dev/topics/reasoning)
- [Inference and efficiency](https://papers.fru.dev/topics/efficiency)
- [Retrieval and data](https://papers.fru.dev/topics/retrieval-data)
- [Multimodal and robotics](https://papers.fru.dev/topics/multimodal)
- [Architectures](https://papers.fru.dev/topics/architectures)
- [Training and scaling](https://papers.fru.dev/topics/training)
- [Foundation models](https://papers.fru.dev/topics/foundation-models)
- [Applied AI](https://papers.fru.dev/topics/applied)

## Most cited

- [Attention Is All You Need](https://papers.fru.dev/papers/attention-is-all-you-need): Introduced the Transformer, the attention-only architecture behind nearly every large language model since. (193,621 citations)
- [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://papers.fru.dev/papers/bert): Pre-training a Transformer to fill in masked words, then fine-tuning it, beat task-specific models across language understanding tests. (121,314 citations)
- [Language Models are Few-Shot Learners](https://papers.fru.dev/papers/gpt-3): GPT-3, a 175B-parameter model, does new tasks from a few examples in the prompt, with no fine-tuning. (63,622 citations)
- [GPT-4 Technical Report](https://papers.fru.dev/papers/gpt-4-technical-report): A multimodal model that passes a simulated bar exam with a score around the top 10% of test takers. (27,281 citations)
- [Training language models to follow instructions with human feedback](https://papers.fru.dev/papers/instructgpt): InstructGPT: fine-tuning on human feedback made a 1.3B model preferred over the 175B GPT-3. (24,241 citations)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://papers.fru.dev/papers/lora): LoRA fine-tunes a large model by training small low-rank matrices, cutting trainable parameters by 10,000 times. (23,235 citations)
- [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://papers.fru.dev/papers/chain-of-thought-prompting): Asking a model to write out its intermediate steps (chain of thought) sharply improves its math and logic answers. (22,107 citations)
- [LLaMA: Open and Efficient Foundation Language Models](https://papers.fru.dev/papers/llama): Open models from 7B to 65B trained only on public data; the 13B model beats GPT-3 on most benchmarks. (21,675 citations)

## Latest

- [RRSI: Regularized Recursive Self-Improvement of Agent Harnesses](https://papers.fru.dev/papers/rrsi): RRSI keeps self-improving agent harnesses from memorising their training tasks, so gains carry over to new benchmarks.
- [DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression](https://papers.fru.dev/papers/deepseek-v4-1-flash): A 552B mixture-of-experts model with a 1M-token context, built to shrink the KV cache for agent workloads.
- [StudentSim: Training LLM-based Student Simulators](https://papers.fru.dev/papers/studentsim): StudentSim trains per-student simulators that answer like a given learner and change their answers under a tutor's guidance.
- [On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability](https://papers.fru.dev/papers/qwen3-8-next-architecture): Qwen3.8-Flash-Next: a 125B mixture-of-experts model with 6B active that nearly matches its 397B predecessor at 1/9 the training compute.
- [Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes](https://papers.fru.dev/papers/towards-physics-of-multimodal-pretraining): Controlled experiments on when text, image understanding and image generation help or compete when trained together.
- [Gemma 4 Technical Report](https://papers.fru.dev/papers/gemma-4-technical-report): Open multimodal models from 2.3B to 31B parameters, with a thinking mode and image and audio input.
- [Agents' Last Exam](https://papers.fru.dev/papers/agents-last-exam): Over 1,000 real, checkable professional tasks written with 250+ industry experts; the hardest tier is far from solved.
- [Cosmos 3: Omnimodal World Models for Physical AI](https://papers.fru.dev/papers/cosmos-3): One model that reads and generates text, images, video, audio and robot actions for physical AI.

## API

- [For AI agents: how to use this data](https://papers.fru.dev/agents)
- [OpenAPI 3.1 spec](https://papers.fru.dev/openapi.json)
- GET https://papers.fru.dev/api/papers?lab=&topic=&since=&sort=date|citations|upvotes&limit=200&offset=0 : papers with their line, arXiv link and counts
- GET https://papers.fru.dev/api/papers/<slug> : one paper, its change history and matching product releases from releases.fru.dev
- GET https://papers.fru.dev/api/labs and https://papers.fru.dev/api/labs/<slug> : labs and their papers
- GET https://papers.fru.dev/api/topics : topics with counts
- GET https://papers.fru.dev/api/changes?since= : append-only history of every changed value
- GET https://papers.fru.dev/api/companies?since= : labs with registry slug, domain, page and dated papers (for companies.fru.dev)
- GET https://papers.fru.dev/api/search?q= : ranked papers, labs and pages
All GET, CORS open, cached for an hour. Free to read; please cite "Papers (papers.fru.dev)".

## Data

- [Full list as plain text](https://papers.fru.dev/llms-full.txt)
- [Sitemap](https://papers.fru.dev/sitemap.xml)