Research Library

Library

Papers, patents & whitepapers, annotated

Preprint2026

Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

Siyuan Huang, Pengyu Cheng, Haotian Liu et al. · arXiv · 2026

Self-play is how systems surpassed human data in games, and this is a serious attempt to port that to general model capability. Read it alongside the other co-evolution work here to see the pattern forming.

Preprint2026

Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems

Gaurav Dadhich · arXiv · 2026

Read this before you build another retrieval layer for your agent. The quadratic cost argument alone reframes context management as an architecture decision rather than a tuning detail.

Preprint2026

SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation

Junsong Chen, Jincheng Yu, Yitong Li et al. · arXiv · 2026

The clearest recent evidence that linear attention works as one half of a hybrid rather than as a wholesale replacement — the same trade-off now shaping long-context language models.

Preprint2026

K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs

Hao Liang, Qihan Lin, Zhaoyang Han et al. · arXiv · 2026

A good demonstration that a benchmark score answers only the question the benchmark asks. Models that ace exam items still fumble the structure underneath them — a real gap for anyone building tutoring products.

Preprint2026

Sample-Efficient Learning from Agent Experience

Chenhui Gou, Haoqin Tu, Yunhao Fang et al. · arXiv · 2026

The gap between 65 and 4 percent is the entire paper. How you train on collected experience matters far more than whether you collect it in the first place.

Preprint2026

Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning

Jian Hu, Huiying Li, Hao Zhang et al. · arXiv · 2026

Worth reading as an argument about research velocity rather than as a tool announcement: when every algorithm change has to thread through layers of framework glue, the framework has become the bottleneck.

Preprint2026

Interactive Training 2: Auditable Control Plane for Live Model Training

Wentao Zhang, Xuanhe Pan, Han Zhou et al. · arXiv · 2026

The auditability is the point. Once agents can steer training runs, a written record of who changed what and what happened next stops being a nicety and becomes the only way to debug.

Preprint2026

Multi-Turn On-Policy Distillation with Prefix Replay

Baohao Liao, Hanze Dong, Christof Monz et al. · arXiv · 2026

Pairs naturally with the relay-distillation work in this collection — both attack the same tension from opposite sides. Read them together if you are training agents on teacher supervision.

Paper2025

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-AI, Daya Guo, Dejian Yang et al. · Nature volume 645, pages 633-638 (2025) · 2025

The result that reframed reasoning as something trained rather than prompted. That the behaviour emerges from outcome-only rewards — nobody demonstrated the reasoning — is the part worth sitting with.

Whitepaper2024

DeepSeek-V3 Technical Report

DeepSeek-AI, Aixin Liu, Bei Feng et al. · DeepSeek (Technical Report) · 2024

Read it for the training economics rather than the benchmarks. It is the clearest public account of how a frontier-scale run is actually made affordable, and much of what followed in open-weight modelling traces back to choices documented here.

Whitepaper2024

Qwen2.5 Technical Report

Qwen, :, An Yang et al. · Alibaba Qwen (Technical Report) · 2024

The most practical open-weight family to build on, because a consistent recipe across sizes means you can prototype small and scale up without changing behaviour underneath you.

Whitepaper2024

The Llama 3 Herd of Models

Meta AI · Meta (Technical Report) · 2024

A rare open look at how a frontier-scale model is actually built end to end. Read it when you want the engineering reality behind the benchmarks.

Whitepaper2024

Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

Marah Abdin, Jyoti Aneja, Hany Awadalla et al. · Microsoft (Technical Report) · 2024

The strongest evidence that data curation substitutes for parameters at the small end. Relevant if you are deciding between a hosted frontier model and something you can run on-device.

Whitepaper2024

StarCoder 2 and The Stack v2: The Next Generation

Anton Lozhkov, Raymond Li, Loubna Ben Allal et al. · BigCode (Technical Report) · 2024

Notable for treating training-data provenance as a design problem with an actual mechanism rather than a disclaimer. If you care where code models get their material, this is the reference implementation.

Paper2024

OLMo: Accelerating the Science of Language Models

Dirk Groeneveld, Iz Beltagy, Pete Walsh et al. · ACL · 2024

'Open-weights' and 'open' are not the same claim, and this is the paper that makes the distinction concrete. If you want to study how a model acquires a capability during training rather than just probe the finished artefact, this is the one that lets you.

Whitepaper2023

Gemini: A Family of Highly Capable Multimodal Models

Gemini Team, Rohan Anil, Sebastian Borgeaud et al. · Google DeepMind (Technical Report) · 2023

The clearest statement of the native-multimodal bet: that joint training beats bolting encoders onto a language model. Worth pairing with the scaling study in this collection, which examines whether that bet pays off under a fixed compute budget.

Whitepaper2023

Mistral 7B

Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch et al. · Mistral AI (Technical Report) · 2023

The paper that made small open models credible. Both attention tricks are now near-universal, so it doubles as the clearest short explanation of why modern models serve cheaply.

Whitepaper2023

Llama 2: Open Foundation and Fine-Tuned Chat Models

Hugo Touvron, Louis Martin, Kevin Stone et al. · Meta (Technical Report) · 2023

For years this was the reference text on how a chat model is actually aligned, because it showed the process rather than just the result. The two-reward-model split is the detail most worth carrying away.

Paper2023

Direct Preference Optimization: Your Language Model is Secretly a Reward Model

Rafael Rafailov, Archit Sharma, Eric Mitchell et al. · NeurIPS · 2023

It made preference tuning something a small team can run, which is why most open fine-tunes now use this rather than the reinforcement-learning pipeline it replaced. A good example of a derivation, not a bigger model, being the contribution.

Whitepaper2023

PaLM 2 Technical Report

Rohan Anil, Andrew M. Dai, Orhan Firat et al. · Google (Technical Report) · 2023

A useful marker of when the industry stopped equating scale with parameters. Read it next to the Chinchilla paper here — this is that argument applied at production scale.

Showing 2140 of 66 sources · newest first