AI RESEARCH

OpenAI's next model, codenamed Astra, produced results on ten problems open for at least a decade - including non-sofic groups and Erdős problems 146, 180, and 183 - with every argument formalized as a Lean certificate.

A SCAM 2026 study of 100 top-starred repos catalogs six configuration smells in AGENTS.md and CLAUDE.md files: Lint Leakage in 62%, Context Bloat in 42%, Skill Leakage in 35%. Only 9 of 100 files were smell-free.

A new 106-issue benchmark across 49 repositories finds frontier coding agents rarely retrieve AI contribution rules on their own - and never refuse to contribute in AI-banned repositories, no matter the prompt. Disclosure and verification can be fixed; bans cannot.

A controlled ablation across Claude Code and Codex, 17 real tasks, and 288 evaluated runs finds context-injection strategy does not measurably change correctness (bounded to under 10-15pp). The failures are implementation skill, not missing repository knowledge.

A calibrated study on real ConflictBench Java conflicts finds LLM agents match the developer's own resolution on 55-59% of true conflicts versus 36.7% for the best structured tool. The edge is coverage, not accuracy: the tools abstain on 20-90% of conflicts, the LLM on none.

Moonshot AI's Kimi Linear paper introduces KDA, a hybrid linear attention that beats full attention at all scales - 75% less KV cache, 6x decoding at 1M context, and open-source checkpoints.

A researcher's 10-page domain-expert prompt helped GPT-5.6 produce a Lean-verified proof closing a complexity gap that stood since 1996. The paper is now on arXiv.

OpenAI claims GPT-5.6 Sol Ultra has generated a proof for a 50-year-old graph theory conjecture in under an hour. The math community is now verifying whether it holds up.

Anthropic's new research reveals LLMs have an internal 'workspace' for silent reasoning - and it could change how we build safer AI.

Mistral releases Leanstral 1.5, an Apache-2.0 licensed 119B parameter model (6B active) for Lean 4 theorem proving that saturates miniF2F and achieves SOTA on FATE benchmarks.

A new paper shows a 3B parameter model hitting 94.3 on AIME26 and 96.1% on LeetCode contests - matching or exceeding models 100x its size. The catch: it traded general knowledge for pure reasoning ability.