AREX-Skill boosts MLE-bench medal rate from 31% to 73% with distilled agent skills
TheTuringPost · x · 2026-09-05
AREX-Skill distills knowledge from GitHub repos into reusable skill packages that teach AI agents to operate software in practice:
- Scale: the library holds 5,000+ verified skills from 1,000 ML repos; agents load only task-relevant skills.
- Structure: each skill centers on a SKILL.md plus references, scripts, validation checks and failure-recovery guidance, using the open Agent Skills format exportable to Codex, Claude Code and others.
- Results: with the same GPT-5.5 model, harness and execution budget, MLE-bench "Any Medal" rate jumped from 31.11% to 72.89%.
- Behind it: DisCo, a research agent with Creator and Researcher modes that creates, verifies and routes skills — now the most popular paper of the week.
Related event: DisCo Distills 5000+ Skills from ML Repos, Boosting Agent Benchmark Scores(2 posts)→
More from coding & agent
- Managing AI agents is basically like managing a 4-year-old, developer says — RachelVT42 · 2026-09-05
- GPT-6 Astra Auto-Preps Entire Final Cut Project: Import, Color Grade, Sync Clips — minchoi · 2026-09-05
- GPT-6 Astra Ultra Generates an App in 26 Minutes Using 10.46M Tokens — minchoi · 2026-09-05
- GPT-6 Astra rolls out to all users; dev cuts LEGO modeling from hours to 10 minutes — iamrobotbear · 2026-09-05
- Devs call to abolish context compaction: agent engineering debate heats up — sull · 2026-09-05
- GPT-6 Astra arrives in GitHub Copilot, built for long-horizon autonomous coding — OpenAIDevs · 2026-09-05