BAAI's DisCo: distilling 1,000 ML repos into 5,000+ skills lifts research agents 134% on MLE-bench
dair_ai · x · 2026-09-03
dairai highlights a BAAI paper on DisCo, aimed at anyone building research agents. Adding distilled skills yields:
- MLE-bench: +134.3%
- PaperBench: +34.4%
- FrontierCS: +9.2%
- PassNet: +14.0%
Approach: even a strong backbone with planning/execution/memory/verification harnesses doesn't know how to make a given method actually work — that know-how lives in repos and papers too large to load during a task. DisCo distills it: task-agnostic distillation condenses 1,000 widely used ML repositories into the AREX-Skill Library (5,000+ verified skills across 20 areas and 178 capability families), while task-oriented distillation writes the skills a concrete task calls for.
Related event: BAAI Open-Sources DisCo: Distilling GitHub Repos into Reusable AI Skills(3 posts)→
More from coding & agent
- LatentPress compresses context into continuous memory tokens for frozen decoders — Zhengze Zhou · 2026-09-04
- Tencent Hunyuan's environment evolution keeps terminal agents learning via rising task difficulty — Tencent-Hunyuan · 2026-09-04
- Qwen's Terminal-Universe turns agent trajectories into scalable terminal training environments — Qwen · 2026-09-04
- Jim Fan on World of Bits: OpenAI's 2016 booking-agent dream finally realized by GPT-6 — DrJimFan · 2026-09-04
- DHH secures $1.95M in model tokens for Omarchy, with Meta as lead patron — sull · 2026-09-04
- "AI slop is so bad I now keep private forks of my whole dev stack," says PySR author Miles Cranmer — MilesCranmer · 2026-09-04