DisCo distills GitHub repos into agent skills, doubling MLE-bench to 72.89%
rohanpaul_ai · x · 2026-09-06
The DisCo paper argues agents waste effort rediscovering operational know-how — which tool to use, how to set it up, how to recover from failures.
- Approach: distill GitHub repos and papers into compact, verified reusable skills, both task-agnostic (a general library) and task-oriented.
- Scale: the AREX-Skill Library holds 5,353 skills from 1,000 widely used ML repos, organized into 20 areas and 178 capability families.
- Results: with the GPT-5.5 backbone, harness, and execution budget fixed, MLE-bench jumps from 31.11% to 72.89% (+134.3%), PaperBench +34.4%, with FrontierCS and PassNet improving too.
Takeaway: better agents may come from teaching proven workflows, not just scaling the underlying model.
More from coding & agent
- Even with command confirmation, coding agents still try 'cd /; rm -rf' — julianharris · 2026-09-06
- web-to-app: build Android APKs from web projects entirely on your phone — tom_doerr · 2026-09-06
- With full StarCraft source code, how fast could you vibecode StarCraft 3? — jakedahn · 2026-09-06
- Engineer's agent has been hunting rogue agent activity on Wikidata for 24 hours — chrisalbon · 2026-09-06
- Humanize + GPT-5.5 solves 670/672 Lean-verified proofs, tops PutnamBench at 99.7% — songhan_mit · 2026-09-06
- Astra Demo Built With Just 400 Lines of PlayCanvas, 3D Assets via Blender MCP — willeastcott · 2026-09-06