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
- Rork claims App Store publishing is now 2.5x to 4x faster — rudrank · 2026-09-07
- Hands-on test: AI ops agent Ghost diagnoses and fixes three simultaneous service failures — BLUECOW009 · 2026-09-07
- Dev runs a live show produced entirely by an agentic AI research team built on a wiki — SurvivedDravoswater · 2026-09-07
- Open-source Codex Skill turns one sentence into a holographic 3D card with Blender and Three.js — Promptmethus · 2026-09-07
- Jim Koppel: the vibe coder's way to understand code is running everything; the agentic engineer reads it — jimmykoppel · 2026-09-07
- OpenAI reportedly testing ChatGPT feature that mimics your writing style from emails and docs — VraserX · 2026-09-07