🔬Agent Skills Lift MLE-Bench Scores 134% at Fixed Compute
TL;DR
DisCo distills 1,000 widely used ML repos into 5,000-plus verified skills, then hands them to a research agent with backbone and budget unchanged. MLE-bench scores rise 134.3%, PaperBench 34.4%, FrontierCS 9.2% and PassNet 14.0%.
DisCo distills 1,000 widely used ML repos into 5,000-plus verified skills, then hands them to a research agent with backbone and budget unchanged. MLE-bench scores rise 134.3%, PaperBench 34.4%, FrontierCS 9.2% and PassNet 14.0%.
Key Points
arXiv 2609.02749, submitted Sep 2 by Jianlyu Chen and 10 co-authors
AREX-Skill Library holds 5,000+ verified skills across 20 areas and 178 capability families
Backbone fixed at GPT-5.5 with the harness and execution budget held constant
Gains: MLE-bench +134.3%, PaperBench +34.4%, FrontierCS +9.2%, PassNet +14.0%
Two distillation modes: task-agnostic repo condensation and task-oriented skill synthesis
Why It Matters
The cheapest capability gain on the table right now may be packaging operational know-how as skills rather than buying a bigger model.
Quick Facts
Frequently Asked Questions
Why does this matter?
The cheapest capability gain on the table right now may be packaging operational know-how as skills rather than buying a bigger model.
What happened?
DisCo distills 1,000 widely used ML repos into 5,000-plus verified skills, then hands them to a research agent with backbone and budget unchanged. MLE-bench scores rise 134.3%, PaperBench 34.4%, FrontierCS 9.2% and PassNet 14.0%.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,463 builders reading daily.