EvoSkill: Automated Skill Discovery for Multi-Agent Systems

ArXiv ID: 2603.02766

Date: March 3, 2026

Authors: Salaheddin Alzubi, et al.

Abstract: EvoSkill is a self-evolving framework that automatically discovers and refines agent skills through iterative failure analysis. It analyzes execution failures, proposes new skills or edits, and materializes them into structured, reusable folders. A Pareto frontier governs selection, retaining only skills that improve performance while the underlying model remains frozen.

RSI Bench Relevance:

Methodology: Demonstrates "Recursive Optimization" at the skill/workflow level rather than raw model weights.
Metrics: Shows +7.3% on OfficeQA and +12.1% on SealQA.
Transferability: Skills evolved on one task showed zero-shot transfer (+5.3% on BrowseComp), validating that high-level logic evolution produces generalizable capabilities.

View Original on ArXiv

Back to Paper Index