Microsoft's ScholarEvolve lets agents improve by reading papers, lifting Qwen's AppWorld score from 69% to 81.4%

bendee983 · x · 2026-10-22

Microsoft and UCSB researchers unveiled ScholarEvolve, a framework for "self-improving AI harnesses": agents upgrade their own harness by ingesting new research papers, no retraining required.
Pipeline: collect trajectories on target tasks → failure audit and categorization → convert failures into research questions → retrieve relevant papers and code → cluster and rank them per harness module (Tool/Context/Skill/Memory/Workflow) → generate research-grounded mutation blueprints.
Evolution runs a genetic-style Mutation → Selection → Crossover loop: candidate variants per module, best-single-branch selection with regression checks, then module crossover; stops on iteration budget or performance convergence.
Reported result: Qwen's AppWorld task completion rose from 69% to 81.4%. The approach aims to break past the limitation that current harness optimization is bounded by the backbone model's internal knowledge.

More from coding & agent