Microsoft's ScholarEvolve evolves agent harnesses with frozen models, +14 points on AppWorld
jiqizhixin · x · 2026-10-31
UCSB and Microsoft introduce ScholarEvolve, which keeps the LLM's parameters frozen and instead lets an agent read research literature to evolve its own software harness across five modules: tools, context, skills, memory, and workflow.
Key idea
- Learning only from observed failures limits search to familiar fixes; useful techniques often already exist in papers but must be translated into working code and tested in combination.
- Pipeline: failure audit → convert failures to research questions → literature retrieval → cluster papers by mechanism across modules → generate research-grounded mutation blueprints → validate candidates → module crossover across generations, with regression checks.
Results
- With Qwen3.5-27B, AppWorld Challenge task-goal completion rises from 49.6% to 63.6%, a 14.0-point gain.
- The evolved harness includes tool-action validation & repair, trajectory-induced pseudocode skills, planner/solver/critic-verifier workflows, query-aware context budgeting, and structured episodic recall, with no test-set leakage.
More from coding & agent
- Claude finds a real 0-day in project dependencies from a single prompt — DanielLockyer · 2026-10-31
- Indie dev ships one threejs codebase across Steam, iOS, Mac, Apple TV and Vision Pro — ezshine · 2026-10-31
- Codex chief teases two banked usage resets within two days — JeremyNguyenPhD · 2026-10-31
- Trusted hook, untrusted code: git pre-commit can execute agent-modified package.json scripts — izgorodin · 2026-10-31
- Two-agent workflow: build a design system first, then lock UI rules to 10x vibe-coded dashboards — pramodk73 · 2026-10-31
- Framework vs Harness: How to Choose Between LangGraph, Deep Agents and Claude Code — Saboo_Shubham_ · 2026-10-31