Survey maps how LLM agents recursively self-improve via prompts, memory, skills, harness
机器之心 · wechat · 2026-11-15
A survey from CASIA, Meituan and other institutions systematizes Recursive Self-Improvement (RSI) in LLM agents around three questions (When/Where/How to update), four evolution tracks, and a seven-step lifecycle. It covers 40 representative works and compares 20 update loops.
Four tracks: Prompt optimization (search → feedback-driven revision → optimizing the optimizer itself); Memory evolution (storage → distilled rules →时效/矛盾管理与效用估计); Skill evolution (reusing success sequences → continuous repair → skill-library curation); Harness evolution (making the runtime framework itself feedback-driven, e.g. auto-adding a "tests must pass" completion check).
Lifecycle: Collect→Evaluate→Diagnose→Abstract→Update→Verify→Consolidate, turning one-off fixes into verified, reusable structure. Open challenges: cross-component attribution, long-term maintenance, and independently verifying whether inherited experience actually improves subsequent improvement.
More from coding & agent
- Design engineering tip: show the first frame before motion content is ready — floguo · 2026-11-15
- FreeToken open-source engine runs 284B DeepSeek locally at 25 tok/s on RTX 5090 — aigclink · 2026-11-15
- Solo lawyer runs an AI-native law firm on a Claude agent team organized like company departments — prasanna_says · 2026-11-15
- Claude builds a colorful 3D city in three.js with zero assets, all code — Eric520CC · 2026-11-15
- fate 2.2.1 ships massive perf gains: up to 343x faster record deletion — cnakazawa · 2026-11-15
- METR open-sources its AI agent evaluation platform for safely testing frontier models — BethMayBarnes · 2026-11-15