Agents Learn to Grow Skills from Failure: EvolveR Accepted by ICML 2026
Agent学会自己「长」Skill了!从失败里长出经验,比人类写的更好用|ICML 2026
EvolveR lets agents distill reusable experience from successful and failed trajectories, maintain a scored experience library, and train retrieval behavior with GRPO; the paper reports the best average performance on seven complex QA benchmarks using Qwen2.5-3B and 7B.
Why it matters: EvolveR scores agent successes and failures into a library and uses GRPO to train when to retrieve, tested on Qwen2.5-3B and 7B across seven complex QA benchmarks.