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: HKR-H/K/R all pass: the agent self-growing-skill angle is clickable, with mechanism and benchmark specifics. Since only a media summary is available and no repo, absolute scores, or reproduction details are disclosed, it stays in the 78–84 research band.