SkillEvolBench: Benchmarking the Evolution from Episodic Experience to Procedural Skills

Researchers introduced SkillEvolBench, a benchmark of 180 tasks across six real-world agent environments.

It tests whether LLM agents can turn episodic trajectories into reusable procedural skills.

Across models and harnesses, agents often adapt to local tasks but fail to build robust, generalizable skills.

In many cases, reusing raw trajectories works better than distilled skill representations.

The results highlight a major limitation in current agent learning and skill formation.