SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents by Discovering Reusable Structure
SkillZip compresses agent skills by extracting shared structure without running evaluation rollouts.
The paper frames growing skill libraries as a maintenance problem: repeated rules, copied action sequences, and scattered exceptions make skills costly to inject into agents. SkillZip uses a typed minimum-description-length objective to state common requirements once, factor repeated workflows, and preserve rare rules under a hard coverage constraint. It offers a one-shot mode and a continual “Zip-on-Write” mode for integrating new self-evolution patches without replaying tasks. The authors report stronger compression, generalization, and cost overhead results in their experiments. ArXiv · AI/CL/LG's note
The paper frames growing skill libraries as a maintenance problem: repeated rules, copied action sequences, and scattered exceptions make skills costly to inject into agents. SkillZip uses a typed minimum-description-length objective to state common requirements once, factor repeated workflows, and preserve rare rules under a hard coverage constraint. It offers a one-shot mode and a continual “Zip-on-Write” mode for integrating new self-evolution patches without replaying tasks. The authors report stronger compression, generalization, and cost overhead results in their experiments. ArXiv · AI/CL/LG's note
score 5