SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents by Discovering Reusable Structure
SkillZip compresses agent skills by sharing repeated structure without running evaluation rollouts.
The paper frames self-evolving agents as accumulating bloated skill records: repeated rules, copied workflows, and scattered exceptions. SkillZip turns those into a shorter structural explanation, keeping required triggers, workflow edges, tool constraints, obligations, and output fields covered. It has a one-shot mode and a continual “Zip-on-Write” mode for integrating new patches without replaying tasks or reparsing full history. The authors report better compression, generalization, and cost overhead in their experiments. HF Daily Papers' note
The paper frames self-evolving agents as accumulating bloated skill records: repeated rules, copied workflows, and scattered exceptions. SkillZip turns those into a shorter structural explanation, keeping required triggers, workflow edges, tool constraints, obligations, and output fields covered. It has a one-shot mode and a continual “Zip-on-Write” mode for integrating new patches without replaying tasks or reparsing full history. The authors report better compression, generalization, and cost overhead in their experiments. HF Daily Papers' note
score 5