SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries
The paper claims agents can shrink large skill libraries without breaking the executable contracts those skills rely on.
SkillZip compresses skills as section-level execution graphs, replacing repeated valid patterns with reversible macros. The method is designed to preserve boundary signatures, dependencies, verifier paths, and the ability to expand back to source-level routines. At inference time, it loads a compact dependency-closed context and expands macros only when needed. The authors report up to a 12.2-point gain over the strongest baseline, with 3.46x compression and high dependency and verifier preservation. HF Daily Papers' note
SkillZip compresses skills as section-level execution graphs, replacing repeated valid patterns with reversible macros. The method is designed to preserve boundary signatures, dependencies, verifier paths, and the ability to expand back to source-level routines. At inference time, it loads a compact dependency-closed context and expands macros only when needed. The authors report up to a 12.2-point gain over the strongest baseline, with 3.46x compression and high dependency and verifier preservation. HF Daily Papers' note
score 4