GraphSkillEvo: Evolutionary Optimization of Graph-Structured Agent Skills
The paper tests graph-shaped agent skills as an optimization target, not just as cleaner prompts.
GraphSkillEvo represents each skill as steps and directed transitions, then uses evolutionary mutation and crossover to refine those structures. The authors argue this narrows the search space compared with free-form natural-language skill rewriting. Across five agent benchmarks, it beats SkillOpt by 4.01% average accuracy on GPT-5.4-nano and 1.76% on GPT-5.4. Code is listed as available. HF Daily Papers' note
GraphSkillEvo represents each skill as steps and directed transitions, then uses evolutionary mutation and crossover to refine those structures. The authors argue this narrows the search space compared with free-form natural-language skill rewriting. Across five agent benchmarks, it beats SkillOpt by 4.01% average accuracy on GPT-5.4-nano and 1.76% on GPT-5.4. Code is listed as available. HF Daily Papers' note
score 4