GraphSkillEvo evolves graph-structured skills for LLM agents, +4% on benchmarks

Rui Sun · hf · 2026-09-21

GraphSkillEvo tackles two flaws of natural-language agent skills: lack of workflow-level guidance with redundancy, and an unconstrained search space that makes optimization ineffective. Skills are represented as graphs—nodes are execution steps with operational guidance, directed edges encode context-dependent transitions. A population-based evolutionary framework with mutation and crossover operators explores this structured skill space more broadly than LLM-only self-refinement, beating the SkillOpt baseline by +4.01% average accuracy on GPT-5.4-nano and +1.76% on GPT-5.4 across five agent benchmarks. Code: github.com/ruisun7/GraphSkillEvo.

Original post →

More from coding & agent

coding & agent channel →