Researchers have introduced GraphSkillEvo, a method that rethinks how procedural skills are represented and optimized for Large Language Model (LLM) agents. The work is published as a paper on Hugging Face.
The paper argues that existing skill optimization approaches typically encode skills as unstructured natural-language instructions, which creates two problems: such skills often lack explicit workflow-level guidance and carry substantial redundancy, making them hard for LLMs to execute, and the unconstrained natural-language search space makes optimization ineffective.
To address this, the authors represent skills as graph-structured natural-language artifacts. Each node holds an execution step along with its operational guidance, while directed edges encode context-dependent transitions between steps. This structure is intended to supply clearer workflow-level guidance and to make the skill space more amenable to search.
Building on that representation, GraphSkillEvo is a population-based evolutionary optimization framework with mutation and crossover operators tailored to graph-structured skills. By maintaining multiple candidate skills and recombining effective components, it aims for broader exploration than purely LLM-based iterative self-refinement.
Experiments across five agent benchmarks show GraphSkillEvo consistently outperforming the strong skill optimization baseline SkillOpt, improving average accuracy by 4.01% on GPT-5.4-nano and 1.76% on GPT-5.4. The authors state that their code is available at https://github.com/ruisun7/GraphSkillEvo.