The paper introduces representing skills for LLM agents as directed graphs, where each node is an execution step with guidance and edges show transitions. This structured representation provides clearer workflow guidance and reduces the search space for optimization. The authors propose GraphSkillEvo, an evolutionary algorithm that mutates and recombines across a population. Experiments on five benchmarks show accuracy improvements of 4.01% on GPT-5.4-nano and 1.76% on GPT-5.4 compared to the unstructured baseline SkillOpt.
Graph-structured skills improve LLM agents through evolutionary optimization
Across five agent benchmarks, evolving workflow graphs outperformed iterative optimization of unstructured natural-language instructions.
Top University
Rui Sun · Zhi Zheng · Zhenkun Wang · Zhichao Lu
City University of Hong Kong · National University of Singapore · Southern University of Science and Technology
Research Digest··2 min read
Thread:Skill Selection for Agents
In one line
Graph-structured skills improve LLM agent performance more than unstructured skill optimization.
Before you read this
Two ideas this paper builds onRead it as
The authors represent agent skills as directed graphs whose nodes specify execution steps and whose edges define context-dependent transitions.
The authors replaced free-form skill instructions with graph-structured natural-language artifacts.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code