Skill documents are becoming the default adaptation surface for frozen large language model (LLM) agents, yet current skill generators suffer from two major limitations. First, humans and different LLMs organize procedural knowledge and solve tasks in distinct ways; skills authored from a single perspective may misalign with the reasoning patterns of other backbones, restricting cross‑model generalization. Second, sequential skill descriptions fail to capture parallel branches and explicit dependencies, and existing skill optimizers operate mainly on prose without native support for graph‑structured evolution.
To address these issues we introduce MultiEvoSkill, a skill‑optimization framework built around three core components:
- Skill Compiler casts a skill into a typed directed graph, employing specialized operators to produce a compact, information‑rich representation with explicit procedural topology.
- Heterogeneous Agent System generates revision advices via collective intelligence, pinpointing minimal yet informative issues based on multiple experience sets.
- Structure‑Aware Debate and Evolution Mechanism organizes candidate edits into a heterogeneous graph, resolves conflicts under dependency and importance constraints, and distills the collective proposal into the final artifact.
A multi‑dimensional gating scheme ensures each evolution step yields information gain while preserving topological correctness. Optimal skills emerge when the system converges to a consistent state. Extensive experiments and ablations against state‑of‑the‑art baselines demonstrate MultiEvoSkill’s effectiveness and its ability to generalize across models.
Review