HiSkill: Hierarchical Skill Graphs for LLM Agents
HiSkill, a novel framework, structures the interaction paths of LLM agents into a directed graph comprising skill nodes, AtomicOp nodes, and typed edges. It effectively represents relationships such as decomposition, temporal transitions, compatibility, support, and recovery among various skills. During the inference process, HiSkill extracts a concise subgraph pertinent to the task, facilitating guided execution through the use of a symbolic task state.
Key facts
- HiSkill is a hierarchical skill graph framework for LLM agents.
- It organizes interaction trajectories into a directed graph.
- The graph includes skill nodes, AtomicOp nodes, and typed edges.
- It captures decomposition, temporal transition, compatibility, support, and recovery relations.
- At inference, it retrieves a compact task-relevant subgraph.
- Execution is guided by the subgraph and a symbolic task state.
- The paper is available on arXiv with ID 2607.25853.
- The research was published in 2025.
Entities
Institutions
- arXiv