HiSkill: Empowering LLM Agents with Hierarchical Skill Graphs

Researchers introduced HiSkill, a hierarchical skill graph for LLM agents that links reusable skills to executable actions.

HiSkill organizes past interaction trajectories into skill nodes and AtomicOp nodes with typed relations, then retrieves relevant subgraphs for task execution.

In experiments across three interactive environments, the method outperformed baselines and used fewer tokens.

The approach improves long-horizon performance by grounding high-level skills in executable steps while reducing inference cost.