SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale

In three linesSkillDAG models inter-skill relationships as a typed directed graph for dynamic LLM agent skill selection at inference time. On ALFWorld and SkillsBench with MiniMax-M2.7, it achieves 67.1% success and 27.3% reward, exceeding Graph-of-Skills baselines by +12.8 and +8.6 points. The graph self-evolves during execution via a propose-then-commit protocol, accumulating structure across episodes.

Summary generated by Claude — human-verified