SkillDAG: Exposing Typed Skill Graphs to LLM Agents Boosts ALFWorld Success by 14.3 Points

SkillDAG represents a new approach to skill selection for Large Language Model (LLM) agents. Instead of treating skills as independent documents for vector retrieval, it constructs a typed skill graph and exposes it as an agent-callable interface. This method addresses the limitations of enumeration or Top-K embedding retrieval, which often fail to capture dependencies, conflicts, and specializations between skills, leading agents to omit necessary prerequisites during task execution.

Experiments across two benchmarks and three model families show that SkillDAG outperforms full-library prompting, vector retrieval, and Graph-of-Skills. Specifically, it achieved a 14.3-point increase in task success rate in the ALFWorld environment. Furthermore, its ranking remained stable even when the skill pool expanded tenfold, whereas fixed pipeline methods degraded under similar conditions.

These results are self-reported by the author team and have not yet been independently reproduced. The core contribution lies in demonstrating that allowing agents to actively query and revise skill structures is more scalable and effective than having a retriever passively provide opaque bundles of skills, offering a new architectural direction for managing large-scale LLM agent capabilities.