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

Abstract:

As Large Language Model (LLM) agents increasingly rely on expansive skill libraries, the challenge of selecting appropriate subsets shifts from simple similarity matching to a complex structural problem. Skills often exhibit dependencies, conflicts, specializations, or redundancies with one another—nuances that remain opaque to both full enumeration techniques and embedding-based similarity searches. To address this, we introduce SkillDAG, a framework that represents inter-skill relationships as a typed directed graph. This graph serves as an inference-time, agent-callable structural retrieval interface, allowing the system to be queried and dynamically evolved during execution rather than being confined to a static retrieval pipeline.

During each search operation, the system retrieves vector matches, neighbors connected by typed edges, and conflict indicators. Furthermore, a "propose-then-commit" protocol enables the agent to register edges backed by actual execution, allowing the graph to accumulate structural knowledge across different episodes.

In evaluations on the ALFWorld and SkillsBench datasets using MiniMax-M2.7, SkillDAG achieved a success rate of 67.1% and a reward score of 27.3%. These results surpass the strongest previously reported Graph-of-Skills baseline by margins of 12.8 and 8.6 points, respectively. This performance advantage extends to the gpt-5.2-codex model as well, where intrinsic SkillsBench Ret@K scores increased from 65.5 to 78.2 under matched query conditions.

These improvements are attributed to distinct, isolable mechanisms. First, candidate ranking remains robust even when the skill pool expands tenfold, whereas fixed seeding-diffusion pipelines typically degrade under similar conditions. Second, the system employs set-monotone online edits that expand ground-truth recall without removing previously identified relevant items.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC