SkillAA Routes a Failed Rollout to a Specific Node in the Skill Graph, Then Gates and Rolls Back the Edit
SkillAA (arXiv 2609.20455, submitted 17 Sep 2026) targets a concrete weakness in agent skill libraries: existing methods edit skills directly from failed rollouts with no structured route from an observed failure to an editable location. It represents skill applicability, execution and composition in one graph so the same structure serves selection, attribution-guided repair and scoped validation, contrasts successful against failed executions to route candidate repairs to specific graph objects, updates only local structure, and screens changes through Local and Big Gates before commitment. With gpt-5.6-sol it reaches 81.5% on SearchQA, 66.7% on LiveMath and 91.2% on DocVQA, the highest observed mean in every main setting, with no parameter updates to the frozen model.
↳ Follow the thread