External skills can supply domain procedures without updating model parameters, yet current approaches often edit skills directly from failed rollouts, lacking a structured route from an observed failure to an editable location. Moreover, existing skill graphs underutilize semantic boundaries, object addresses, and topological dependencies for skill retrieval, targeted updating, and scoped validation. We introduce SkillAA (Skill Abductive Attribution), a structured skill‑optimization framework for frozen language models. SkillAA represents skill applicability, execution, and composition within a unified graph, enabling the same structure to support skill selection, contrast successful and failed executions, and guide repairs to specific graph objects, updating only the selected local sub‑structure. The framework then employs two layers of screening—Local Gate and Big Gate—to validate candidate changes within the graph scope before committing them. Using gpt-5.6-sol, SkillAA achieves 81.5% on SearchQA, 66.7% on LiveMath, and 91.2% on DocVQA, and records the highest mean performance across all main settings, confirming the utility of attribution‑guided graph editing and graph‑scoped validation.
Review: SkillAA provides a concrete pathway for fine‑grained skill repair on frozen models through structured graphs and attribution, offering promising directions for interpretability and safety research.