GraphSkillAA introduces a method for updating skill graphs by attributing failures to specific node fields or edges using a teacher-student setup. It employs a two-level validation: a Local Gate retests affected examples, and a Big Gate evaluates the full update pool, rolling back rejected changes. On SearchQA, LiveMath, and DocVQA with GPT-5.6-sol, it achieves 81.5%, 66.7%, and 91.2% respectively, attaining the highest mean in every main setting. This makes skill graphs locally optimizable, testable, and reversible external states.
Attribution-guided skill graphs improve targeted repairs for frozen language models
SkillAA traces execution failures to specific graph objects, edits only the implicated structure, and validates changes before committing them.
Top University
Ziqiao Shang · Ling-Yue Ge · Lan-Zhe Guo
National Key Laboratory for Novel Software Technology · Nanjing University
Research Digest··2 min read
Thread:Skill Selection for Agents
In one line
GraphSkillAA uses attribution-guided updates and two-level validation to optimize skill graphs, outperforming existing methods.
Before you read this
Three ideas this paper builds ongraph-structured skillsA directed knowledge structure where skills are linked, enabling targeted updates and efficient retrieval for AI agents.5 papers →failure attributionIdentifying which component or step in a system caused an observed failure, often using structured trace representations.140 papers →agent skillsLearned capabilities that enable AI agents to perform specialized tasks with improved effectiveness.118 papers →
Read it as
Shang, Ge, and Guo introduce SkillAA, a framework that improves external procedural skills without updating a language model’s parameters.
The authors represented when skills apply, how they execute, and how they compose as a unified graph.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.