FAILURE2SKILL: LEARNING RESIDUAL SKILLS FROM AGENT FAILURES FOR INTERACTIVE TOOL USE
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Large language model agents can often repair a single mistake when given feedback, yet they rarely turn that mistake into reusable operational knowledge. We study a focused setting for self-improving tool-use agents: after an agent equipped with a current skill library still fails, can the residual failure be compressed into a new skill that transfers to related tasks? We propose Failure2Skill, a failure-driven skill induction pipeline that diagnoses failed trajectories, clusters causal residuals, and writes compact, guarded skill cards for later retrieval. Unlike trace-combination baselines that summarize successful and failed runs together, Failure2Skill treats failures as targeted evidence about missing preconditions, API response shapes, exception paths, and state-update constraints. On a 54-task AppWorld residual- family evaluation using a DeepSeek Chat code agent, Failure2Skill solves 54/54 tasks, while the best combined success/failure skill baseline solves 37/54. Across 18 task families, Failure2Skill wins 8, ties 10, and loses 0; its average verifier pass fraction is 1.000 versus 0.843 for the best combined baseline. Ablations show that removing individual residual rules often collapses final success, including payment-card retry, relation filtering, receipt attachment, note-format preservation, and comment parsing. These results suggest that failed trajectories are not merely negative examples: when localized, they are high-density supervision for building reusable agent skills
- Primary area:
- learning theory
- Keywords:
- LLM Agent, Continue Learning