Language agents improve skills by tracking and fixing errors iteratively
RESKILL: Explicit Failure Attribution and Structured Repair for Interactive Language Agents
Computation and Language
Summary
Language agents often use skills to perform tasks, but when they fail, it can be hard to fix the problem clearly and consistently. The authors present RESKILL, a method that keeps track of why skills fail and tests fixes step by step, learning which repairs work best. This method helps the agents get better at their tasks over time by using structured repair steps instead of random guesses. They tested RESKILL on two text-based benchmarks and found it improved success rates compared to simpler repair methods.
What this means in practice
- •For interactive ai developers: Improve performance of language agents by systematically diagnosing and fixing skill failures through multiple repair rounds.
- •For game ai designers: Enhance text-based game agents by explicitly linking failures to specific skill patches and refining them iteratively for better in-game task success.
Authors
Mengyi Deng, Xin Li, Duyi Pan, Zilin Wang, Zhiwei Li, Zhijiang Guo, Wei Wang
Abstract
Language agents increasingly rely on reusable skills, but post-failure repair is often handled by opaque one-shot reflection: a model generates a skill patch without explicitly maintaining how failure explanations relate to candidate repairs or how unsuccessful retests should influence later edits. We introduce RESKILL, a structured repair framework that maintains an explicit repair state across repair rounds. Given a failed rollout, the framework links failure hypotheses to candidate skill patches, selects local repairs through coverage-based attribution, retests the edited skill set in the environment, and uses retest outcomes to guide subsequent repair updates. The language model supplies structured repair factors, while the repair procedure records them, compares local skill patches by how well they address active failure explanations, and carries unsuccessful retest outcomes into later repair rounds. We evaluate RESKILL on ALFWorld and TextCraft across three model sizes under fixed repair budgets. RESKILL obtains the strongest final success in all six benchmark-model settings, improving average final success by 3.7 percentage points over direct repair and 3.3 points over hypothesis-conditioned repair. These results suggest that explicit attribution alone is insufficient; durable improvement emerges when attribution is integrated with repair selection and persistent retest-conditioned update.