SkillMentor: LLM Agent Self-Evolution via Learning Blind-Spot Diagnosis

2026-07-29

AI Summary

This work addresses a critical limitation in existing agent self-evolution methods, which typically assume failures are already identified and focus solely on repair, thereby neglecting the agent’s capacity for proactive diagnosis of its own blind spots—i.e., unknown deficiencies. The paper introduces, for the first time, blind spot diagnosis as a distinct and learnable capability. It proposes a reinforcement learning–based Mentor policy that autonomously generates diagnostic tasks, identifies recurring failure patterns, and distills them into reusable corrective skills, enabling purely diagnosis-driven self-evolution without requiring updates to the executor’s weights or human annotations. Evaluated on the AppWorld and BFCLv3 benchmarks, the approach improves executor performance by 44.2% on average, demonstrating the learnability and efficacy of explicit blind spot diagnosis.

Abstract

Agent self-evolution has primarily focused on learning how to act, while overlooking an equally important capability: learning to discover what an agent does not know. Existing approaches typically assume that failure discovery is given, focusing on how to repair failures once they are identified. We ask whether blind-spot diagnosis itself can be learned. We thus study diagnosis as an agent capability separate from execution, and exclude two alternative sources of progress: executor adaptation and human supervision. Under these constraints, performance cannot improve through executor updates or annotated examples, forcing all improvements to originate from the learned diagnostic capability. We propose SkillMentor, which trains a Mentor policy via reinforcement learning to generate diagnostic tasks, identify recurrent failure modes, and curate them into reusable corrective skills. Across AppWorld and BFCLv3, SkillMentor improves executor performance by an average of 44.2%. These results suggest that blind-spot diagnosis is a learnable capability, enabling self-evolution without updating executor weights or relying on human-curated data.

Problem

• blind-spot diagnosis
• agent self-evolution
• failure discovery
• diagnostic capability
• self-improvement

Innovation

• blind-spot diagnosis
• agent self-evolution
• reinforcement learning
• diagnostic task generation
• executor-agnostic learning