GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents

Researchers introduce GRASP, a method for improving large language model agents through controlled skill library updates that prevent performance regression. Tested across five base models on clinical benchmarks, GRASP achieves dramatic improvements (40.6% to 88.8% on MedAgentBench) while maintaining stability, outperforming existing self-improvement approaches by significant margins.

GRASP addresses a fundamental problem in autonomous AI agent development: how to iteratively improve procedural knowledge without inadvertently breaking existing capabilities. Traditional self-improvement methods accumulate guidance through natural language instructions, but lack mechanisms to detect when a fix for one task silently degrades performance elsewhere. This creates a reliability crisis in structured environments where agents must execute precise sequences of actions.

The research identifies that LLM agent failures stem primarily from procedural knowledge gaps rather than conversational limitations. By constraining improvements to a bounded skill library with strict regression validation against held-out probes, GRASP enforces a hard budget on acceptable performance loss. This gating mechanism ensures that only net-positive edits enter the library, fundamentally changing how agents can safely learn from experience.

The empirical results demonstrate substantial practical impact. On clinical benchmarks, GRASP produced 48.2 percentage point gains for gpt-oss-120b and 21-point advantages over competing baselines. The ablations reveal that the acceptance gate and regression budget drive gains more than skill generation itself—ungated skill writing provides no benefit, suggesting validation mechanisms are critical. The asymmetric transfer learning finding (stronger model skills transfer to weaker executors but not vice versa) reveals how agent capability hierarchies function and has implications for distributed AI systems.

Beyond clinical domains, GRASP's generalization demonstrates relevance across multiple structured environments, though limitations appear where action spaces become open-ended. This positions controlled self-improvement as essential infrastructure for deploying reliable autonomous agents in high-stakes domains like healthcare, where regression testing and bounded updates directly impact safety and trustworthiness.

• →GRASP prevents regression in AI agent self-improvement by gating skill library updates against performance on held-out validation probes
• →Clinical benchmark results show 48.2 percentage point improvements and 21-point advantages over existing self-improvement methods
• →Validation mechanisms matter more than skill generation itself; ungated approaches provide no benefit over baselines
• →Skill libraries transfer asymmetrically across models, with stronger models improving weaker executors beyond their self-learning capabilities
• →The approach generalizes to non-clinical environments but faces limitations in open-ended action spaces