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

Sep 15, 2026

Uploaded by Patrick

LLM agents operating in structured environments can fail in ways tied to execution procedures rather than conversational quality, causing reliability breakdowns that remain uncorrected. Prior self-improvement approaches add natural-language guidance without verifying that new edits preserve previously correct behavior, so fixes may silently regress other trajectories. GRASP treats improvement as gated, regression-aware edits to a bounded skill library and accepts candidates only under a hard regression budget, evaluated on FHIR-based clinical benchmarks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages26

Reading Time1 h 6 min

AccessFree to download

Introduction

Structured environments and operational reliability

Limits of prior self-improvement methods

GRASP approach and gated regression-aware editing

Evaluation setup and results

GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents Johannes Moll1 , Jean-Philippe Corbeil2 , Jiazhen Pan1 , Martin Hadamitzky3 , Daniel Rueckert1 , Lisa Adams4,†, Keno Bressem3,4,† 1Technical University of Munich and TUM University Hospital, 2Microsoft Healthcare & Life Sciences, 3Department of Cardiovascular Radiology and Nuclear Medicine, German Heart Center, TUM University Hospital, 4Department of Diagnostic and Interventional Radiology, Klinikum rechts der Isar, TUM University Hospital Correspondence: [johannes.moll@tum.de](johannes.moll@tum.de) †Equal senior authorship Project page: [https://jomoll.github.io/GRASP](https://jomoll.github.io/GRASP) Code: [https://github.com/jomoll/GRASP](https://github.com/jomoll/GRASP) arXiv : 2605 .29668v 3 [cs .AI] 2 1 Aug 2026 Abstract LLM agents acting in structured environments fail in operational rather than conversationalways, and reliability depends on procedural knowledge of the environment. Prior self-improvement methods accumulate naturallanguage guidance without checking that each new item preserves previously correct behavior, so a note that fixes one trajectory can silently regress another. We introduce GRASP (Gated Regression-Aware Skill Proposer), which treats agent improvement as a sequence of edits toa bounded skill library, admitting each candidate only if it produces a net improvement on a balanced held-out probe under a hard regression budget. We evaluate GRASP across five base models on two FHIR-based clinical benchmarks, which score procedural reliability against FHIR state rather than clinical correctness or patient outcomes. On MedAgentBench, GRASP lifts gpt-oss-120b from 40.6% to 88.8%, exceeds the strongest of five selfimprovement baselines by 21.0 points, and improves every other base model by 17.2 to 40.3 points. Ablations attribute the gain to comparative proposal generation, the acceptance gate, and the hard regression budget rather than to skill writing itself, which without validation is no better than using no skills. Granting the same acceptance gate to all five baselines lifts each of them in-domain and none of them out of distribution, isolating the gain to the gate applied to a bounded, editable library rather than to held-out validation itself. The mechanism helps in non-clinical environments where tasks recur with verifiable structure and is flat where the action space is open-ended. Frozen libraries transfer across models and across benchmarks that share a tool-calling convention and degrade under interface mismatch. 1 Introduction Structured environments such as electronic health records, databases, and web interfaces impose strict operational rules on LLM agents. Violations of these rules produce recurring errors that often go uncorrected. A clinical assistant working with an EHR, for instance, must retrieve the right resources, apply temporal eligibility criteria, avoid duplicate medication orders, and verify that changes were written (Jiang et al., 2025 ; Moll et al., 2026) . These failures reflect the structure of the environment rather than the particular case, so reliability depends on procedural knowledge of how to act ina specific environment, not on general language ability (Hsiao et al., 2025 ; Roig, 2025) . To acquire this knowledge, prior selfimprovement methods accumulate naturallanguage guidance at inference, either as reflections on individual failures (Shinn et al., 2023) or as rules extracted across many (Zhao et al., 2024), admitting each item as soon as it is produced (Appendix A) . Without a check against existing behavior, a note that fixes one trajectory can silently regress another, and overtime monotonic accumulation dilutes task-relevant content and degrades the agent’s context (Shi et al., 2023 ; Liu et al., 2024a) . Avoiding both failures requires held-out validation of each candidate against previously correct behavior, together with a bounded library that supports modification

FAQs

Why do LLM agents fail in structured environments even when they generate good text?

Failures stem from operational rules imposed by environments like EHRs, databases, and web interfaces. Violations recur and often remain uncorrected, so reliability depends on procedural knowledge of how to act in that environment.

What problem does GRASP address in prior self-improvement methods?

Prior methods accumulate natural-language guidance without checking whether each new item preserves previously correct behavior. A note that fixes one trajectory can silently regress another, and monotonic accumulation can dilute task-relevant context.

How does GRASP decide whether a proposed skill edit should be accepted?

GRASP proposes edits to a bounded skill library and re-runs each candidate on a balanced held-out probe. It accepts the edit only if it yields a net improvement, fixing more failures than it causes regressions under a hard regression budget.

Sep 15, 2026

Uploaded by Patrick