GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents
GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents
2Microsoft Healthcare & Life Sciences
†Equal senior authorship
Abstract
LLM agents acting in structured environments fail in operational rather than conversational ways, and reliability depends on procedural knowledge of the environment. Prior self-improvement methods accumulate natural-language 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 to a 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 (gpt-oss-120b, DeepSeek V4 Flash, Gemini 3.1 Flash Lite, GPT-4.1, GPT-5.4) on two FHIR-based clinical benchmarks. On MedAgentBench, GRASP lifts gpt-oss-120b from 40.6% to 88.8%, exceeds the strongest of five self-improvement 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. The mechanism generalizes beyond the clinical domain, improving agents on three of four non-clinical environments, and frozen libraries transfer across models. Skills from a stronger model improve weaker executors beyond what they learn for themselves while the reverse does not, an asymmetry that no ungated baseline reproduces.
How GRASP Works
GRASP maintains a small, versioned library of behavioral skills injected into the agent's context. After each training batch it proposes edits and admits them only if they pass a regression-aware probe. The example below traces a single skill being learned on MedAgentBench.
Extend GRASP
The codebase is designed to be extended. You can plug in your own self-improvement method and benchmark it against GRASP on the same tasks, or deploy GRASP on a new environment by implementing a single interface.
Results
Clinical Benchmarks
Test accuracy on MedAgentBench across five base models. GRASP is the strongest method on every model, with gains of +17.2 to +48.2 points over no-skills and +9.4 to +21.0 points over the strongest self-improvement baseline.
| Model | No Skills | Best Baseline | GRASP (ours) | Gain vs. No Skills | Gain vs. Best Baseline |
|---|---|---|---|---|---|
| gpt-oss-120b | 40.6 ±3.9 | 67.8 (Evo-MedAgent) | 88.8 ±5.8 | +48.2 | +21.0 |
| GPT-5.4 (low) | 45.1 ±1.1 | 47.4 (Batch Memory) | 85.4 ±10.4 | +40.3 | +38.0 |
| GPT-4.1 | 45.5 ±0.5 | 47.4 (Batch Memory) | 84.9 ±0.9 | +39.4 | +37.5 |
| DeepSeek V4 Flash | 47.7 ±2.1 | 55.9 (SkillX) | 70.0 ±5.5 | +22.3 | +14.1 |
| Gemini 3.1 Flash Lite | 54.2 ±1.5 | 55.2 (Evo-MedAgent) | 71.4 ±1.8 | +17.2 | +16.2 |
Non-Clinical Environments
GRASP on four AgentBench environments (gpt-oss-120b). The mechanism generalizes where tasks recur with verifiable structure, and is flat only where the action space is open-ended.
| Benchmark | No Skills | GRASP | Gain |
|---|---|---|---|
| ALFWorld | 23.3 ±2.9 | 51.7 ±7.6 | +28.4 |
| WebShop | 20.7 ±1.2 | 41.3 ±8.1 | +20.6 |
| DBBench | 65.6 ±3.9 | 70.6 ±1.0 | +5.0 |
| OS Interaction | 48.6 ±2.9 | 49.5 ±1.7 | +0.9 |
Cross-Model Skill Transfer
Frozen GRASP skill libraries applied across models on MedAgentBench (gpt-oss-120b). Skills written by a stronger model improve weaker executors beyond what they learn for themselves, but the reverse does not hold, an asymmetry no ungated baseline reproduces.
| Skill source | Split | Executor: gpt-oss-120b | Executor: Gemini 3.1 | Executor: GPT-5.4 |
|---|---|---|---|---|
| None (baseline) | Test | 40.6 ±3.9 | 54.2 ±1.5 | 45.1 ±1.1 |
| OOD | 8.7 ±3.2 | 18.1 ±3.5 | 1.7 ±1.6 | |
| gpt-oss-120b | Test | 88.8 ±5.8 | 79.7 ±7.8 | 72.4 ±15.8 |
| OOD | 56.3 ±14.5 | 36.1 ±14.0 | 51.7 ±2.9 | |
| Gemini 3.1 | Test | 65.6 ±8.3 | 71.4 ±1.8 | 69.3 ±6.5 |
| OOD | 57.8 ±31.9 | 41.7 ±20.9 | 58.9 ±37.5 | |
| GPT-5.4 (low) | Test | 76.0 ±7.7 | 76.6 ±10.9 | 85.4 ±10.4 |
| OOD | 77.8 ±1.0 | 71.1 ±1.9 | 80.6 ±6.7 |
Ablations
MedAgentBench (gpt-oss-120b, 3 seeds). Removing the acceptance gate entirely (applying every edit unconditionally) falls to the no-skills baseline. The matched-compute variants confirm the gain is in the gate's decision, not the compute it spends.
| Method variant | Val* | Test |
|---|---|---|
| GRASP (full) | 86.0 ±4.4 | 88.8 ±5.8 |
| w/o failure grouping | 82.1 ±5.2 | 84.4 ±3.1 |
| w/o regression budget | 81.2 ±6.5 | 81.8 ±1.8 |
| fixes-only selection | 78.8 ±14.4 | 80.2 ±13.7 |
| append-only | 73.3 ±8.3 | 80.2 ±9.4 |
| w/o acceptance gate (K=4) | 65.4 ±9.7 | 63.5 ±3.9 |
| w/o acceptance gate (K=1) | 38.8 ±8.2 | 40.1 ±11.3 |
| matched compute (proposer) | 67.1 ±14.8 | 70.8 ±14.0 |
| matched compute (random) | 62.1 ±7.5 | 67.2 ±10.2 |
Learned Skill Library
13 representative skills learned by GRASP across clinical FHIR and non-clinical environments. Each was admitted by the regression gate, fixing more failures than regressions on the held-out probe.
BibTeX
@article{moll2026grasp,
title = {GRASP: Gated Regression-Aware Skill Proposer for Self-Improving LLM Agents},
author = {Moll, Johannes and Corbeil, Jean-Philippe and Pan, Jiazhen and Hadamitzky, Martin and Rueckert, Daniel and Adams, Lisa and Bressem, Keno},
journal={arXiv preprint arXiv:2605.29668},
year={2026}
}