Verify-repair loops are a standard means for large language model (LLM) agents to correct faulty plans in code generation, mathematical reasoning, and tool use. However, when both the verifier and the repairer are noisy, repair can damage already-correct plans, leading to rising reported acceptance while true validity declines. Thus, existing methods lack a principled basis for deciding when repair should stop.
We propose VRR-Stop, a robust stopping framework for noisy verify-repair-repeat (VRR) loops. A four-parameter noise model separates verifier false acceptance and false rejection from the repair and damage behavior of the repairer. Belief filtering converts repeated verification votes into an estimate of committed validity, and the loop commits or repairs according to the sign of the true marginal gain, which requires only sign identifiability rather than accurate recovery of all parameters.
When verifier discrimination approaches zero, calibration itself fails, and estimation error can flip the stopping sign. Therefore, we pair VRR-Stop with VRR-Guard, an estimation-free fallback that replaces the incumbent candidate only under a sufficient verification margin.
In a GSM8K stress setting, VRR-Stop improves final true validity by 60.6 percentage points at an average cost of 0.72 repair rounds. Across settings, stopping reliability is governed jointly by verifier discrimination and the decision margin rather than by the absolute size of estimation error.
Blogger's Review: VRR-Stop significantly enhances the decision-making capability of LLM agents in uncertain environments by introducing belief filtering and noise modeling. Its integration with VRR-Guard provides fresh insights for future research, warranting further exploration.