Search papers, labs, and topics across Lattice.
This paper introduces VRR-Stop, a robust stopping framework designed to enhance the performance of noisy verify-repair loops in large language model agents. By employing a four-parameter noise model that distinguishes between verifier false acceptance and rejection, the method effectively estimates committed validity through belief filtering and optimally decides when to stop repairs based on the true marginal gain. In experiments on the GSM8K stress setting, VRR-Stop achieved a remarkable 60.6 percentage point increase in true validity compared to traditional fixed repair methods, while maintaining a low average repair cost of 0.72 rounds.
Stopping repairs too soon can lead to a 60% drop in true validity, but VRR-Stop ensures LLM agents know exactly when to commit or repair.
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. When both the verifier and the repairer are noisy, repair can damage already-correct plans, and reported acceptance keeps rising while true validity falls, so 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 turns 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, so we pair VRR-Stop with VRR-Guard, an estimation-free fallback that replaces the incumbent candidate only under a sufficient verification margin. On a GSM8K stress setting, VRR-Stop improves final true validity by 60.6 percentage points over fixed five-round repair 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.
