VERDICT Verifies Multimodal Reasoning with Disagreement-Aware Consensus

Multimodal large language models (LLMs) often produce reasoning chains that contain subtle errors, leading to incorrect final answers. Existing verification methods either demand expensive labeled supervision or rely on simple aggregation of scores from multiple sources, overlooking the crucial information embedded in disagreements among these sources. This research introduces VERDICT (VERification via Disagreement-Informed Coupled Thresholding), a novel training-free and domain-agnostic approach for step-wise verification of multimodal reasoning. VERDICT formalizes cross-modal disagreement among disparate verifiers as a coordination game, where agreement signals valid steps and disagreement highlights instability. By computing consensus scores through a unique closed-form solution, VERDICT enables both disagreement-aware filtering and stability-conscious ranking of reasoning steps. Evaluated across six benchmarks, VERDICT consistently improves base model performance and competes effectively with domain-specific critics that require extensive training data, demonstrating the power of leveraging cross-modal agreement for robust verification without task-specific adaptation.