VERDICT is a training-free, domain-agnostic approach for step-wise verification of multimodal reasoning chains. It treats the disagreement between multiple frozen verifiers as a signal to identify invalid reasoning steps through a consensus-based coordination game.
HOW THIS AFFECTS YOU
●
builderYou can improve the reliability of multimodal agent outputs by implementing disagreement-aware consensus checks.
●
researcherYou can verify multimodal reasoning steps without the need for expensive, task-specific labeled supervision.