ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding
2026-07-20 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors study a new way to improve reasoning in Vision Language Models that generate answers step-by-step but can be slow and make mistakes. They focus on diffusion-based models that consider all parts of the answer at once, allowing easier correction and faster reasoning. Their method, called Spatio-Temporal Token Veto (ST-Veto), identifies uncertain or poorly supported parts of the answer and replaces them to ensure better quality. They tested ST-Veto on several models and tasks, finding that it boosts accuracy without extra training or slower generation. Overall, the authors show how better token selection in these models leads to more reliable reasoning.
Vision Language ModelsChain-of-Thought promptingDiffusion modelsMultimodal reasoningToken vetoConfidence dynamicsImage-attentionIterative refinementDecoding policies
Authors
Keuntae Kim, Beomseok Lee, Hyunwoo Kim, Yong Suk Choi
Abstract
Vision Language Models (VLMs) achieve strong reasoning with Chain-of-Thought (CoT) prompting but incur high sequential-generation cost, error accumulation, and limited self-correction. Diffusion Multimodal Large Language Models (dMLLMs) unmask tokens in an order-agnostic process, improving efficiency and enabling iterative refinement, yet their reasoning and how to enhance it remain underexplored. We propose a training-free method, Spatio-Temporal Token Veto (ST-Veto), which leverages the ability to observe all token positions at each diffusion step. Rather than relying only on current-step confidence, ST-Veto vetoes temporally unstable tokens via second-order Taylor prediction of confidence dynamics and filters weakly grounded tokens using image-attention mass, swapping them with safer candidates. Across multiple dMLLMs and multimodal reasoning benchmarks, ST-Veto consistently outperforms standard decoding policies and prior VLM reasoning methods, improving accuracy by up to 9% with no additional training or generation cost. Analyses show that ST-Veto steers generation toward higher-confidence, better-grounded paths.