ST-Veto: Spatio-Temporal Token Veto for Diffusion MLLMs via Taylor Prediction and Visual Grounding

TL;DR - ST-Veto is a training-free decoding method for diffusion multimodal LLMs that replaces temporally unstable or weakly image-grounded tokens. It improves multimodal reasoning accuracy by up to 9% without additional training or generation cost.

• Uses second-order Taylor prediction to identify tokens with unstable confidence across diffusion steps.
• Measures image-attention mass to filter tokens lacking strong visual grounding.
• Swaps vetoed tokens with safer candidates during iterative unmasking.
• Consistently outperforms standard decoding policies and prior VLM reasoning methods across multiple models and benchmarks.