Temporal Dynamics in Video Grounding

Background: Multimodal Diffusion Large Language Models (MDLLMs) using parallel masked decoding lack explicit mechanisms to verify visual support during token commitment, leading to generation of content not present in the image.

Question / Future Work: Extending the entropy-based consensus mechanism to explicitly model motion dynamics and enforce cross-frame temporal consistency for video-based parallel decoding remains an important research direction for improving temporal grounding.