Visual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models

Researchers proposed a training-free decoding method for diffusion-based multimodal LLMs that reduces redundant visual grounding during parallel token selection.

The paper shows that confidence-based parallel decoding can overuse the same visual cues, and introduces the Visual Redundancy Index to measure this overlap.

Their Visual-Redundancy-Controlled Decoding (VRCD) prioritizes complementary token positions at inference time to better cover the image.

The approach improves multimodal benchmark accuracy, including M3CoT and MMBench, with only modest runtime overhead.