arXiv:2605.25820v2 Announce Type: replace 
Abstract: Diffusion-based multimodal large language models (dMLLMs) decode by iteratively predicting tokens at multiple masked positions in parallel. This turns each decoding step into a position-selection problem: the model must choose not only which predictions are reliable in isolation, but also which positions should be committed together as context for later decoding steps. Existing confidence-based decoding ranks masked positions independently and commits the top-K positions, largely ignoring whether the committed tokens provide complementary visual grounding. We identify a step-level limitation of this strategy in multimodal settings: high-confidence tokens selected in the same step can rely on overlapping visual grounding, introducing visual redundancy among the committed tokens and leaving less complementary visual grounding available for later decoding. To quantify this effect, we introduce the Visual Redundancy Index (VRI), which measures visual grounding overlap among tokens committed in parallel. To control this redundancy during decoding, we propose Visual-Redundancy-Controlled Decoding (VRCD), a training-free inference-time decoding method that uses token-to-image attention to prioritize visually complementary positions. Across diverse multimodal benchmarks, VRCD reduces visual redundancy and remaining-position entropy with modest runtime overhead. In longer decoding experiments, it also achieves relative accuracy gains of up to 18.8% on M^3CoT and 6.9% on MMBench over confidence-based decoding. Code is available at https://github.com/infiniteYuanyl/VRCD.

قدمت دراسة حديثة مفهوم فك الشيفرة المتوازية الخاضع للرقابة على التكرار البصري لنماذج اللغة متعددة الوسائط المعتمدة على الانتشار (dMLLM)، والذي يعزز فك الشيفرة من خلال معالجة التكرار البصري في اختيار الرموز. تستخدم هذه الطريقة مؤشر التكرار البصري (VRI) لتحسين اختيار الرموز في عدة مواقع مخفية، مما يضمن أن الرموز ذات الثقة العالية لا تعتمد على أساس بصري متداخل.

Un estudio reciente presentó el Decodificación Paralela Controlada por Redundancia Visual para Modelos de Lenguaje Multimodal Basados en Difusión (dMLLM), que mejora la decodificación al abordar la redundancia visual en la selección de tokens. Este método utiliza un Índice de Redundancia Visual (VRI) para optimizar la selección de tokens en múltiples posiciones enmascaradas, asegurando que los tokens de alta confianza no dependan de un anclaje visual superpuesto.

Une étude récente a introduit le décodage parallèle contrôlé par redondance visuelle pour les modèles de langage multimodaux basés sur la diffusion (dMLLM), qui améliore le décodage en s'attaquant à la redondance visuelle dans la sélection des tokens. Cette méthode utilise un indice de redondance visuelle (VRI) pour optimiser la sélection des tokens à plusieurs positions masquées, garantissant que les tokens de haute confiance ne reposent pas sur un ancrage visuel chevauchant.

A recent study introduced Visual-Redundancy-Controlled Parallel Decoding for Diffusion-Based Multimodal Large Language Models (dMLLMs), which enhances decoding by addressing visual redundancy in token selection. This method utilizes a Visual Redundancy Index (VRI) to optimize the selection of tokens at multiple masked positions, ensuring that high-confidence tokens do not rely on overlapping visual grounding.