The paper identifies a failure mode called spatial lexical bias in MLLMs, where adding a spatial word to options biases the model's choice, and demonstrates that this failure originates primarily from…

This paper systematically analyzes how different architectural components of Large Vision-Language Models (LVLMs) contribute to hallucination robustness, finding that joint enhancement of visual fidel…

The paper proposes VLM3, a simple, scalable method that demonstrates standard Vision Language Models (VLMs) can natively learn 3D understanding by focusing on architectural simplicity and specific dat…

The paper proposes CSMR, a cognitive scheduling framework that allows a language model to dynamically decide when to acquire task-relevant visual evidence, significantly improving multimodal reasoning…

dMoE proposes a block-level Mixture-of-Experts (MoE) framework for Diffusion Large Language Models (dLLMs) that aggregates token-level expert distributions into a unified block-level distribution, sig…

The paper introduces Partial Information Decomposition (PID) to quantitatively separate unique, redundant, and synergistic contributions of different modalities (e.g., vision, language) in multimodal…

The paper introduces MLLM-Microscope, a system that analyzes the internal structure of multimodal large language models (MLLMs), finding that modality fusion significantly impacts the linearity and di…

The paper proposes a unified framework that decouples long-video reasoning into semantic and visual evidence, significantly improving performance on the HD-EPIC VQA Challenge.

The paper introduces a multilingual benchmark (MentalMap) to test if LLMs build internal spatial world models from text, finding a universal 'L3 reasoning cliff' suggesting that text-only working memo…

The paper introduces PAR3D, a unified part-aware 3D-MLLM framework, to enhance 3D scene understanding by enabling models to reason about and ground both whole objects and their fine-grained parts.

The paper proposes BRACS, a training-free steering framework that adaptively corrects visual grounding failures in large vision-language models, significantly reducing object hallucination without sac…

The paper analyzes token reduction for efficient unified VLM training, finding that while task-specific acceleration saves computation, it destroys the mutual performance gains achieved through joint…

The paper introduces EASE, a method that enhances multimodal Reinforcement Learning with Verifiable Rewards (RLVR) by providing spatial attention supervision anchored to visual evidence, significantly…

The paper evaluates the performance of Vision-Language Models (VLMs) in a collaborative dialogue task requiring spatial reconstruction, finding that while detailed text representations improve results…

The paper introduces UniKE, a benchmark showing that successful knowledge edits in text-only multimodal models do not reliably transfer to image generation, revealing a significant modality gap.

VidPrism introduces a novel heterogeneous Mixture-of-Experts framework that specializes temporal processing by dividing labor among experts, achieving state-of-the-art performance in image-to-video tr…

The paper proposes Visual Gradient Steering (VGS), a method that decomposes the distillation loss into language and visual components and steers the optimization to prioritize visual grounding, signif…

ROVER is a lightweight, learnable plugin that efficiently routes and integrates object-centric visual evidence across multiple images and objects, significantly improving performance on grounded multi…

The paper introduces RA-MoE, a novel fine-tuning framework that leverages the internal routing structure of Mixture-of-Experts (MoE) models to improve performance on multilingual downstream tasks by a…