The field of multimodal reasoning is moving towards more interpretable and evidence-driven models. Recent developments focus on strengthening visual grounding and reducing hallucination in Large Vision-Language Models (LVLMs). A key direction is the development of quantitative measures to evaluate the model's reliance on visual evidence, allowing for more targeted and effective refinement of model responses. Another area of innovation is the design of training-free workflows that can improve the accuracy of zero-shot vision tasks. Noteworthy papers include:

• Draft and Refine, which proposes a novel agent framework that utilizes a question-conditioned utilization metric to refine model responses with targeted feedback from external visual experts.
• Binary Verification for Zero-Shot Vision, which introduces a simple and unified workflow that emphasizes inference-time design over task-specific training, yielding significant improvements across various tasks.
• Direct Visual Grounding by Directing Attention of Visual Tokens, which proposes a novel loss function that directly supervises the attention of visual tokens, leading to improved performance on visual tasks.