The field of multimodal reasoning and vision-language models is moving towards more robust and reliable methods for integrating visual perception with language understanding. Researchers are focusing on developing frameworks that can effectively guide reasoning and improve performance on tasks such as visual grounding and question answering. One direction is the use of self-training frameworks that jointly improve perception and reasoning, while another is the development of mechanisms for identifying the most critical modality in multimodal emotion understanding. Additionally, there is a growing interest in creating more flexible and scalable tool-based reasoning approaches that can interactively use tools to reason about visual inputs. Noteworthy papers include:

• PhotoFramer, which introduces a multi-modal composition instruction framework for providing guidance on image composition.
• See, Think, Learn, which proposes a self-training framework for enhancing multimodal reasoning ability.
• Learning What to Attend First, which presents a framework for improving the reliability of reasoning-based multimodal emotion understanding.
• Thinking with Programming Vision, which proposes a flexible and scalable code-as-tool framework for robust tool-based reasoning.
• Visual Reasoning Tracer, which introduces a benchmark for evaluating visual reasoning and requires models to explicitly predict the intermediate objects that form the reasoning path.