MLLMs often lack transparent reasoning, merely providing final predictions without intermediate steps or visual evidence. This paper introduces the Visual Reasoning Tracer (VRT) task and associated benchmarks (VRT-Bench, VRT-80k) to explicitly require models to localize intermediate objects in their reasoning paths, significantly enhancing model interpretability and reliability.

DraCo introduces a novel interleaved reasoning paradigm, Draft-as-CoT, for text-to-image generation that leverages both textual and visual content. This approach addresses limitations of existing methods by generating low-resolution draft images for visual planning and verification, significantly improving the generation of rare attribute combinations and overall image quality.

This paper introduces TempR1, a novel temporal-aware multi-task reinforcement learning framework designed to significantly enhance the temporal understanding capabilities of Multimodal Large Language Models (MLLMs). By integrating diverse temporal tasks and tailored reward functions, TempR1 achieves state-of-the-art performance across various video understanding benchmarks and improves generalization.