MathVis-Fine Improves Multimodal Math Reasoning with Visual Dependency Training.

Multimodal Chain-of-Thought (CoT) reasoning has expanded to include scenarios involving both linguistic and visual inputs, particularly in mathematical problem-solving. However, current approaches often treat visual information as uniformly auxiliary or homogeneous, failing to capture the specific, sample-dependent relationships between text and images. This leads to two main problems: coarse-grained visual supervisory signals that don't adapt to actual visual necessity, and inaccurate training feedback due to uniform application of visual rewards. To overcome these limitations and achieve more precise multimodal reasoning, this research introduces the MathVis-Fine framework. A key component is the creation of the MathVis-Fine dataset, which augments standard mathematical problems with fine-grained visual annotations and explicit visual dependency ratings. This dataset provides the necessary granular information for more targeted training. Building on this, the framework proposes a two-stage progressive visual enhancement training paradigm. This paradigm dynamically balances answer correctness rewards with visual grounding rewards, adjusting their emphasis according to the intrinsic visual dependency level of each sample. This adaptive approach mitigates reward bias and significantly improves supervision accuracy. Extensive experiments confirm that MathVis-Fine effectively enhances visual perception progressively, leading to a more precise training framework for multimodal mathematical reasoning.