Multimodal models improve reasoning by adding a step between seeing and thinking
MM-OPD: Towards One More Bottleneck Between Perception and Reasoning
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Usually, AI models look at images and immediately try to answer questions about them. The authors found that if the model first converts the image into words or symbols, it actually answers questions better. This means there is a hidden problem: the model struggles to pick the right visual details when moving from seeing the image to reasoning about it. To fix this, the authors teach the AI to focus better on important visual clues using the symbol-based approach as guidance, which improves performance on many reasoning tasks.
What this means in practice
- •For multimodal ai developers: Improve AI systems that answer complex questions about images by enhancing how visual information is selected for reasoning.
- •For document analysis teams: Enhance understanding of charts and documents by directing AI attention to correct visual evidence during reasoning.
Authors
Jintao Tong, Yujing Lou, Zhanming Shen, Jiaqi Gu, Lubin Fan, Ruixuan Li, Yue Wu, Jieping Ye, Yixiong Zou
Abstract
Recent multimodal large language models (MLLMs) advance visual reasoning by strengthening both perception and reasoning, implicitly assuming a process that transitions seamlessly from perception to reasoning. However, we observe a counterintuitive phenomenon that challenges this assumption: holding the model, question, and decoding fixed, we replace images with their caption or code representations (symbolic views), which seems to be redundant given the clear image structures, but the performance surprisingly improves by 10.2% to 23.6% across model scales and datasets. We term this performance gap as the Symbolic Visual Gap and then take a closer look at it. Through experiments, we find that although the visual evidence can already appear in the reasoning trace for the image-input model, the symbolic-view-input model shows much higher attention to the correct evidence than the image-input model. This suggests that despite good capabilities from current works in perception and reasoning themselves, another bottleneck exists between perception and reasoning in selecting perceived visual information as appropriate evidence for subsequent reasoning. To handle this bottleneck, since the symbolic view steers attention toward correct evidence and is readily obtained at scale, it provides supervision for evidence selection without manually labeled evidence. Building on this, we introduce MM-OPD, a multimodal on-policy self-distillation framework for symbolic-to-visual correction that transfers guidance from symbolic-conditioned behavior to the image-conditioned policy through residual token-level targets, steering the model toward correct visual evidence. Experiments across benchmarks and model scales show that MM-OPD improves a broad range of multimodal abilities, with gains in visual perception, chart and document understanding, mathematical reasoning, and general VQA.