Multimodal large language models can process both text and images. Some models use internal visual states, mental images, to reason about problems. But do these visual states actually help? The authors tested several models by removing or changing those visual states. They found that such changes barely affected the model's final answers, meaning the visual states were decorative, not functional.
The study introduces a training method called AnchorSim. It trains models to rely on their generated visual states. AnchorSim forces the model to use visual information by comparing its predictions with and without those visual states.
On nine reasoning benchmarks, AnchorSim improved performance by 13.5 points. The key finding is that visual states now matter: changes to them meaningfully alter predictions. This shows that making visual states consequential, not just present, is crucial for true multimodal reasoning.