Researchers have identified a significant disparity in the reasoning capabilities of vision-language models (VLMs) compared to large language models (LLMs), particularly in visual reasoning tasks such as geometry problems. VLMs often struggle to solve problems presented in diagram form, even when the equivalent text or combined diagram+text views are provided. Interestingly, these different views can elicit distinct behaviors from the model, with success in one view not guaranteeing success in another. This limitation is addressed through the introduction of MIRROR, a novel approach that enables VLMs to learn from multiple views, enhancing their multi-modal reasoning capabilities. By leveraging this technique, VLMs can improve their performance on visual reasoning tasks, bringing them closer to the capabilities of LLMs1. This development matters to practitioners as it has significant implications for the development of more robust and versatile AI models, which can impact various aspects of society, including policy, security, and workforce dynamics.