MIRROR: Teach Vision-Language Models to Reason by Learning From the Other Modality's View

arXiv 2607.21552·medium signalXBlueskyLinkedInCopy link

Wen Ye, Yuxiao Qu and Aviral Kumar tackle the gap where VLMs fail at visual reasoning tasks that LLMs handle easily in text, proposing MIRROR — a method that transfers reasoning capability by having each modality learn from the other's view of the same problem. The premise is that the reasoning ability already exists in the language half of the model and simply doesn't route through the visual pathway. For builders shipping multimodal features, this targets the specific failure where a model reads a chart correctly but reasons about it worse than if you'd transcribed it.