Language models can do vision tasks if you give them a bridge — and some layers work without fine-tuning

Researchers found that language model parameters can be adapted to handle vision tasks, contrary to the assumption that the two modalities are fundamentally incompatible. A simple retraining stage called random label bridge training aligns language model weights to work on visual problems, and some internal layers work for vision even without additional tuning.

This is a narrow technical finding about parameter reuse, not a structural shift in how AI models work. The paper shows that language and vision parameters are more compatible than researchers thought, which could reduce the amount of training needed to build multi-modal systems. But the practical effect remains confined to the research setting — it does not change what models can do in the world, alter costs at scale, or affect deployment.