3 November 2026

Separately trained image and text models converge on shared representations without paired data

• Researchers Balestriero and LeCun report that separately trained image and text models can end up with essentially the same internal representation.
• This convergence happens without the models ever being shown matched image and text pairs.
• The authors claim the approach is more sample-efficient than next-token prediction, the method most chatbots are trained with.

How it was covered

Latent Spaceswyx & Alessio

Platonic representation hypothesis, JEPA image and text training, sample efficiency comparison, mechanism of shared views