Can a compact text encoder learn representations that agree with vision more closely than those of a much larger language model? This preprint investigates that question through a conditional mathematical theorem and experiments comparing Joint-Embedding Predictive Architectures (JEPAs) with next-token prediction. Its central proposal is to train on different descriptions of the same scene, encouraging models to preserve shared information. A 109-million-parameter text JEPA achieves stronger visual alignment than the evaluated language models, including models with up to 72 billion parameters. The reported gains concern representation geometry, retrieval and selected classification tasks. The paper also combines a JEPA with an LLM decoder to generate image descriptions. This publication examines those contributions alongside the assumptions of the theorem, the meaning of the reported sample-efficiency advantage, the role of generated training captions and the limits of comparisons with general-purpose language models. The findings suggest a practical direction for representation learning while leaving broader claims about reasoning, planning and autonomous world models open.
aiweekly.co.in
“Where Machines Make Headlines.”