Byte-level language models process text as raw bytes rather than tokens. The Byte Latent Transformer (BLT) is one such model: it runs a large global model once per patch, and decides where to start a new patch based on a small model's prediction of next-byte entropy. If entropy is high, a patch begins.

A new arXiv preprint (2610.22100) offers a compression perspective on this design. The authors argue that the decision of where to start patches can be viewed as a compression problem, which may clarify where compute should be spent. The abstract notes that this rule rarely starts a patch at in-line computed locations, though the full details are not in the abstract.

The paper is available on arXiv under ID 2610.22100. As with any preprint, the findings have not yet been peer-reviewed. The abstract itself is truncated, so the full implications of the compression view are not yet clear from the announcement.