TL;DR of the paper, generated automatically.

DeepMem, a byte-level language model that compresses sequences for efficient reasoning.

• Results: beats a matched byte-level Transformer on FineWeb-Edu and PG19 language modeling.
• Cost: about 7,600x fewer counted FLOPs than full-resolution reasoning.
• Also: best macro average on FLORES, PIQA accuracy up, code competitive.