Byte Latent Transformer: Patches Scale Better Than Tokens

Authors:
  • APArtidoro Pagnoni
    Artidoro Pagnoni
  • RPRam Pasunuru
    Ram Pasunuru
  • PRPedro Rodriguez
    Pedro Rodriguez
  • JNJohn Nguyen
    John Nguyen
  • BMBenjamin Muller
    Benjamin Muller
  • MLMargaret Li
    Margaret Li
  • CZChunting Zhou
    Chunting Zhou
  • LYLili Yu
    Lili Yu
  • JWJason Weston
    Jason Weston
  • LZLuke Zettlemoyer
    Luke Zettlemoyer
  • GGGargi Ghosh
    Gargi Ghosh
  • MLMike Lewis
    Mike Lewis

Abstract
We introduce the Byte Latent Transformer (BLT), a new byte-level LLM architecture that, for the first time, matches tokenization-based LLM performance at scale with significant improvements in inference efficiency and robustness. BLT encodes bytes into dynamically sized patches, which serve as the primary units of computation. Patches are segmented based on the entropy of the next byte, allocating more compute and model capacity where increased data complexity demands it. We present the first FLOP controlled scaling study of byte-level models up to 8B parameters and 4T training bytes. Our results demonstrate the feasibility of scaling models trained on raw bytes without a fixed vocabulary. Both training and inference efficiency improve due to dynamically selecting long patches when data is predictable, along with qualitative improvements on reasoning and long tail generalization. Overall, for fixed inference costs, BLT shows significantly better scaling than tokenization-based models, by simultaneously growing both patch and model size.
All Related Versions

This publication has 10 references indexed in Scilit: