Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Zheng, Lin, Bashlovkina, Vasilisa, Dozat, Timothy, Garrette, Dan, Rimell, Laura, Maynez, Joshua
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.09630
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910207252103168
author Zheng, Lin
Bashlovkina, Vasilisa
Dozat, Timothy
Garrette, Dan
Rimell, Laura
Maynez, Joshua
author_facet Zheng, Lin
Bashlovkina, Vasilisa
Dozat, Timothy
Garrette, Dan
Rimell, Laura
Maynez, Joshua
contents Tokenizer-free language models eliminate the tokenizer step of the language modeling pipeline by operating directly on bytes; patch-based variants further aggregate contiguous byte spans into patches for efficiency. However, the average patch size chosen at the model design stage governs a tight trade-off: larger patches reduce compute and KV-cache footprint, but degrade modeling quality. We trace this trade-off to patch lag: until a patch is fully observed, byte predictions within it must rely on a stale representation from the previous patch to preserve causality; this lag widens as patches grow larger. We introduce Scratchpad Patching (SP), which inserts transient scratchpads inside each patch to aggregate the bytes seen so far and refresh patch-level context for subsequent predictions. SP triggers scratchpads using next-byte prediction entropy, selectively allocating compute to information-dense regions and enabling post-hoc adjustment of inference-time compute. Across experiments on natural language and code, SP improves model quality at the same patch size; for example, even at $16$ bytes per patch, SP-augmented models match or closely approach the byte-level baseline on downstream evaluations while using a $16\times$ smaller KV cache over patches and $3$-$4\times$ less inference compute.
format Preprint
id arxiv_https___arxiv_org_abs_2605_09630
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models
Zheng, Lin
Bashlovkina, Vasilisa
Dozat, Timothy
Garrette, Dan
Rimell, Laura
Maynez, Joshua
Computation and Language
Machine Learning
Tokenizer-free language models eliminate the tokenizer step of the language modeling pipeline by operating directly on bytes; patch-based variants further aggregate contiguous byte spans into patches for efficiency. However, the average patch size chosen at the model design stage governs a tight trade-off: larger patches reduce compute and KV-cache footprint, but degrade modeling quality. We trace this trade-off to patch lag: until a patch is fully observed, byte predictions within it must rely on a stale representation from the previous patch to preserve causality; this lag widens as patches grow larger. We introduce Scratchpad Patching (SP), which inserts transient scratchpads inside each patch to aggregate the bytes seen so far and refresh patch-level context for subsequent predictions. SP triggers scratchpads using next-byte prediction entropy, selectively allocating compute to information-dense regions and enabling post-hoc adjustment of inference-time compute. Across experiments on natural language and code, SP improves model quality at the same patch size; for example, even at $16$ bytes per patch, SP-augmented models match or closely approach the byte-level baseline on downstream evaluations while using a $16\times$ smaller KV cache over patches and $3$-$4\times$ less inference compute.
title Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2605.09630