AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

ByteFlow: Language Modeling through Adaptive Byte Compression without a Tokenizer

Chunyuan Deng, Sanket Lokegaonkar, Colin Lockard, Besnik Fetahu, Nasser Zalmout, Xian Li

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Language model pretraining and induction · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use lossy coding-rate Top-K chunking instead of cosine-similarity chunking (ablation)ByteFlow Net (600M) · Chunking ablationvalidation BPB: 0.92 -> 0.86Improved
use coding-rate chunking instead of word-boundary chunking (ablation)ByteFlow Net (600M) · Chunking ablationtask accuracy: 49.38 -> 50.89Improved
use random chunk boundaries (ablation)ByteFlow Net (600M) · Chunking ablationtask accuracy: 50.89 -> 41.34Hurt
use fixed-stride chunking (ablation)ByteFlow Net (600M) · Chunking ablationvalidation BPB: 0.86 -> 0.96Hurt
use word-boundary chunking instead of LLaMA BPE tokenization (ablation)byte-level hierarchical model (600M) · zero-shot task accuracytask accuracy: 49.15 -> 49.38No clear effect
add Canon layers to local encoder/decoder (ablation)ByteFlow Net (600M) · Zero-shot average (600M)accuracy: 49.04 -> 50.89Improved
add Canon layers to local encoder/decoder (ablation)ByteFlow Net (1.3B) · Zero-shot average (1.3B)accuracy: 61.06 -> 63.19Improved
increase global sequence length (lower compression) from 1600 to 4096 (ablation)ByteFlow Net (600M) · Zero-shot average (600M)accuracy: 48.48 -> 51.74Improved
increase global sequence length from 1600 to 4096 (ablation)ByteFlow Net (1.3B) · Zero-shot average (1.3B)accuracy: 60.91 -> 64.44Improved
replace BPE LLaMA with ByteFlow Net at 600MLLaMA · Zero-shot average (600M)accuracy: 49.15 -> 50.89Improved
replace BPE LLaMA with ByteFlow Net at 1.3BLLaMA · Zero-shot average (1.3B)accuracy: 60.15 -> 63.19Improved
replace Llama 3 with ByteFlow Net 1.3BLLM · CUTEaccuracy: 27.5 -> 51.2Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy