Research map / Large language models / Language, safety and interpretability
Sampling from Your Language Model One Byte at a Time
Jonathan Hayase, Alisa Liu, Noah A. Smith, Sewoong Oh
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use ByteSampler instead of no mitigation for next-character prediction | OLMo2-1B · English next-character prediction | next character accuracy: 29.490 -> 81.560 | Improved |
| use ByteSampler instead of token healing/alignment backtracking | OLMo2-1B · English next-character prediction | next character accuracy: 76.281 -> 81.560 | Improved |
| increase backtracking from 2 to 4 tokens in Token Alignment (ablation) | Token Alignment · English next-character prediction | next character accuracy: 76.281 -> 75.407 | Hurt |
| apply ByteSampler to convert the LM to character level | OLMo2-1B · English text (OLMo2 corpus) | bits per character: 6.94 -> 0.87 | Improved |
| apply ByteSampler to convert the LM to character level | Qwen3-1.7B-Base · Chinese text (MADLAD-400 zh) | bits per character: 5.47 -> 3.43 | Improved |
| use ByteSampler instead of Token Alignment backtracking | Qwen3-1.7B-Base · Chinese next-character prediction | next character accuracy: 49.6 -> 52.7 | Improved |
| use ByteSampler instead of prior exact method with prefix caching for solving PBP | OLMo2 tokenizer · PBP inference cost | overhead vs BPE (inference tokens): +2.09 -> +0.72 | Improved |
| ensemble Qwen3, OLMo2 and Llama3.2 at byte level | byte-level ensemble · seven downstream tasks | accuracy: LAMBADA 0.727 -> 0.755; TriviaQA 0.535 -> 0.526 | Mixed |
| proxy-tune Llama3.1-8B with OLMo2-1B-Instruct expert and OLMo2-1B anti-expert at byte level | Llama3.1-8B · GSM8K | accuracy: 55.3 -> 76.6 | Improved |
| proxy-tune Llama3.1-8B with OLMo2 expert pair at byte level | Llama3.1-8B · MMLU | accuracy: 27.8 -> 59.5 | Improved |
| proxy-tune Llama3.1-8B with OLMo2 expert pair at byte level | Llama3.1-8B · AlpacaEval 2 | LC winrate: 0.88 -> 33.5 | Improved |
Closest papers by meaning
- Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles (ICLR 2025 · similarity 0.90)
- From Language Models over Tokens to Language Models over Characters (ICML 2025 · similarity 0.88)
- The Foundations of Tokenization: Statistical and Computational Concerns (ICLR 2025 · similarity 0.87)
- From Bytes to Ideas: Language Modeling with Autoregressive U-Nets (NeurIPS 2025 · similarity 0.87)
- Enhancing Large Language Models through Adaptive Tokenizers (NeurIPS 2024 · similarity 0.87)
- BatchPrompt: Accomplish more with less (ICLR 2024 · similarity 0.86)
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.
