Research map / Large language models / Language, safety and interpretability
Lossless Vocabulary Reduction for Auto-Regressive Language Models
Daiki Chijiwa, Taku Hasegawa, Kyosuke Nishida, Shin'ya Yamaguchi, Tomoya Ohba, Tamao Sakao, Susumu Takeuchi
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 lossless vocabulary reduction instead of naive restriction to 1-byte tokens | Qwen2.5-3B · GSM8K | accuracy: 0.00 -> 71.19 | Improved |
| use lossless vocabulary reduction instead of naive restriction to <=4-byte tokens | Falcon3-7B · GSM8K | accuracy: 34.72 -> 79.30 | Improved |
| use lossless vocabulary reduction instead of naive restriction to <=8-byte tokens | Qwen2.5-3B · GSM8K | accuracy: 65.96 -> 72.18 | Improved |
| use lossless vocabulary reduction instead of naive restriction to <=8-byte tokens | Llama3.2-3B · GSM8K | accuracy: 26.31 -> 26.16 | No clear effect |
| reduce Falcon3-7B vocabulary to 1 byte with K-LVR | Falcon3-7B · GSM8K | accuracy: 76.65 -> 78.92 | No clear effect |
| ensemble via K-LVR over MCV instead of union vocabulary (PoE) | Qwen2.5-3B + Falcon3-7B · GSM8K | accuracy: 27.60 -> 81.12 | Improved |
| ensemble via K-LVR over MCV instead of naive restriction to MCV (PoE) | Qwen2.5-3B + Falcon3-7B · MMLU-Pro | accuracy: 2.07 -> 42.00 | Improved |
| ensemble with K-LVR (MCV) rather than best single model (PoE) | Qwen2.5-3B + Falcon3-7B · GSM8K | accuracy: 76.65 -> 81.12 | Improved |
| use MCV instead of 1-byte vocabulary for ensemble (PoE) | Qwen2.5-3B + Falcon3-7B · GSM8K | accuracy: 82.49 -> 81.12 | No clear effect |
| use MCV instead of 1-byte vocabulary for ensemble speed | Qwen2.5-3B + Falcon3-7B · GSM8K | bytes/sec: 17.03 -> 33.50 | Improved |
| use MCV instead of 1-byte vocabulary for ensemble speed | OLMo2-1B + Qwen2.5-0.5B · GSM8K | bytes/sec: 25.72 -> 46.78 | Improved |
| use Union baseline with PoE instead of MoE | Qwen2.5-3B + Falcon3-7B · GSM8K | accuracy: 80.74 -> 27.60 | Hurt |
Closest papers by meaning
- The Foundations of Tokenization: Statistical and Computational Concerns (ICLR 2025 · similarity 0.89)
- An Analysis of Tokenization: Transformers under Markov Data (NeurIPS 2024 · similarity 0.88)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.88)
- Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling (ICML 2025 · similarity 0.87)
- Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles (ICLR 2025 · similarity 0.87)
- From Bytes to Ideas: Language Modeling with Autoregressive U-Nets (NeurIPS 2025 · similarity 0.87)
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.
