AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Controlled text generation · 3 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 lossless vocabulary reduction instead of naive restriction to 1-byte tokensQwen2.5-3B · GSM8Kaccuracy: 0.00 -> 71.19Improved
use lossless vocabulary reduction instead of naive restriction to <=4-byte tokensFalcon3-7B · GSM8Kaccuracy: 34.72 -> 79.30Improved
use lossless vocabulary reduction instead of naive restriction to <=8-byte tokensQwen2.5-3B · GSM8Kaccuracy: 65.96 -> 72.18Improved
use lossless vocabulary reduction instead of naive restriction to <=8-byte tokensLlama3.2-3B · GSM8Kaccuracy: 26.31 -> 26.16No clear effect
reduce Falcon3-7B vocabulary to 1 byte with K-LVRFalcon3-7B · GSM8Kaccuracy: 76.65 -> 78.92No clear effect
ensemble via K-LVR over MCV instead of union vocabulary (PoE)Qwen2.5-3B + Falcon3-7B · GSM8Kaccuracy: 27.60 -> 81.12Improved
ensemble via K-LVR over MCV instead of naive restriction to MCV (PoE)Qwen2.5-3B + Falcon3-7B · MMLU-Proaccuracy: 2.07 -> 42.00Improved
ensemble with K-LVR (MCV) rather than best single model (PoE)Qwen2.5-3B + Falcon3-7B · GSM8Kaccuracy: 76.65 -> 81.12Improved
use MCV instead of 1-byte vocabulary for ensemble (PoE)Qwen2.5-3B + Falcon3-7B · GSM8Kaccuracy: 82.49 -> 81.12No clear effect
use MCV instead of 1-byte vocabulary for ensemble speedQwen2.5-3B + Falcon3-7B · GSM8Kbytes/sec: 17.03 -> 33.50Improved
use MCV instead of 1-byte vocabulary for ensemble speedOLMo2-1B + Qwen2.5-0.5B · GSM8Kbytes/sec: 25.72 -> 46.78Improved
use Union baseline with PoE instead of MoEQwen2.5-3B + Falcon3-7B · GSM8Kaccuracy: 80.74 -> 27.60Hurt

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