AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Proxy Compression for Language Modeling

Lin Zheng, Xinyu Li, Qian Liu, Xiachong Feng, Lingpeng Kong

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Language model pretraining and induction · 0 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
train with proxy (tokenizer) compression mixed with raw bytes instead of pure byte-levelEvaByte 14B · HumanEval-Pluspass@1: 24.4 -> 30.5Improved
train with proxy (neural) compression instead of pure byte-levelEvaByte 14B · MBPP-Pluspass@1: 42.1 -> 49.2Improved
proxy (tokenizer) vs tokenizer-based model at 0.5BEvaByte 0.5B · HumanEval-Pluspass@1: 17.7 -> 12.2Hurt
proxy (tokenizer) vs tokenizer-based model at 14BEvaByte 14B · MBPP-Pluspass@1: 48.1 -> 49.5Improved
longer training (320B symbols) with neural proxy vs byte-levelEvaByte 7B · HumanEval-Pluspass@1: 14.6 -> 21.3Improved
add proxy (tokenizer) compression vs byte-level at 1.5B, 320B symbolsEvaByte 1.5B · HumanEval-Pluspass@1: 9.1 -> 12.8Improved
add always-on translation pairing instead of no pairs (tokenizer proxy) (ablation)Proxy (Tokenizer) 1.5B · HumanEval-Plus oracle translationoracle-translation pass@1: 45.7 -> 96.3Improved
use pairing in warm-up only instead of always-on (tokenizer proxy) (ablation)Proxy (Tokenizer) 1.5B · HumanEval-Pluspass@1: 17.0 -> 20.7Improved
warm-up-only pairing vs no pairs (neural proxy) (ablation)Proxy (Neural) 1.5B · HumanEval-Pluspass@1: 14.6 -> 18.9Improved
mix gzip-compressed data into training (increase gzip proportion) (ablation)EvaByte 1.5B · HumanEval-Pluspass@1Hurt
use proxy (tokenizer) rather than tokenizer-based modelEvaByte 7B · ReCode HumanEvalrobust pass@1: 14.9 -> 19.1Improved
use neural proxy instead of byte-levelEvaByte 7B · ReCode HumanEvalrobust pass@1: 18.7 -> 19.8Improved

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