AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Proteus: Lookup-Free Trellis-Coded Quantization by Lattice-Breaking Compute Codes for 2-Bit LLMs

Zhengwu Yang, Xunchao Li, Ke Cheng, Kunlong Liu, Jianfeng Yang, Haoshuang Wang, Kaipeng Deng, Qingqing Dang, Yanlin Sha, Yanjun Ma, Dianhai Yu

ICML 2026 · Architectures and efficiency › Quantization · approach: KV cache quantization

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
replace QTIP-HYB lookup codebook with compute-only MUL-BAL generator (ablation)Llama2-7B 2-bit · WikiText2perplexity: 5.9100 -> 5.8792Improved
add Stage I and Stage II (layer-wise generator selection and channel scales) (ablation)MUL-BAL · WikiText2perplexity: 5.8792 -> 5.8594Improved
add Stage III global few-shot relaxation (ablation)MUL-BAL + Stage I + II · WikiText2perplexity: 5.8594 -> 5.8521Improved
use MUL-BAL instead of QTIP-1MAD generator with Stages I-III (ablation)Proteus · WikiText2perplexity: 5.8794 -> 5.8521Improved
use QTIP-1MAD plus Stages I-III instead of QTIP-HYB (ablation)Llama2-7B 2-bit · WikiText2perplexity: 5.9100 -> 5.8794Improved
use Proteus instead of QTIPLlama 2-70B 2-bit · WikiText2perplexity: 3.78 -> 3.69Improved
use Proteus instead of QTIPLlama 2-13B 2-bit · C4perplexity: 6.99 -> 6.82Improved
use Proteus instead of QTIP at batch size 1Llama-2-7B W2 decoding · decode throughputtokens/s: 159 -> 161No clear effect
use Proteus instead of QTIP at batch size 16Llama-2-7B W2 decoding · decode throughputtokens/s: 2126 -> 2011Hurt
use Proteus instead of QTIP at batch size 32Llama-2-7B W2 decoding · decode throughputtokens/s: 2797 -> 2940Improved
use Proteus instead of QTIP at batch size 64Llama-2-7B W2 decoding · decode throughputtokens/s: 3605 -> 3602No clear effect
use Proteus instead of QTIP for 2-bit decoding bandwidthLlama 2-7B · observed memory bandwidthGB/s: 710 -> 572Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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