AI Paper Atlas

Research map / Architectures and efficiency / Quantization

UniSVQ: 2-bit Unified Scalar-Vector Quantization

Haoyu Wang, Haiyan Zhao, Xingyu Yu, Zhangyang Yao, Xu Han, Zhiyuan Liu, Maosong Sun

ICML 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs

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
add block-wise fine-tuning of the affine grid (ablation)UniSVQ · 0-shot QA avg, Qwen-3-8Baccuracy: 66.99 -> 67.95Improved
initialize with random orthogonal matrix instead of D4 lattice generator (ablation)UniSVQ · 0-shot QA avg, Qwen-3-8Baccuracy: 61.00 -> 67.95Improved
apply Randomized Hadamard Transform before quantization (ablation)UniSVQ · WikiText-2, Llama-3-8Bperplexity: 8314.12 -> 20.04Improved
increase vector dimension d from 4 to 8 (no fine-tuning) (ablation)UniSVQ · 0-shot QA avg, Qwen-3-8Baccuracy: 66.99 -> 67.34No clear effect
use linear-constrained codebook on top of Hessian compensation and RHT (ablation)UniSVQ · weight-level SNR, Qwen-3-4B 2-bitSNR: -3.06 -> 8.48Improved
increase bit-width from 2 to 3 (ablation)UniSVQ · 0-shot QA avg, Llama-3-8Baccuracy: 63.42 -> 67.71Improved
replace scalar quantization with UniSVQ gridUniSVQ vs OSTQuant · WikiText-2, Qwen-3-8B 2-bitperplexity: 26.08 -> 14.82Improved
use UniSVQ instead of AQLMUniSVQ vs AQLM · WikiText-2, Qwen-3-8B 2-bitperplexity: 18.26 -> 14.82Improved
UniSVQ vs QuIP# (E8P lattice VQ)UniSVQ · 0-shot QA avg, Qwen-3-8Baccuracy: 67.55 -> 67.95Mixed
UniSVQ vs Qtip trellis-coded VQUniSVQ · 0-shot QA avg, Qwen-3-8Baccuracy: 71.08 vs 67.95Hurt
use UniSVQ kernels instead of FP16UniSVQ · Llama-3-8B throughputtok/s: 60.38 -> 101.65Improved
UniSVQ instead of Quip#/AQLM for decoding speedUniSVQ · Llama-3-8B throughputtok/s: 79.60 -> 101.65Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 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