AI Paper Atlas

Research map / Architectures and efficiency / Quantization

TwinQuant: Learnable Subspace Decomposition for 4-Bit LLM Quantization

Haodong WANG, Junjie Liu, Zicong Hong, Qianli Liu, Jian Lin, Song Guo, Xu Chen

ICML 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of LLMs · 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
add SVD low-rank decomposition to naive W4A4 (ablation)Naive 4-bit · LLaMA3-8B W4A40-shot avg accuracy: 41.7 -> 61.1Improved
add SVD low-rank decomposition to naive W4A4 (ablation)Naive 4-bit · LLaMA3-8B W4A4WikiText2 perplexity: 91.6 -> 19.6Improved
add SVD low-rank decomposition to naive W4A4 (ablation)Naive 4-bit · Qwen3-8B W4A40-shot avg accuracy: 41.5 -> 62.8Improved
replace SVD low-rank with fixed Hadamard transform in decomposed branches (ablation)+Low-Rank · LLaMA3-8B W4A40-shot avg accuracy: 61.1 -> 64.9Improved
replace fixed Hadamard with learnable global and layer-specific transforms (ablation)+Hadamard · LLaMA3-8B W4A40-shot avg accuracy: 64.9 -> 70.9Improved
replace fixed Hadamard with learnable global and layer-specific transforms (ablation)+Hadamard · Qwen3-8B W4A40-shot avg accuracy: 66.0 -> 70.2Improved
replace fixed Hadamard with learnable transforms (ablation)+Hadamard · Qwen3-8B W4A4WikiText2 perplexity: 16.3 -> 13.2Improved
increase low-rank rank from 32 to 128 (ablation)TwinQuant · LLaMA3-8B W4A40-shot avg accuracy: 68.0 -> 70.9Improved
increase low-rank rank from 32 to 128 (ablation)TwinQuant · LLaMA3-8B W4A4WikiText2 perplexity: 11.3 -> 9.4Improved
increase rank from 128 to 256 (ablation)TwinQuant · LLaMA3-8B W4A40-shot avg accuracy: 70.9 -> 71.1No clear effect
increase rank from 128 to 256 (ablation)TwinQuant · LLaMA3-8B W4A4memory overhead: 4.9% -> 9.9%Hurt
use learnable decomposition instead of SVDQuant rank-32 decompositionTwinQuant · Qwen3-32B W4A40-shot avg accuracy: 72.2 -> 73.3Improved

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