AI Paper Atlas

Research map / Architectures and efficiency / Quantization

WUSH: Near-Optimal Adaptive Transforms for LLM Quantization

Jiale Chen, Vage Egiazarian, Roberto L. Castro, Torsten Hoefler, Dan Alistarh

ICML 2026 · Architectures and efficiency › Quantization · approach: Ultra-low-bit LLM quantization · 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 WUSH transform with GPTQ instead of Hadamard GPTQ (MR-GPTQ)W4A4 quantization · Llama-3.1-8B-Instruct MXFP4average accuracy: 73.65 -> 74.35Improved
add WUSH transform with GPTQ instead of Hadamard GPTQ (MR-GPTQ)W4A4 quantization · Llama-3.1-8B-Instruct NVFP4average accuracy: 75.84 -> 76.10Improved
add WUSH transform to RTN instead of HadamardRTN · Llama-3.1-8B-Instruct MXFP4average accuracy: 70.45 -> 73.21Improved
add WUSH transform to RTN instead of HadamardRTN · Llama-3.1-8B-Instruct NVFP4average accuracy: 74.05 -> 75.28Improved
add Hadamard transform to RTNRTN · Llama-3.1-8B-Instruct NVFP4average accuracy: 74.73 -> 74.05Hurt
add Hadamard transform to RTNRTN · Llama-3.1-8B-Instruct MXFP4average accuracy: 69.32 -> 70.45Improved
add Hadamard backbone to the data-aware component (WUS to WUSH) (ablation)WUSH · Qwen3-8B INT4 layerwise loss, QL2 loss x1e-3: 213 -> 2.39Improved
add Hadamard backbone to the data-aware component (WUS to WUSH) (ablation)WUSH · Qwen3-8B MXFP4 layerwise loss, QL2 loss x1e-3: 6.27 -> 3.34Improved
add Hadamard backbone to the data-aware component (WUS to WUSH) (ablation)WUSH · Qwen3-8B NVFP4 layerwise loss, QL2 loss x1e-3: 2.26 -> 2.40No clear effect
add data-aware component on top of Hadamard (ablation)Hadamard · Qwen3-8B MXFP4 layerwise loss, QL2 loss x1e-3: 7.24 -> 3.34Improved
use GPTQ instead of RTN with WUSHWUSH · Llama-3.1-8B-Instruct MXFP4average accuracy: 73.21 -> 74.35Improved
use WUSH+Quant kernel instead of Hadamard+Quant kernelFP4 inference · kernel throughputthroughput: average difference of only 1.3%No clear effect

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