AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Achieving low-bit Muon through subspace preservation and grid quantization

Huaijin Wu, Bingrui Li, Yebin Yang, Yi Tu, Zhanpeng Zhou, Jianfei Chen, Junchi Yan

ICLR 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
add top-singular-subspace preservation (GRASP) to 4-bit Muon (ablation)4bit-Muon-base · SlimPajama validation (350M)perplexity: 15.96 -> 15.67Improved
add top-singular-subspace preservation (GRASP) to 4-bit Muon (ablation)4bit-Muon-base · SlimPajama validation (130M)perplexity: 19.73 -> 19.35Improved
add top-singular-subspace preservation (GRASP) to 4-bit Muon (ablation)4bit-Muon-base · SlimPajama validation (1.1B)perplexity: 12.76 -> 12.48Improved
add GRASP to 4-bit Muon (ablation)4bit-Muon-base · Downstream average (350M)accuracy: 43.7 -> 44.5Improved
use 4-bit GRASP instead of fp32 MuonMuon · Total memory (1.1B)memory (GB): 13.22 -> 10.14Improved
use 4-bit GRASP instead of fp32 MuonMuon · Step time (1.1B)step time (s): 61.3 -> 61.9No clear effect
use 4-bit GRASP instead of fp32 Muon in SFTQwen2.5-7B-Math · Minerva MathPass@1: 26.8 -> 29.0Improved
use 4-bit GRASP instead of fp32 Muon in SFTQwen2.5-7B · MBPPPass@1: 71.4 -> 70.9No clear effect
preserve only top singular space, discard residual (ablation)GRASP · Pretraining losstraining loss: accuracy loss exceeds 2%Hurt
use grid quantization instead of group quantization (ablation)4-bit Muon · Pretraining loss (350M)accuracy loss: reduces the accuracy loss of group quantization by halfImproved
increase top singular space rank toward 1/2 (ablation)GRASP · LLaMA-350M pretrainingtraining loss: rank 1/2 shows no difference from baselineImproved
run Newton-Schulz iterations on quantized top-subspace momentum (k=64)Muon · 1.1B LLaMA first 100 iterationsrelative error RE(Mtop): 0.08 -> 3.31Hurt

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