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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add top-singular-subspace preservation (GRASP) to 4-bit Muon (ablation) | 4bit-Muon-base · SlimPajama validation (350M) | perplexity: 15.96 -> 15.67 | Improved |
| add top-singular-subspace preservation (GRASP) to 4-bit Muon (ablation) | 4bit-Muon-base · SlimPajama validation (130M) | perplexity: 19.73 -> 19.35 | Improved |
| add top-singular-subspace preservation (GRASP) to 4-bit Muon (ablation) | 4bit-Muon-base · SlimPajama validation (1.1B) | perplexity: 12.76 -> 12.48 | Improved |
| add GRASP to 4-bit Muon (ablation) | 4bit-Muon-base · Downstream average (350M) | accuracy: 43.7 -> 44.5 | Improved |
| use 4-bit GRASP instead of fp32 Muon | Muon · Total memory (1.1B) | memory (GB): 13.22 -> 10.14 | Improved |
| use 4-bit GRASP instead of fp32 Muon | Muon · Step time (1.1B) | step time (s): 61.3 -> 61.9 | No clear effect |
| use 4-bit GRASP instead of fp32 Muon in SFT | Qwen2.5-7B-Math · Minerva Math | Pass@1: 26.8 -> 29.0 | Improved |
| use 4-bit GRASP instead of fp32 Muon in SFT | Qwen2.5-7B · MBPP | Pass@1: 71.4 -> 70.9 | No clear effect |
| preserve only top singular space, discard residual (ablation) | GRASP · Pretraining loss | training 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 half | Improved |
| increase top singular space rank toward 1/2 (ablation) | GRASP · LLaMA-350M pretraining | training loss: rank 1/2 shows no difference from baseline | Improved |
| run Newton-Schulz iterations on quantized top-subspace momentum (k=64) | Muon · 1.1B LLaMA first 100 iterations | relative error RE(Mtop): 0.08 -> 3.31 | Hurt |
Closest papers by meaning
- Sharp Capacity Scaling of Spectral Optimizers in Learning Associative Memory (NeurIPS 2026 · similarity 0.89)
- Towards Understanding the Power and Limits of the Muon Optimizer: A River-Valley Perspective (NeurIPS 2026 · similarity 0.89)
- LittleBit: Ultra Low-Bit Quantization via Latent Factorization (NeurIPS 2025 · similarity 0.88)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.88)
- Powering Up Zeroth-Order Training via Subspace Gradient Orthogonalization (NeurIPS 2026 · similarity 0.87)
- Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression (NeurIPS 2025 · similarity 0.87)
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.
