Research map / Architectures and efficiency / Quantization
Beyond Outliers: A Study of Optimizers Under Quantization
Georgios Vlassis, Saleh Ashkboos, Alexandra Volkova, Torsten Hoefler, Dan Alistarh
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 |
|---|---|---|---|
| use Muon instead of AdamW in full precision | OLMo2 1.5B · avg zero-shot accuracy | accuracy: 67.93 -> 69.19 | Improved |
| use Muon in PTQ W4A4 | OLMo2 1.5B · avg zero-shot accuracy after PTQ | accuracy: 62.51 (AdamW) -> 47.75 (Muon) | Hurt |
| use Shampoo instead of AdamW for PTQ | OLMo2 1.5B · avg zero-shot accuracy after PTQ | accuracy: 62.51 -> 63.88 | Improved |
| use outlier metrics MMR/kurtosis to predict PTQ accuracy | 760M models · correlation with accuracy | correlation | No clear effect |
| increase learning rate | all optimizers · max-to-median ratio | MMR | Hurt |
| use Shampoo for 4-bit QAT | OLMo2 1.5B · accuracy degradation vs FP | accuracy drop: -1.20 vs -1.63 AdamW, -2.11 Muon | Improved |
| FP ranking as predictor of QAT ranking | six optimizers · avg zero-shot accuracy | accuracy | Mixed |
| use Shampoo, parameter efficiency under 4-bit QAT | scaling law · rho_4bit | parameter efficiency | Improved |
| use Shampoo vs Muon for QAT at 350M | OLMo2 350M · accuracy degradation | accuracy drop: SOAP best 56.61; Shampoo -2.64 vs Muon -4.98 | Mixed |
Closest papers by meaning
- Predicting Quantization Price for Selecting PTQ Configurations Before Deployment (NeurIPS 2026 · similarity 0.87)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.87)
- Scaling Laws for Precision (ICLR 2025 · similarity 0.87)
- LOTION: Smoothing the Optimization Landscape for Quantized Training (NeurIPS 2026 · similarity 0.86)
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.86)
- Coverage-Based Calibration for Post-Training Quantization via Weighted Maximum Coverage over Outlier Channels (NeurIPS 2026 · similarity 0.86)
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.
