Research map / Architectures and efficiency / Quantization
Outlier Suppression: Pushing the Limit of Low-bit Transformer Language Models
Xiuying Wei, Yunchen Zhang, Xiangguo Zhang, Ruihao Gong, Shanghang Zhang, Qi Zhang, Fengwei Yu, Xianglong Liu
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 Gamma Migration to MinMax calibration (ablation) | RoBERTa 6-bit PTQ · QNLI | accuracy: 62.13 -> 78.56 | Improved |
| add Gamma Migration to MinMax calibration (ablation) | RoBERTa 6-bit PTQ · CoLA | Matthews corr.: 0.0 -> 0.0 | No clear effect |
| add Token-Wise Clipping (coarse stage) to baseline (ablation) | RoBERTa 6-bit PTQ · QNLI | accuracy: 62.13 -> 79.46 | Improved |
| add fine-grained stage after coarse token-wise clipping (ablation) | Token-Wise Clipping · SST-2 | accuracy: 91.63 -> 91.63 | No clear effect |
| add fine-grained stage after coarse token-wise clipping (ablation) | Token-Wise Clipping · CoLA | Matthews corr.: 34.95 -> 37.64 | Improved |
| add Gamma Migration to Token-Wise Clipping (ablation) | RoBERTa 6-bit PTQ · CoLA | Matthews corr.: 37.64 -> 46.35 | Improved |
| add Gamma Migration to Token-Wise Clipping (ablation) | RoBERTa 6-bit PTQ · QNLI | accuracy: 79.66 -> 86.82 | Improved |
| use outlier suppression framework vs best baseline calibration | BERT 6-6-6 PTQ · GLUE average | score: 74.49 -> 81.19 | Improved |
| use outlier suppression framework vs best baseline calibration | BART 6-6-6 PTQ · GLUE average | score: 67.31 -> 79.10 | Improved |
| use outlier suppression framework with 4-4-4 QAT vs LSQ+ | BERT QAT · GLUE average | score: 71.49 -> 81.13 | Improved |
| use framework with KD at 2-2-4 vs LSQ+(+KD) | BERT QAT · GLUE average | score: 66.63 -> 80.56 | Improved |
| use framework at 8-bit on SQuAD v1.1 BERT vs OMSE | BERT 8-8-8 PTQ · SQuAD v1.1 | F1/EM: 87.90/80.16 -> 87.60/79.80 | Hurt |
Closest papers by meaning
- Rethinking Channel Dimensions to Isolate Outliers for Low-bit Weight Quantization of Large Language Models (ICLR 2024 · similarity 0.91)
- DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024 · similarity 0.91)
- ASER: Activation Smoothing and Error Reconstruction for Large Language Model Quantization (AAAI 2025 · similarity 0.89)
- Systematic Outliers in Large Language Models (ICLR 2025 · similarity 0.89)
- Outlier-Efficient Hopfield Layers for Large Transformer-Based Models (ICML 2024 · similarity 0.89)
- Structured Transforms for Low-Overhead Quantization of Language Models (NeurIPS 2026 · similarity 0.88)
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.
