Research map / Architectures and efficiency / Quantization
Dissecting Quantization Error: A Concentration-Alignment Perspective
Marco Federici, Boris van Breugel, Paul N. Whatmough, Markus Nagel
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 alignment-optimizing block transform before Hadamard (CAT block) vs Hadamard-only QuaRot | W4A4 RTN · WikiText-2, Llama 3 8B | perplexity: 10.99 -> 8.29 | Improved |
| train CAT (block) transform on calibration data | CAT (block) W4A4 RTN · WikiText-2, Llama 3 8B | perplexity: 8.29 -> 7.33 | Improved |
| train CAT (block) transform | CAT (block) W4A4 RTN · 0-shot avg, Llama 3 8B | accuracy (%): 67.77 -> 69.66 | Improved |
| CAT (block) trained instead of FlatQuant | W4A4 RTN · WikiText-2 / 0-shot, Llama 3 8B | perplexity / accuracy: 7.61 -> 7.33 ppl; 68.34 -> 69.66 acc | Mixed |
| replace RTN with GPTQ weight quantization (ablation) | CAT (block) trained and FlatQuant · WikiText-2, Llama 3 8B | perplexity: 7.33 -> 7.25 (CAT) | No clear effect |
| replace RTN with GPTQ weight quantization (ablation) | QuaRot · WikiText-2, Llama 3 8B | perplexity: 10.99 -> 8.21 | Improved |
| increase CAT block size (ablation) | CAT (block) · Llama 3.2 1B, WikiText-2 | alignment, SQNR, perplexity | Improved |
| apply Hadamard or other rotations (ablation) | rotation-based transforms · alignment on LLM layers | alignment [dB] | No clear effect |
| apply channel scaling (SmoothQuant) to move outliers into weights (ablation) | W4A4 · LLM layers | activation/weight concentration | Mixed |
| use asymmetric instead of symmetric activation quantization (ablation) | uniform integer quantization · LLM layers | concentration | No clear effect |
| change calibration set size or dataset for CAT (ablation) | CAT (block) · calibration sensitivity | accuracy | No clear effect |
| use SmoothQuant at W4A4 without rotation | SmoothQuant RTN · WikiText-2, Llama 3 8B | perplexity: 321.22 -> 233.61 vs no transform; still far above FP 6.14 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · better on 8 of 8 shared results
- SmoothQuant — SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models (ICML 2023) · better on 1 of 1 shared result
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
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.88)
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.88)
- QERA: an Analytical Framework for Quantization Error Reconstruction (ICLR 2025 · similarity 0.88)
- ASER: Activation Smoothing and Error Reconstruction for Large Language Model Quantization (AAAI 2025 · similarity 0.88)
- OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting (ICLR 2025 · similarity 0.88)
- Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization (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.
