Research map / Architectures and efficiency / Quantization
SERQ: Saliency-Aware Low-Rank Error Reconstruction for LLM Quantization
Yeonsik Park, Hyeonseong Kim, Seungkyu Choi
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 SERQ single saliency-aware low-rank matrix instead of L2QER two-factor SVD reconstruction | SERQ · WikiText2 W4A4 INT4, LLaMA-3 8B | perplexity: 11.44 -> 7.75 | Improved |
| use SERQ instead of L2QER under W4A8 | SERQ (RTN) · WikiText2 W4A8, LLaMA-3 8B | perplexity: 7.16 -> 6.71 | Improved |
| use SERQ instead of SpinQuant/QuaRot rotation methods | SERQ · MMLU W4A4, LLaMA-3 8B | accuracy: 49.93 -> 53.8 | Improved |
| use SERQ instead of SpinQuant on LLaMA-2 13B | SERQ · WikiText2 W4A4, LLaMA-2 13B | perplexity: 5.2 -> 5.2 | No clear effect |
| use SERQ instead of SpinQuant on LLaMA-3.2 1B MMLU | SERQ · MMLU W4A4, LLaMA-3.2 1B | accuracy: 26.38 -> 26.34 | No clear effect |
| use SERQ instead of L2QER on generation | SERQ · GSM8K 5-shot W4A4, LLaMA-3 8B | accuracy: 7.96 -> 23.65 | Improved |
| use MXFP4 with L2QER instead of INT4 L2QER | L2QER · MMLU W4A4, LLaMA-3 8B | accuracy: 38.33 -> 53.82 | Improved |
| increase low-rank size from 0 to 256 salient rows (ablation) | SERQ · WikiText2 W4A4, LLaMA-3 8B | perplexity: 9.8 -> 7.98 | Improved |
| add static activation flattening to SERQ (ablation) | SERQ · WikiText2 W4A4 INT RTN, Qwen-2.5 3B | perplexity: 10.83 -> 9.57 | Improved |
| add static activation flattening to SERQ (ablation) | SERQ · WikiText2 W4A4 INT RTN, LLaMA-2 7B | perplexity: 6.05 -> 6.03 | No clear effect |
| add error reconstruction to static flattening (ablation) | SERQ · WikiText2 W4A4 INT RTN, LLaMA-3.2 1B | perplexity: 56.92 -> 13.49 | Improved |
| change calibration data size or dataset (Wiki vs Pile, 32-512 samples) (ablation) | SERQ · WikiText2 W4A4, LLaMA-3 8B | perplexity: 8.07 -> 7.98 | No clear effect |
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.
- SpinQuant — SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025) · better on 3 of 3 shared results
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · better on 3 of 3 shared results
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
- RILQ: Rank-Insensitive LoRA-Based Quantization Error Compensation for Boosting 2-Bit Large Language Model Accuracy (AAAI 2025 · similarity 0.91)
- Exploring Post-training Quantization in LLMs from Comprehensive Study to Low Rank Compensation (AAAI 2024 · similarity 0.91)
- LQER: Low-Rank Quantization Error Reconstruction for LLMs (ICML 2024 · similarity 0.90)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.90)
- SliM-LLM: Salience-Driven Mixed-Precision Quantization for Large Language Models (ICML 2025 · similarity 0.90)
- Coverage-Based Calibration for Post-Training Quantization via Weighted Maximum Coverage over Outlier Channels (NeurIPS 2026 · similarity 0.89)
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.
