Research map / Architectures and efficiency / Quantization
UniSVQ: 2-bit Unified Scalar-Vector Quantization
Haoyu Wang, Haiyan Zhao, Xingyu Yu, Zhangyang Yao, Xu Han, Zhiyuan Liu, Maosong Sun
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 block-wise fine-tuning of the affine grid (ablation) | UniSVQ · 0-shot QA avg, Qwen-3-8B | accuracy: 66.99 -> 67.95 | Improved |
| initialize with random orthogonal matrix instead of D4 lattice generator (ablation) | UniSVQ · 0-shot QA avg, Qwen-3-8B | accuracy: 61.00 -> 67.95 | Improved |
| apply Randomized Hadamard Transform before quantization (ablation) | UniSVQ · WikiText-2, Llama-3-8B | perplexity: 8314.12 -> 20.04 | Improved |
| increase vector dimension d from 4 to 8 (no fine-tuning) (ablation) | UniSVQ · 0-shot QA avg, Qwen-3-8B | accuracy: 66.99 -> 67.34 | No clear effect |
| use linear-constrained codebook on top of Hessian compensation and RHT (ablation) | UniSVQ · weight-level SNR, Qwen-3-4B 2-bit | SNR: -3.06 -> 8.48 | Improved |
| increase bit-width from 2 to 3 (ablation) | UniSVQ · 0-shot QA avg, Llama-3-8B | accuracy: 63.42 -> 67.71 | Improved |
| replace scalar quantization with UniSVQ grid | UniSVQ vs OSTQuant · WikiText-2, Qwen-3-8B 2-bit | perplexity: 26.08 -> 14.82 | Improved |
| use UniSVQ instead of AQLM | UniSVQ vs AQLM · WikiText-2, Qwen-3-8B 2-bit | perplexity: 18.26 -> 14.82 | Improved |
| UniSVQ vs QuIP# (E8P lattice VQ) | UniSVQ · 0-shot QA avg, Qwen-3-8B | accuracy: 67.55 -> 67.95 | Mixed |
| UniSVQ vs Qtip trellis-coded VQ | UniSVQ · 0-shot QA avg, Qwen-3-8B | accuracy: 71.08 vs 67.95 | Hurt |
| use UniSVQ kernels instead of FP16 | UniSVQ · Llama-3-8B throughput | tok/s: 60.38 -> 101.65 | Improved |
| UniSVQ instead of Quip#/AQLM for decoding speed | UniSVQ · Llama-3-8B throughput | tok/s: 79.60 -> 101.65 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- SpinQuant — SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025) · better on 7 of 7 shared results
- OSTQuant — OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting (ICLR 2025) · better on 7 of 7 shared results
- Quip# — QuIP: 2-Bit Quantization of Large Language Models With Guarantees (NeurIPS 2023) · better on 5 of 7 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
- RSAVQ: Riemannian Sensitivity-Aware Vector Quantization for Large Language Models (NeurIPS 2025 · similarity 0.90)
- Learning Grouped Lattice Vector Quantizers for Low-Bit LLM Compression (NeurIPS 2025 · similarity 0.89)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.89)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.89)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.89)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · 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.
