Research map / Architectures and efficiency / Quantization
TwinQuant: Learnable Subspace Decomposition for 4-Bit LLM Quantization
Haodong WANG, Junjie Liu, Zicong Hong, Qianli Liu, Jian Lin, Song Guo, Xu Chen
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 SVD low-rank decomposition to naive W4A4 (ablation) | Naive 4-bit · LLaMA3-8B W4A4 | 0-shot avg accuracy: 41.7 -> 61.1 | Improved |
| add SVD low-rank decomposition to naive W4A4 (ablation) | Naive 4-bit · LLaMA3-8B W4A4 | WikiText2 perplexity: 91.6 -> 19.6 | Improved |
| add SVD low-rank decomposition to naive W4A4 (ablation) | Naive 4-bit · Qwen3-8B W4A4 | 0-shot avg accuracy: 41.5 -> 62.8 | Improved |
| replace SVD low-rank with fixed Hadamard transform in decomposed branches (ablation) | +Low-Rank · LLaMA3-8B W4A4 | 0-shot avg accuracy: 61.1 -> 64.9 | Improved |
| replace fixed Hadamard with learnable global and layer-specific transforms (ablation) | +Hadamard · LLaMA3-8B W4A4 | 0-shot avg accuracy: 64.9 -> 70.9 | Improved |
| replace fixed Hadamard with learnable global and layer-specific transforms (ablation) | +Hadamard · Qwen3-8B W4A4 | 0-shot avg accuracy: 66.0 -> 70.2 | Improved |
| replace fixed Hadamard with learnable transforms (ablation) | +Hadamard · Qwen3-8B W4A4 | WikiText2 perplexity: 16.3 -> 13.2 | Improved |
| increase low-rank rank from 32 to 128 (ablation) | TwinQuant · LLaMA3-8B W4A4 | 0-shot avg accuracy: 68.0 -> 70.9 | Improved |
| increase low-rank rank from 32 to 128 (ablation) | TwinQuant · LLaMA3-8B W4A4 | WikiText2 perplexity: 11.3 -> 9.4 | Improved |
| increase rank from 128 to 256 (ablation) | TwinQuant · LLaMA3-8B W4A4 | 0-shot avg accuracy: 70.9 -> 71.1 | No clear effect |
| increase rank from 128 to 256 (ablation) | TwinQuant · LLaMA3-8B W4A4 | memory overhead: 4.9% -> 9.9% | Hurt |
| use learnable decomposition instead of SVDQuant rank-32 decomposition | TwinQuant · Qwen3-32B W4A4 | 0-shot avg accuracy: 72.2 -> 73.3 | 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.
- FlatQuant — FlatQuant: Flatness Matters for LLM Quantization (ICML 2025) · better on 4 of 4 shared results
- SpinQuant — SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025) · better on 4 of 4 shared results
- QuaRot — QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs (NeurIPS 2024) · better on 2 of 2 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
- ResQ: Mixed-Precision Quantization of Large Language Models with Low-Rank Residuals (ICML 2025 · similarity 0.90)
- any4: Learned 4-bit Numeric Representation for LLMs (ICML 2025 · similarity 0.90)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.90)
- LittleBit: Ultra Low-Bit Quantization via Latent Factorization (NeurIPS 2025 · similarity 0.89)
- DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs (NeurIPS 2024 · similarity 0.89)
- OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting (ICLR 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.
