Research map / Architectures and efficiency / Quantization
ParoQuant: Pairwise Rotation Quantization for Efficient Reasoning LLM Inference
Yesheng Liang, Haisheng Chen, Song Han, Zhijian 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 8 independent rotations to channel-wise scaling (ablation) | ParoQuant (w/ Stage 2) · C4 perplexity, LLaMA-3-8B | perplexity: 7.41 -> 7.27 | Improved |
| add channel-wise scaling to 8 independent rotations (ablation) | ParoQuant (w/ Stage 2) · C4 perplexity, LLaMA-3-8B | perplexity: 7.40 -> 7.27 | Improved |
| add Stage 2 fine-tuning (ablation) | ParoQuant (8 IR + S) · C4 perplexity, LLaMA-3-8B | perplexity: 7.35 -> 7.27 | Improved |
| add transforms (8 IR + S) to RTN without stage 2 (ablation) | RTN · C4 perplexity, LLaMA-3-8B | perplexity: 7.56 -> 7.35 | Improved |
| add Stage 2 when no transform is used (ablation) | RTN · C4 perplexity, LLaMA-3-8B | perplexity: 7.56 -> 7.42 | Improved |
| increase independent rotations from 0 to 8 (2048 samples) (ablation) | ParoQuant · C4 perplexity, LLaMA-3-8B | perplexity: 7.41 -> 7.27 | Improved |
| increase independent rotations from 4 to 8 (ablation) | ParoQuant · C4 perplexity, LLaMA-3-8B | perplexity: 7.27 -> 7.27 | No clear effect |
| increase calibration samples from 128 to 2048 (ablation) | ParoQuant · C4 perplexity, LLaMA-3-8B | perplexity: 7.30 -> 7.27 | Improved |
| use mixed calibration set instead of RedPajama only (ablation) | ParoQuant · MMLU-Pro, Qwen3-4B | accuracy: 69.5 -> 70.1 | Improved |
| use ParoQuant instead of AWQ | Reasoning LLMs (4-bit) · Reasoning average | accuracy: 59.5 -> 61.9 | Improved |
| use ParoQuant instead of EfficientQAT | Reasoning LLMs (4-bit) · Reasoning average | accuracy: 55.6 -> 61.9 | Improved |
| use ParoQuant instead of AWQ for decoding throughput | Qwen3-4B · Decoding throughput | tokens/s: 176 -> 160 | Hurt |
Later papers that compared with this method
2 later accepted papers report ParoQuant++ as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss (NeurIPS 2026 · better on 17 of 17 shared results)
- SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models (NeurIPS 2026 · better on 4 of 4 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
- SpinQuant: LLM Quantization with Learned Rotations (ICLR 2025 · similarity 0.93)
- LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss (NeurIPS 2026 · similarity 0.91)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.89)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.89)
- ConQuR: Corner Aligned Activation Quantization via Optimized Rotations for LLMs (NeurIPS 2026 · similarity 0.88)
- DartQuant: Efficient Rotational Distribution Calibration for LLM Quantization (NeurIPS 2025 · 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.
