- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes a quantization method called High-Degree Node Energy Smearing, which suppresses the numerical magnitudes of hub nodes via rotation transformations, thereby reducing the variance in activations for quantization and improving quantization accuracy. Experiments show that the method achieves good results across various tasks.
Critical strengths
- The paper is well written and easy to follow.
- The motivation is clear and well grounded: the existence of hub nodes and tail nodes indeed makes GNN quantization difficult.
Critical weaknesses
- The novelty is limited. First, the outlier phenomenon in GNNs has been extensively discussed in prior work. Second, rotation-matrix-based solutions have been widely adopted for outlier suppression in LLM quantization. What is the fundamental difference between the rotation method proposed in this paper and existing rotation-based methods?
- The experiments are insufficient. The paper does not conduct direct comparisons with current state-of-the-art methods, such as DQ and , which makes it hard to demonstrate the effectiveness of the proposed method.
Questions for authors
- What is the fundamental difference between the rotation method proposed in this paper and rotation methods applied in LLM quantization (e.g., SpinQuant: LLM quantization with learned rotations)?
- Please supplement the experiments with direct comparisons against previous state-of-the-art methods, such as DQ and .