Research map / Architectures and efficiency / Quantization
S2D: Selective Spectral Decay for Quantization-Friendly Conditioning of Neural Activations
Arnav Chavan, Nahush Lele, Udbhav Bamba, Sankalp Dayal, Aditi Raghunathan, Deepak Gupta
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 S2D regularization to AdamW fine-tuning, then quantize with ERQ W4A4 | SigLIP2-Base-384 · ImageNet-1k PTQ | top-1 accuracy: 65.58 -> 72.99 | Improved |
| add S2D regularization, then quantize with RepQ-ViT W5A5 | SigLIP2-Base-384 · ImageNet-1k PTQ | top-1 accuracy: 46.04 -> 78.07 | Improved |
| add S2D regularization, then quantize with PTQ4ViT W4A4 | SigLIP2-Base · ImageNet-1k PTQ (384 and 512 res) | top-1 accuracy: 3.4 -> 3.9 (384); 4.2 -> 3.8 (512) | No clear effect |
| add S2D regularization at 512 resolution, ERQ W4A4 | SigLIP2-Base · ImageNet-1k PTQ (512 res) | top-1 accuracy: 58.6 -> 68.1 | Improved |
| add S2D regularization to full-precision fine-tuning | SigLIP2-Base · ImageNet-1k FP16 | top-1 accuracy: 85.0 -> 85.0 | No clear effect |
| add S2D regularization to QAT W4A4 | SigLIP2-Base-384 · ImageNet-1k QAT | top-1 accuracy: 65.76 -> 69.70 | Improved |
| add S2D regularization to QAT W3A4 | SigLIP2-Base-384 · ImageNet-1k QAT | top-1 accuracy: 59.85 -> 62.37 | Improved |
| add S2D regularization for detection fine-tuning, ERQ W5A5 | SigLIP2-Base-384 Detectron2 · MS-COCO detection | AP: 10.8 -> 40.7 | Improved |
| add S2D regularization, ERQ W4A4 | SigLIP2-Base-384 Detectron2 · MS-COCO instance segmentation | AP: 0.50 -> 13.0 | Improved |
| add S2D to LLaVA-1.5 fine-tuning | SigLIP2-Base-384 + Qwen2.5-0.5B · GQA FP16 | accuracy: 47.6 -> 55.4 | Improved |
| add S2D to LLaVA-1.5 fine-tuning, quantized | SigLIP2-Base-384 + Qwen2.5-0.5B · POPE | accuracy: 67.4 -> 66.6 (W5A5) | No clear effect |
| add S2D to LLaVA-1.5 fine-tuning, W6A6 | SigLIP2-Base-384 + Qwen2.5-0.5B · TextVQA | accuracy: 19.2 -> 28.0 | Improved |
Closest papers by meaning
- SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.87)
- DiRotQ: Rotation-Aware Quantization for 4-bit Diffusion Transformers (NeurIPS 2026 · similarity 0.86)
- The Quantization Benefits of Residual-Free Transformers (NeurIPS 2026 · similarity 0.86)
- PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · similarity 0.86)
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.86)
- Understanding and Minimising Outlier Features in Transformer Training (NeurIPS 2024 · similarity 0.85)
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.
