AI Paper Atlas

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

CVPR 2026 · Architectures and efficiency › Quantization · approach: Post-training quantization of ViTs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add S2D regularization to AdamW fine-tuning, then quantize with ERQ W4A4SigLIP2-Base-384 · ImageNet-1k PTQtop-1 accuracy: 65.58 -> 72.99Improved
add S2D regularization, then quantize with RepQ-ViT W5A5SigLIP2-Base-384 · ImageNet-1k PTQtop-1 accuracy: 46.04 -> 78.07Improved
add S2D regularization, then quantize with PTQ4ViT W4A4SigLIP2-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 W4A4SigLIP2-Base · ImageNet-1k PTQ (512 res)top-1 accuracy: 58.6 -> 68.1Improved
add S2D regularization to full-precision fine-tuningSigLIP2-Base · ImageNet-1k FP16top-1 accuracy: 85.0 -> 85.0No clear effect
add S2D regularization to QAT W4A4SigLIP2-Base-384 · ImageNet-1k QATtop-1 accuracy: 65.76 -> 69.70Improved
add S2D regularization to QAT W3A4SigLIP2-Base-384 · ImageNet-1k QATtop-1 accuracy: 59.85 -> 62.37Improved
add S2D regularization for detection fine-tuning, ERQ W5A5SigLIP2-Base-384 Detectron2 · MS-COCO detectionAP: 10.8 -> 40.7Improved
add S2D regularization, ERQ W4A4SigLIP2-Base-384 Detectron2 · MS-COCO instance segmentationAP: 0.50 -> 13.0Improved
add S2D to LLaVA-1.5 fine-tuningSigLIP2-Base-384 + Qwen2.5-0.5B · GQA FP16accuracy: 47.6 -> 55.4Improved
add S2D to LLaVA-1.5 fine-tuning, quantizedSigLIP2-Base-384 + Qwen2.5-0.5B · POPEaccuracy: 67.4 -> 66.6 (W5A5)No clear effect
add S2D to LLaVA-1.5 fine-tuning, W6A6SigLIP2-Base-384 + Qwen2.5-0.5B · TextVQAaccuracy: 19.2 -> 28.0Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy