Research map / Vision-language and multimodal / Multimodal LLMs
Activation Quantization of Vision Encoders Needs Prefixing Registers
Seunghyeon Kim, Taesun Yeom, Jinho Kim, Wonpyo Park, Kyuyeun Kim, Jaeho Lee
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 RegCache on top of RepQ-ViT (CLIP-B/16) | RepQ-ViT · ImageNet-1k zero-shot W4A4 | accuracy: 1.83 -> 21.50 | Improved |
| add RegCache on top of FIMA-Q (CLIP-B/16) | FIMA-Q · ImageNet-1k zero-shot W4A4 | accuracy: 50.41 -> 62.08 | Improved |
| add RegCache on top of ERQ (SigLIP2-B/16) | ERQ · ImageNet-1k zero-shot W4A4 | accuracy: 10.56 -> 28.98 | Improved |
| add RegCache on naive W8A8 quantization (SigLIP2) | Naive RTN · ImageNet-1k zero-shot W8A8 | accuracy: 26.04 -> 72.35 | Improved |
| add RegCache to DINOv2 with advanced PTQ at W8A8 | PTQ4ViT/RepQ-ViT/ERQ · ImageNet-1k zero-shot W8A8 (DINOv2) | accuracy: average -0.47 | Hurt |
| add RegCache at high bitwidth with strong PTQ | PTQ baselines · ImageNet-1k zero-shot W8A8 | accuracy: average +0.25 to +1.36 | No clear effect |
| add prefix caching only (no token deletion) (ablation) | RegCache · ImageNet-1k W8A8 (SigLIP2) | accuracy: 26.04 -> 23.82 | Hurt |
| add token deleting only (no prefix caching) (ablation) | RegCache · ImageNet-1k W8A8 (SigLIP) | accuracy: 69.71 -> 42.41 | Hurt |
| combine prefix caching with token deleting (ablation) | RegCache · ImageNet-1k W8A8 (SigLIP2) | accuracy: 26.04 -> 72.35 | Improved |
| add RegCache to RepQ-ViT in 4-bit VLM vision encoder | Qwen3-VL-2B · VQAv2 | accuracy: 32.03 -> 42.72 | Improved |
| add RegCache to RepQ-ViT in 4-bit VLM vision encoder | Qwen3-VL-8B · MLVU | accuracy: 44.89 -> 45.40 | Improved |
| add RegCache with ImageNet-searched prefix to other datasets | CLIP-B/16 W4A4 naive · StanfordCars | accuracy: 29.76 -> 49.96 | Improved |
Closest papers by meaning
- Fine-Grained Post-Training Quantization for Large Vision Language Models with Quantization-Aware Integrated Gradients (CVPR 2026 · similarity 0.88)
- Q-VLM: Post-training Quantization for Large Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Quant Experts: Token-aware Adaptive Error Reconstruction with Mixture of Experts for Large Vision-Language Models Quantization (CVPR 2026 · similarity 0.87)
- SPHQuant: Efficient extreme low bit weight quantization for Vision-Language Models (NeurIPS 2026 · similarity 0.87)
- Amortized-Precision Quantization for Early-Exit Vision Transformers (NeurIPS 2026 · similarity 0.87)
- Vision Transformers Don't Need Trained Registers (NeurIPS 2025 · similarity 0.87)
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.
