Research map / Architectures and efficiency / Quantization
AdaSVD: Singular Value Decomposition with Adaptive Mechanisms for Large Multimodal Models
Zhiteng Li, Mingyuan Xia, Jingyuan Zhang, Zheng Hui, Haotong Qin, Linghe Kong, Yulun Zhang, Xiaokang Yang
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 adaComp (adaptive compensation) (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 60% compression | perplexity: 78.82 -> 50.33 | Improved |
| add adaComp (adaptive compensation) (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 50% compression | perplexity: 30.00 -> 25.58 | Improved |
| add adaCR (adaptive compression ratio vs constant) (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 60% compression | perplexity: 69.46 -> 50.33 | Improved |
| add adaCR (adaptive compression ratio vs constant) (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 50% compression | perplexity: 27.33 -> 25.58 | Improved |
| increase adaComp iterations from 1 to 15 (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 60% compression | perplexity: 50.33 -> 62.34 | Hurt |
| increase adaComp iterations from 1 to 15 (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 40% compression | perplexity: 14.76 -> 15.84 | Hurt |
| raise minimum retention ratio from 0.30 to 0.40 (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 60% compression | perplexity: 50.33 -> 60.08 | Hurt |
| raise minimum retention ratio from 0.40 to 0.50 (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 40% compression | perplexity: 15.01 -> 14.76 | No clear effect |
| use Moore-Penrose pseudoinverse update and stack-of-batch calibration (ablation) | AdaSVD on LLaMA2-7B · compression error | update error | Improved |
| combine with GPTQ-INT4 quantization (ablation) | AdaSVD on LLaMA2-7B · WikiText-2, 40% compression | perplexity: 14.76 -> 22.55 | Hurt |
| use AdaSVD instead of SVD-LLM, combined with GPTQ-INT4 | AdaSVD on LLaMA2-7B · WikiText-2, 40% compression | perplexity: 33.56 -> 22.55 | Improved |
| use AdaSVD instead of SVD-LLM | AdaSVD on LLaMA2-7B · WikiText-2, 60% compression | perplexity: 89.90 -> 50.33 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 10 of 10 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
- Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives (ICLR 2025 · similarity 0.88)
- CASP: Compression of Large Multimodal Models Based on Attention Sparsity (CVPR 2025 · similarity 0.86)
- Efficiently Serving Large Multimodal Models Using EPD Disaggregation (ICML 2025 · similarity 0.86)
- SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025 · similarity 0.85)
- WSVD: Weighted Low-Rank Approximation for Fast and Efficient Execution of Low-Precision Vision-Language Models (ICLR 2026 · similarity 0.85)
- Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025 · similarity 0.84)
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.
