Research map / Large language models / Language, safety and interpretability
CGSVD: Cascaded Granular Singular Value Decomposition for Large Language Model Compression
Yuli Chen, Shuhao Zhang, Jiale Han, Fanshen Meng, Haishen Jiang, Bo Cheng, Qiang Tong, Xiulei Liu
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 |
|---|---|---|---|
| use CGSVD cascaded non-uniform allocation instead of uniform SVD-LLM | SVD-LLM · LLaMA3.1-8B, 30% compression | WikiText2 perplexity: 32.92 -> 17.46 | Improved |
| use CGSVD instead of SVD-LLM at 60% compression | SVD-LLM · LLaMA2-7B WikiText2 | perplexity: 89.95 -> 49.92 | Improved |
| add layer importance via angular distance (inter-layer allocation) (ablation) | SVD-LLM uniform allocation · LLaMA2-7B, WikiText2, 60% | perplexity: 89.95 -> 65.85 | Improved |
| add module compressibility via spectral entropy (intra-layer allocation) (ablation) | SVD-LLM uniform allocation · LLaMA2-7B, WikiText2, 60% | perplexity: 89.95 -> 73.83 | Improved |
| combine layer importance and module compressibility (ablation) | SVD-LLM uniform allocation · LLaMA2-7B, WikiText2, 60% | perplexity: 89.95 -> 50.48 | Improved |
| add Iterative Residual Filling (ablation) | CGSVD (LI + MC) · LLaMA2-7B, WikiText2, 60% | perplexity: 50.48 -> 49.92 | Improved |
| add Iterative Residual Filling (ablation) | CGSVD (LI + MC) · LLaMA2-7B, WikiText2, 30% | perplexity: 9.56 -> 9.37 | Improved |
| add LoRA fine-tuning to CGSVD | CGSVD · LLaMA2-7B, WikiText2, 60% | perplexity: 49.92 -> 16.72 | Improved |
| increase calibration sequence length (ablation) | CGSVD · WikiText2 | perplexity: 20.28 -> 8.65 | Improved |
| increase calibration batch size (ablation) | CGSVD · WikiText2 | perplexity: 11.31 -> 9.29 | Improved |
| add 8-bit GPTQ quantization on top of CGSVD | CGSVD · LLaMA2-7B WikiText2, 30% | perplexity: 9.37 -> 9.37 | No clear effect |
| add 4-bit GPTQ quantization on top of CGSVD | CGSVD · LLaMA2-7B WikiText2, 30% | perplexity: 9.37 -> 14.31 | Hurt |
Closest papers by meaning
- SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025 · similarity 0.92)
- LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression (NeurIPS 2026 · similarity 0.92)
- Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025 · similarity 0.91)
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.91)
- Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.90)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.90)
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.
