Research map / Architectures and efficiency / Pruning and efficient tuning
Large Language Model Compression with Global Rank and Sparsity Optimization
Changhai Zhou, Qian Qiao, Yuhua Zhou, Yuxin Wu, Shichao Weng, WEIZHONG ZHANG, Cheng Jin
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 CAP (RPCA + learned global allocation) instead of Wanda | LLaMA-3.1-8B-Instruct 50% sparse · GSM8K | accuracy (%): 45.6 -> 56.8 | Improved |
| use CAP instead of OATS at 50% sparsity | LLaMA-3 8B · WikiText | perplexity: 10.87 -> 10.35 | Improved |
| use CAP instead of SparseGPT at 50% sparsity | Phi-3 Mini · zero-shot avg | accuracy (%): 66.36 -> 69.12 | Improved |
| use CAP instead of layerwise allocation (OWL) on Wanda | LLaMA-3 8B 50% · zero-shot avg | accuracy (%): 63.95 -> 66.85 | Improved |
| add sparse component to pure SVD compression | LLaMA-7B 20% · WikiText-2 | perplexity: 6.08 -> 5.85 | Improved |
| add fine-tuning to CAP | DeBERTa-V3-base · MNLI | accuracy (%): 78.2 -> 86.8 | Improved |
| training-free CAP vs LoSparse (with FT) | DeBERTa-V3-base · MNLI | accuracy (%): 83.8 -> 78.2 | Hurt |
| replace learned global allocation with uniform thresholding of RPCA output (ablation) | CAP · LLaMA2-7B | performance | Hurt |
| use CAP's high-sparsity sparse component for inference | LLaMA-3.1-8B · throughput | tok/s: 163.4 -> 176.5 | Improved |
| calibrate on WikiText-2 instead of C4 (ablation) | CAP · LLaMA-3.1-8B | perplexity: PPL 7.05 -> 6.88; acc 76.8 -> 75.4 | Mixed |
| calibrate on GitHub Code instead of C4 (ablation) | CAP · LLaMA-3.1-8B | accuracy (%): 76.8 -> 76.1 | No clear effect |
| use CAP on GLUE with BERT-base at 16% budget vs CoFi | BERT-base · MNLI | accuracy (%): 80.5 -> 82.8 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- OATS — OATS: Outlier-Aware Pruning Through Sparse and Low Rank Decomposition (ICLR 2025) · better on 4 of 4 shared results
- SVD-LLM v2 — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 1 of 1 shared result
- Dobi-SVD — Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives (ICLR 2025) · better on 1 of 1 shared result
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
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.93)
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 2025 · similarity 0.92)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.92)
- Compressing Large Language Models by Joint Sparsification and Quantization (ICML 2024 · similarity 0.91)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.90)
- Dynamic Low-Rank Sparse Adaptation for Large Language Models (ICLR 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.
