Research map / Architectures and efficiency / Quantization
Adaptive Nonlinear Compression for Large Foundation Models
Liang Xu, Shufan Shen, Qingming Huang, Yao Zhu, Xiangyang Ji, Shuhui Wang
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 adaptive budget allocation (ablation) | NLA on Swin-Tiny · ImageNet-1K | top-1 accuracy (%): 77.40 -> 77.63 | Improved |
| add adaptive budget allocation (ablation) | NLA on Swin-Base · ImageNet-1K | top-1 accuracy (%): 82.55 -> 82.73 | Improved |
| add adaptive budget allocation (ablation) | NLA on OPT-125M · WikiText-2 | perplexity: 31.52 -> 31.34 | Improved |
| add adaptive budget allocation (ablation) | NLA on OPT-1.3B · WikiText-2 | perplexity: 20.13 -> 19.77 | Improved |
| use cubic instead of linear budget schedule (ablation) | NLA on OPT-1.3B · WikiText-2 | perplexity: 43.24 -> 19.77 | Improved |
| use cubic instead of linear budget schedule (ablation) | NLA on OPT-125M · WikiText-2 | perplexity: 63.04 -> 31.34 | Improved |
| use piecewise-linear kernel instead of Gaussian kernel (ablation) | NLA on OPT-1.3B · inference | time (s): 0.45 -> 0.25 | Improved |
| use piecewise-linear kernel instead of Gaussian kernel (ablation) | NLA on LLaMA-7B · inference | time (s): 1.45 -> 0.82 | Improved |
| use piecewise-linear kernel instead of Gaussian kernel (ablation) | NLA on OPT-1.3B · peak GPU memory | memory (GB): 2.7 -> 2.8 | No clear effect |
| add memory-saving forward Algorithm 1 (ablation) | NLA on OPT-1.3B · inference memory | memory (MB): 18053.07 -> 2204.22 | Improved |
| add memory-saving forward Algorithm 1 (ablation) | NLA on OPT-1.3B · training memory | memory (MB): >81920 -> 12623.75 | Improved |
| use nonlinear NLA instead of linear SVD approximation (no training) (ablation) | LLaMA-7B compression · WikiText-2 | perplexity: 32194 -> 16774 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SVD-LLM 50% — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 9 of 9 shared results
- PELA 29.2% — PELA: Learning Parameter-Efficient Models with Low-Rank Approximation (CVPR 2024) · 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 0.89)
- COALA: Numerically Stable and Efficient Framework for Context-Aware Low-Rank Approximation (NeurIPS 2025 · similarity 0.87)
- Dynamic Low-Rank Sparse Adaptation for Large Language Models (ICLR 2025 · similarity 0.86)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.86)
- From Low Rank Gradient Subspace Stabilization to Low-Rank Weights: Observations, Theories, and Applications (ICML 2025 · similarity 0.86)
- A Stronger Mixture of Low-Rank Experts for Fine-Tuning Foundation Models (ICML 2025 · similarity 0.86)
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.
