Research map / Architectures and efficiency / Quantization
SAES-SVD: Self-Adaptive Suppression of Accumulated and Local Errors for SVD-based LLM Compression
Xing Hu, Dawei Yang, Yuan Cheng, Zhixuan Chen, Zukang Xu
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 cumulative error-aware compensation (CEALC) (ablation) | whitened SVD compression · WikiText2 PPL, LLaMA2-7B | perplexity: 9.34 -> 7.66 | Improved |
| add cumulative error-aware compensation (CEALC) (ablation) | whitened SVD compression · WikiText2 PPL, LLaMA-3-8B | perplexity: 16.59 -> 12.25 | Improved |
| add adaptive coefficient selection (ACES) on top of CEALC (ablation) | SAES-SVD · WikiText2 PPL, LLaMA2-7B | perplexity: 7.66 -> 7.37 | Improved |
| add ACES on top of CEALC (ablation) | SAES-SVD · Avg accuracy, LLaMA-3-8B | accuracy (%): 58.82 -> 60.18 | Improved |
| restrict adaptive alpha range to [0.25, 0.75] rather than [0,1] (ablation) | ACES · WikiText2 PPL, LLaMA2-7B ratio 0.2 | perplexity: 8.96 -> 7.37 | Improved |
| narrow alpha range to [0.4, 0.6] (ablation) | ACES · WikiText2 PPL, LLaMA2-7B ratio 0.2 | perplexity: 7.37 -> 7.63 | Hurt |
| use SAES-SVD instead of SVD-LLM at high compression | SAES-SVD · WikiText2 PPL, LLaMA-3-8B ratio 0.8 | perplexity: 6971.65 -> 181.84 | Improved |
| use SAES-SVD instead of Dip-SVD with no fine-tuning or mixed rank | SAES-SVD · Avg accuracy drop, LLaMA-7B ratio 0.2 | accuracy drop (%): 9.2% -> 3.9% | Improved |
| add ASVD mixed-rank allocation to SAES-SVD (ablation) | SAES-SVD · WikiText2 PPL, LLaMA-7B ratio 0.4 | perplexity: 10.42 -> 7.48 | Improved |
| SAES-SVD vs Dip-SVD on 7-task accuracy at ratio 0.4 (7B) | SAES-SVD · Avg accuracy drop, LLaMA-7B ratio 0.4 | accuracy drop (%): 22.8% -> 21.1% | Improved |
| SAES-SVD versus structured pruning at 7GB | SAES-SVD · WikiText2 PPL, LLaMA-7B 7GB | perplexity: 43.05 -> 10.15 | 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 — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 8 of 8 shared results
- Dobi-SVD — Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives (ICLR 2025) · better on 3 of 3 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
- LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression (NeurIPS 2026 · similarity 0.93)
- Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives (ICLR 2025 · similarity 0.91)
- SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025 · similarity 0.91)
- Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025 · similarity 0.89)
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 2025 · similarity 0.89)
- EvoPress: Accurate Dynamic Model Compression via Evolutionary Search (ICML 2025 · similarity 0.89)
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.
