Research map / Architectures and efficiency / Pruning and efficient tuning
Zero Sum SVD: Balancing Loss Sensitivity for Low Rank LLM Compression
Ali Abbasi, Chayne Thrash, Haoran Qin, Shansita Sharma, Sepehr Seifi, Soheil Kolouri
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 1 round of truncate-correct-re-truncate (ablation) | ZS-SVD · WikiText2 (ratio 0.4) | perplexity: 45.17 -> 26.92 | Improved |
| increase correction rounds from 1 to 10 (ablation) | ZS-SVD · WikiText2 (ratio 0.4) | perplexity: 26.92 -> 18.49 | Improved |
| add 1 correction round at light compression (ablation) | ZS-SVD · WikiText2 (ratio 0.8) | perplexity: 6.74 -> 6.61 | Improved |
| use zero-sum selection instead of smallest |dL| (ablation) | ZS-SVD · WikiText2 (ratio 0.4) | perplexity: 51.8 -> 45.2 | Improved |
| use zero-sum selection instead of most negative dL (ablation) | ZS-SVD · WikiText2 (ratio 0.4) | perplexity: 182452 -> 45.2 | Improved |
| use zero-sum selection instead of smallest sigma (ablation) | ZS-SVD · WikiText2 (ratio 0.4) | perplexity: 803599 -> 45.2 | Improved |
| enforce per-matrix sigma order with magnitude of dL (ablation) | ZS-SVD · WikiText2 (ratio 0.4) | perplexity: 341.3 -> 51.8 | Improved |
| first-order sensitivity instead of diagonal Fisher (ablation) | ZS-SVD · WikiText2 (60% pruning) | perplexity: 46.32 -> 45.17 | Improved |
| first-order sensitivity instead of diagonal Fisher (ablation) | ZS-SVD · WikiText2 (20% pruning) | perplexity: 6.83 -> 6.74 | Improved |
| project residual onto gradient instead of gradient onto residual (ablation) | ZS-SVD correction · WikiText2 (60% pruning) | perplexity: 48.8 -> 26.9 | Improved |
| project residual onto gradient instead of plain GD step (ablation) | ZS-SVD correction · WikiText2 (60% pruning) | perplexity: 38.3 -> 26.9 | Improved |
| use HQ (half-prune plus quantize) instead of remapping (ablation) | ZS-SVD · WikiText2 (60% pruning) | perplexity: 9.25 -> 6.73 | 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.
- Dobi-SVD — Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives (ICLR 2025) · better on 7 of 8 shared results
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 6 of 6 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 0.92)
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.91)
- Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025 · similarity 0.90)
- SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025 · similarity 0.90)
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 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.
