Research map / Architectures and efficiency / Quantization
SkipCat: Rank-Maximized Low-Rank Compression of Large Language Models via Shared Projection and Block Skipping
Yu-Chen Lu, Sheng-Feng Yu, Hui-Hsien Weng, Pei-Shuo Wang, Yu-Fang Hu, Liang Hung-Chun, Hung-Yueh Chiang, Kai-Chiang Wu
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 Cat (shared low-rank projection) to naive SVD low-rank compression (ablation) | SVD-LLM style low-rank compression · WikiText-2 (LLaMA2-7B, 20% compression) | Wikitext-2 perplexity: 8.82 -> 7.84 | Improved |
| add Skip (block skipping) to naive SVD low-rank compression (ablation) | SVD-LLM style low-rank compression · WikiText-2 (LLaMA2-7B, 20% compression) | Wikitext-2 perplexity: 8.82 -> 6.71 | Improved |
| add Cat on top of Skip (ablation) | Skip-only low-rank compression · WikiText-2 (LLaMA2-7B, 20% compression) | Wikitext-2 perplexity: 6.71 -> 6.29 | Improved |
| add 8-bit HQQ quantization to the compressed model (ablation) | SkipCat · WikiText-2 (LLaMA2-7B, 20% compression) | Wikitext-2 perplexity: 6.29 -> 6.29 | No clear effect |
| add Hadamard transform before 8-bit quantization (ablation) | SkipCat 8-bit · WikiText-2 (LLaMA2-7B, 20% compression, 8-bit) | Wikitext-2 perplexity: 6.36 -> 6.32 | Improved |
| add channel scaling before 8-bit quantization (ablation) | SkipCat 8-bit · WikiText-2 (LLaMA2-7B, 20% compression, 8-bit) | Wikitext-2 perplexity: 6.36 -> 6.30 | Improved |
| add both Hadamard and channel scaling before 8-bit quantization (ablation) | SkipCat 8-bit · WikiText-2 (LLaMA2-7B, 20% compression, 8-bit) | Wikitext-2 perplexity: 6.36 -> 6.29 | Improved |
| add column permutation before block skipping (vs FP16 without it) (ablation) | Block skipping · WikiText-2 (LLaMA2-7B, 20% compression, FP16) | Wikitext-2 perplexity: Overflow -> 6.29 | Improved |
| run block skipping without permutation in BF16 instead of FP16 (10% compression) (ablation) | Block skipping without column permutation · WikiText-2 (LLaMA2-7B, 10% compression) | Wikitext-2 perplexity: 13.34 -> 31628.76 | Hurt |
| add LoRA fine-tuning to SkipCat at 20% compression | SkipCat · Zero-shot 7 tasks (LLaMA2-7B, 20% compression) | accuracy drop vs dense (%): 2.20 -> 0.39 | Improved |
| use SkipCat instead of SVD-LLM after LoRA fine-tuning at 20% compression | SVD-LLM · Zero-shot 7 tasks (LLaMA2-7B, 20% compression, LoRA) | average zero-shot accuracy (%): 51.02 -> 54.41 | Improved |
| use SkipCat instead of SVD-LLM at 30% compression | SVD-LLM · Zero-shot 7 tasks (LLaMA2-7B, 30% compression) | average zero-shot accuracy (%): 41.25 -> 48.46 | 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.
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 12 of 12 shared results
- Basis Sharing — Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025) · better on 4 of 4 shared results
- Dobi-SVD — Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives (ICLR 2025) · better on 2 of 2 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
- Compressing Large Language Models using Low Rank and Low Precision Decomposition (NeurIPS 2024 · similarity 0.92)
- SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression (AAAI 2025 · similarity 0.91)
- LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression (NeurIPS 2026 · similarity 0.90)
- MoDeGPT: Modular Decomposition for Large Language Model Compression (ICLR 2025 · similarity 0.90)
- GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024 · similarity 0.90)
- BackSlash: Rate Constrained Optimized Training of Large Language Models (ICML 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.
