Research map / Large language models / Language, safety and interpretability
RCPU: Rotation-Constrained Error Compensation for Structured Pruning of a Large Language Model
Shuichiro Haruta, Kazunori Matsumoto, Zhi Li, Yanan Wang, Mori Kurokawa
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 rotation-constrained compensation (RCPU) after Wanda-sp pruning | Wanda-sp pruned Llama-7B · WikiText-2, 30% pruning, N_calib=128 | perplexity: 24.13 -> 18.35 | Improved |
| Use rotation compensation instead of FLAP bias compensation | pruned Llama-7B · WikiText-2, 30% pruning, N_calib=512 | perplexity: 18.31 -> 16.99 | Improved |
| Use rotation instead of least-squares + ridge update (ablation) | pruned Llama-7B · 7-task mean accuracy, 30% pruning, N_calib=128 | mean accuracy: 51.73 -> 55.50 | Improved |
| Use rotation instead of ridge LS at small calibration size (ablation) | pruned Llama · WikiText-2 PPL, N_calib=128 | perplexity | Improved |
| Use ridge LS instead of rotation at N_calib=512 (ablation) | pruned Llama · WikiText-2 PPL, N_calib=512 | perplexity | Mixed |
| Add isotropic scaling to the rotation (ablation) | RCPU · WikiText-2 PPL | perplexity: 30%/128: 18.35 -> 18.21 | No clear effect |
| Apply compensation to both o_proj and down_proj instead of one (ablation) | RCPU, Llama-7B · 30% pruning, N_calib=128 | perplexity: 21.94 (none) -> 20.22 (down only) -> 18.91 (o_proj only) -> 18.35 (both) | Improved |
| Rotation-compensated Llama-2-13B versus SliceGPT | pruned Llama-2-13B · 7-task mean accuracy, 30% pruning | mean accuracy: 50.77 -> 59.62 | Improved |
| Use RCPU versus FLAP at low pruning on Llama-7B benchmarks | pruned Llama-7B · 10% pruning, mean accuracy | mean accuracy: 63.80 (FLAP) -> 64.01 (RCPU) | No clear effect |
| Increase calibration samples (ablation) | RCPU · WikiText-2 PPL | perplexity | Improved |
Closest papers by meaning
- SlimGPT: Layer-wise Structured Pruning for Large Language Models (NeurIPS 2024 · similarity 0.90)
- Fluctuation-Based Adaptive Structured Pruning for Large Language Models (AAAI 2024 · similarity 0.90)
- Beware of Calibration Data for Pruning Large Language Models (ICLR 2025 · similarity 0.90)
- DISP-LLM: Dimension-Independent Structural Pruning for Large Language Models (NeurIPS 2024 · similarity 0.90)
- PAT: Pruning-Aware Tuning for Large Language Models (AAAI 2025 · similarity 0.89)
- An Efficient Pruner for Large Language Model with Theoretical Guarantee (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.
