AI Paper Atlas

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

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data

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 changedWhereMeasuredEffect
Add rotation-constrained compensation (RCPU) after Wanda-sp pruningWanda-sp pruned Llama-7B · WikiText-2, 30% pruning, N_calib=128perplexity: 24.13 -> 18.35Improved
Use rotation compensation instead of FLAP bias compensationpruned Llama-7B · WikiText-2, 30% pruning, N_calib=512perplexity: 18.31 -> 16.99Improved
Use rotation instead of least-squares + ridge update (ablation)pruned Llama-7B · 7-task mean accuracy, 30% pruning, N_calib=128mean accuracy: 51.73 -> 55.50Improved
Use rotation instead of ridge LS at small calibration size (ablation)pruned Llama · WikiText-2 PPL, N_calib=128perplexityImproved
Use ridge LS instead of rotation at N_calib=512 (ablation)pruned Llama · WikiText-2 PPL, N_calib=512perplexityMixed
Add isotropic scaling to the rotation (ablation)RCPU · WikiText-2 PPLperplexity: 30%/128: 18.35 -> 18.21No clear effect
Apply compensation to both o_proj and down_proj instead of one (ablation)RCPU, Llama-7B · 30% pruning, N_calib=128perplexity: 21.94 (none) -> 20.22 (down only) -> 18.91 (o_proj only) -> 18.35 (both)Improved
Rotation-compensated Llama-2-13B versus SliceGPTpruned Llama-2-13B · 7-task mean accuracy, 30% pruningmean accuracy: 50.77 -> 59.62Improved
Use RCPU versus FLAP at low pruning on Llama-7B benchmarkspruned Llama-7B · 10% pruning, mean accuracymean accuracy: 63.80 (FLAP) -> 64.01 (RCPU)No clear effect
Increase calibration samples (ablation)RCPU · WikiText-2 PPLperplexityImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy