>-

A new method called DenoiseRotator improves pruning of large language models by redistributing parameter importance through learnable orthogonal transformations, making models more robust to pruning under semi-structured sparsity. Evaluations on models like LLaMA3 and Qwen2.5 show significant reductions in performance degradation, with the approach enhancing perplexity and zero-shot accuracy compared to traditional pruning techniques.