.png){kind=icon}
返回
AEMC: Aggregator-efficient model compression for streamlining LLMs
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.132779.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Large language models (LLMs) are increasingly used in practical applications via simple prompt-based interfaces, but their high inference cost limits deployment. Structured pruning offers a promising solution, yet most effective methods require post-training, which poses barriers in real-world scenarios. We identify amplitude mismatch, a hidden state norm discrepancy caused by residual connections, as a key reason for pruning-induced degradation. To address this, we propose Aggregator-Efficient Model Compression (AEMC), a training-free pruning framework that inserts lightweight Aggregator layers to restore residual amplitude and preserve information flow. Without any fine-tuning, AEMC consistently outperforms all existing training-free pruning methods and achieves state-of-the-art performance even compared to post-trained baselines. AEMC offers a practical path toward efficient and scalable LLM deployment in resource-constrained settings.
期刊
IF:
6.5
论文数:
2.5W
被引数:
6.5W
.png){kind=icon}
