作者位次 #7/11ACL 2026Poster· Natural Language Processing
Projecting Out the Malice: A Global Subspace Approach to LLM Detoxification
Zenghao Duan · Zhiyi Yin · Zhichao Shi · Liang Pang · Shaoling Jing · Zihe Huang · Jiayi Wu · Yu Yan · Jingcheng Deng · Huawei Shen · Xueqi Cheng
摘要
Large language models (LLMs) exhibit exceptional performance but pose inherent risks of generating toxic content, restricting their safe deployment. While traditional methods (e.g., alignment) adjust output preferences, they fail to eliminate underlying toxic regions in parameters, leaving models...