Large Language Models (LLMs) can generate toxic content, motivating effective test-time detoxification. Existing representation steering methods often rely o…
机构:北大
来源:arXiv 2610.41165 | AI4Papers 论文推荐平台