[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-alignment":3},{"tag":4,"articles":9},{"id":5,"name":6,"slug":6,"article_count":7,"description_zh":8,"description_en":8},"04fba315-9313-4e38-9e9d-78e8bb9e1481","alignment",2,null,[10],{"id":11,"slug":12,"title":13,"summary":14,"category":15,"image_url":16,"cover_image":16,"language":17,"created_at":18},"7b7926a8-c1ad-4f35-bc80-a32147ca7f89","llms-harmful-content-unified-mechanism-zh","LLM 的有害行為藏在哪裡","這篇研究用權重剪枝發現，有害輸出可能集中在一小組共享權重裡，且和正常能力分開。這能解釋為什麼 jailbreak 很脆弱，也為什麼窄域微調會引發更廣泛的失配。","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776060240663-k9nm.png","zh","2026-04-13T06:03:37.879104+00:00"]