[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"topic-zh-llm-":3},{"cluster":4,"timeline":16},{"id":5,"slug":6,"title":7,"pinned":8,"status":9,"summary":10,"category":11,"language":12,"created_at":13,"merged_into":14,"article_count":15,"first_seen_at":13,"last_updated_at":13},"a12ef47b-e5a9-493d-9285-34de8da160c4","llm-","LLM 的有害行為藏在哪裡",false,"active","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","zh","2026-07-05T09:03:42.599722+00:00",null,1,[17],{"id":18,"slug":19,"title":7,"summary":10,"category":11,"image_url":20,"cover_image":20,"published_at":21,"is_canonical_seed":22},"7b7926a8-c1ad-4f35-bc80-a32147ca7f89","llms-harmful-content-unified-mechanism-zh","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776060240663-k9nm.png","2026-04-13T06:03:37.743+00:00",true]