[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-grpo":3},{"tag":4,"articles":10},{"id":5,"name":6,"slug":7,"article_count":8,"description_zh":9,"description_en":9},"3e552e10-4a86-4b1d-aba0-e70b3f9f1e59","GRPO","grpo",0,null,[11,20,27,34],{"id":12,"slug":13,"title":14,"summary":15,"category":16,"image_url":17,"cover_image":17,"language":18,"created_at":19},"a4cb421e-464e-4933-9e1c-6371d3cd1503","prevent-catastrophic-forgetting-llm-fine-tuning-zh","如何防止 LLM 微調災難性遺忘","用 Anchored Weight Decay 在 LLM 微調時降低舊任務漂移，保住原有能力並檢查模型是否回復。","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780730281470-8i97.png","zh","2026-06-06T07:17:28.426709+00:00",{"id":21,"slug":22,"title":23,"summary":24,"category":16,"image_url":25,"cover_image":25,"language":18,"created_at":26},"f16d36a6-00e3-4ca8-ac3f-01179eae9490","fixing-bias-in-multimodal-llm-judges-zh","修正多模態 LLM 評審偏誤","這篇論文提出一套訓練方式，讓多模態 LLM 評審更依賴影像證據，而不是被看起來合理的文字帶偏。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1780380203973-x3h7.png","2026-06-02T06:02:32.493674+00:00",{"id":28,"slug":29,"title":30,"summary":31,"category":16,"image_url":32,"cover_image":32,"language":18,"created_at":33},"bc5e34d0-1e65-4ce0-a130-424da873bd5a","vector-policy-optimization-search-diversity-zh","VPO 讓搜尋更會找答案","VPO 把語言模型訓練成會產生多樣解法，讓 test-time search 在多候選搜尋時更容易挑到好答案。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779432361030-ief3.png","2026-05-22T06:45:29.893413+00:00",{"id":35,"slug":36,"title":37,"summary":38,"category":16,"image_url":39,"cover_image":39,"language":18,"created_at":40},"978080ac-12c3-491c-808b-1920bf107fcd","alphagrpo-self-reflective-multimodal-generation-zh","AlphaGRPO 讓多模態模型學會自我修正","AlphaGRPO 把可驗證、可拆解的獎勵訊號加進多模態模型訓練，讓模型能推理、修正輸出，並在不靠 cold-start 的情況下提升生成表現。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778652662823-p0k7.png","2026-05-13T06:10:34.396837+00:00"]