[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-generalization":3},{"tag":4,"articles":10},{"id":5,"name":6,"slug":6,"article_count":7,"description_zh":8,"description_en":9},"df98642b-f381-4474-a0c8-a6c4c45ff29f","generalization",3,"泛化描述模型在未見資料、不同分佈或更長推理條件下仍能維持表現的能力。這個主題常連到邊界穩定性、Hessian 光譜、訓練動態，以及 LLM 在換地圖、拉長序列時的失效模式。","Generalization is the ability of a model to keep working on unseen data, shifted distributions, or longer reasoning paths. Here it connects training stability, Hessian-spectrum sharpness, and LLM failures on new maps or longer sequence lengths.",[11,20],{"id":12,"slug":13,"title":14,"summary":15,"category":16,"image_url":17,"cover_image":17,"language":18,"created_at":19},"1b8be06a-85ea-4cd1-a3c7-ffccdc3eefd5","edge-of-stability-generalization-zh","邊界不穩定為何反而更會泛化","這篇論文把高學習率下的混沌訓練，連到泛化能力，並用 Hessian 光譜定義新的 sharpness dimension。","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776837839747-ism8.png","zh","2026-04-22T06:03:36.116147+00:00",{"id":21,"slug":22,"title":23,"summary":24,"category":16,"image_url":25,"cover_image":25,"language":18,"created_at":26},"46ad5553-2eab-41b1-8602-82bf7fb94933","llm-generalization-shortest-path-scale-zh","LLM 會看地圖，卻撐不住長度","這篇合成最短路徑研究把「會換地圖」和「能拉長題目」拆開看，結果發現 LLM 能跨地圖泛化，卻在長度變長時因遞迴推理不穩而失手。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776406013309-pvmm.png","2026-04-17T06:06:33.258278+00:00"]