[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-glm-5-3-coding-gains-post-training-zh":3,"article-related-glm-5-3-coding-gains-post-training-zh":31,"series-industry-54ada8df-3058-4dd1-94ef-e7e78f611e22":77},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":24,"views":28,"created_at":29,"published_at":30,"topic_cluster_id":11},"54ada8df-3058-4dd1-94ef-e7e78f611e22","glm-5-3-coding-gains-post-training-zh","GLM-5.3 編碼提升，重點在後訓練","\u003Cp>GLM-5.3 的寫程式提升到底從哪裡來？\u003C\u002Fp>\u003Cp data-speakable=\"summary\">GLM-5.3 的編碼能力主要靠後訓練提升，不是換了一個新的基礎模型。\u003C\u002Fp>\u003Ctable>\u003Cthead>\u003Ctr>\u003Cth>項目\u003C\u002Fth>\u003Cth>關鍵規格\u003C\u002Fth>\u003Cth>對編碼的影響\u003C\u002Fth>\u003C\u002Ftr>\u003C\u002Fthead>\u003Ctbody>\u003Ctr>\u003Ctd>基礎模型\u003C\u002Ftd>\u003Ctd>不變\u003C\u002Ftd>\u003Ctd>把改進重心放在後續訓練\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>編碼超參數\u002F資料\u003C\u002Ftd>\u003Ctd>針對性後訓練\u003C\u002Ftd>\u003Ctd>更貼近真實開發任務\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>推理訓練\u003C\u002Ftd>\u003Ctd>多步驟思考\u003C\u002Ftd>\u003Ctd>提升除錯與規劃能力\u003C\u002Ftd>\u003C\u002Ftr>\u003Ctr>\u003Ctd>代理式流程\u003C\u002Ftd>\u003Ctd>可迭代修正\u003C\u002Ftd>\u003Ctd>更適合 code agent\u003C\u002Ftd>\u003C\u002Ftr>\u003C\u002Ftbody>\u003C\u002Ftable>\u003Ch2>1. 後訓練，而不是重做底模\u003C\u002Fh2>\u003Cp>GLM-5.3 的核心訊息很直接：它的寫程式進步，不是靠從頭重訓一個更大的基礎模型，而是靠 pretraining 之後的後訓練，把模型往更會推理、更聽指令、也更會寫 \u003Ca href=\"\u002Fnews\u002Fclaude-code-sessions-message-each-other-zh\">code\u003C\u002Fa> 的方向推。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786773771566-dghc.png\" alt=\"GLM-5.3 編碼提升，重點在後訓練\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這會改變你看模型升級的方式。若只盯著參數量或底模名稱，很容易錯過真正影響編碼表現的環節。對工程團隊來說，該問的不只是「模型多大」，還有「後面怎麼調」。\u003C\u002Fp>\u003Cul>\u003Cli>底模：維持不變\u003C\u002Fli>\u003Cli>改進來源：後訓練\u003C\u002Fli>\u003Cli>結果：編碼行為更穩\u003C\u002Fli>\u003Cli>判斷重點：看訓練流程，不只看規模\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>2. 以程式任務為核心的監督\u003C\u002Fh2>\u003Cp>一個重要來源，是更偏向 code 的監督資料。這類訓練不是只餵一般文字，而是讓模型處理更像開發者日常的任務，例如修 bug、補函式、產生測試、重構，或完成多步驟問題。\u003C\u002Fp>\u003Cp>這種調整之所以有效，是因為寫程式重視的是正確性與約束，不是文筆。模型在聊天裡看起來很會講，不代表它能正確處理 import、邊界條件或 \u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa> 用法。後訓練把這個落差縮小了。\u003C\u002Fp>\u003Cul>\u003Cli>修 bug 與修補程式\u003C\u002Fli>\u003Cli>產生單元測試\u003C\u002Fli>\u003Cli>學習 API 呼叫模式\u003C\u002Fli>\u003Cli>處理跨檔案修改\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>3. 更好的推理軌跡\u003C\u002Fh2>\u003Cp>另一個可能的\u003Ca href=\"\u002Fnews\u002Fdailyarxiv-arxiv-keyword-paper-feed-zh\">關鍵\u003C\u002Fa>，是推理導向的訓練。當模型學會先拆解問題、再輸出答案，它在需要規劃、追蹤依賴關係、或跨檔除錯的任務上，通常會表現更好。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786773761777-9052.png\" alt=\"GLM-5.3 編碼提升，重點在後訓練\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這代表模型學到的不只是語法，還包括如何看懂問題、排出步驟、避免一開始就給出脆弱答案。對 coding assistant 來說，這種能力往往比單一 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 分數更有價值。\u003C\u002Fp>\u003Ccode>典型流程：\u003Cbr>1. 讀錯誤訊息\u003Cbr>2. 判斷可能子系統\u003Cbr>3. 檢查相關程式路徑\u003Cbr>4. 提出修補\u003Cbr>5. 建議驗證測試\u003C\u002Fcode>\u003Ch2>4. 代理式評估與回饋\u003C\u002Fh2>\u003Cp>這篇內容也指向一個更大的趨勢：訓練與評估要貼近實際使用方式。對寫程式而言，這意味著讓模型在能規劃、能檢查輸出、能根據失敗訊號修正答案的流程裡接受訓練。\u003C\u002Fp>\u003Cp>這點對 code \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> 特別重要。只會一次產碼的模型有用，但能在 compiler error 或 test failure 出現後自行改稿的模型，價值高得多。後訓練可以把這種迴圈納入獎勵。\u003C\u002Fp>\u003Cul>\u003Cli>先生成程式\u003C\u002Fli>\u003Cli>再跑檢查或測試\u003C\u002Fli>\u003Cli>讀取失敗訊號\u003C\u002Fli>\u003Cli>回頭修正 patch\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>5. 對模型採購者真正有用的地方\u003C\u002Fh2>\u003Cp>GLM-5.3 提醒我們，benchmark 圖表常常看不出性能是怎麼來的。兩個架構接近的模型，如果其中一個在 code、指令遵循或 agent 行為上做了更強的後訓練，實際表現就可能差很多。\u003C\u002Fp>\u003Cp>所以採購或評測時，該問的是訓練配方，不只是模型名稱。對團隊來說，最好的 coding model 可能不是最新、也不一定是最大，而是後訓練做得更對路的那一個。\u003C\u002Fp>\u003Cul>\u003Cli>先確認底模有沒有變\u003C\u002Fli>\u003Cli>再問後訓練資料是什麼\u003C\u002Fli>\u003Cli>也要看怎麼評測 coding\u003C\u002Fli>\u003Cli>最後確認是否針對 agent 工作流調整\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>哪種適合你\u003C\u002Fh2>\u003Cp>如果你最在意的是寫程式、修 bug、產測試這類\u003Ca href=\"\u002Fnews\u002Fgrok-bot-cloud-agent-copy-template-zh\">工作\u003C\u002Fa>，優先看有明確 code supervision 與 agent 式評估的模型。若你更在意通用聊天品質，底模本身的強弱仍然很重要。\u003C\u002Fp>\u003Cp>對要採購或做 benchmark 的團隊來說，GLM-5.3 的重點很清楚：真正拉開差距的，常常是後訓練選擇，而不是你第一眼看到的參數量或標題分數。\u003C\u002Fp>","4 個後訓練動作解釋 GLM-5.3 為何在不換底模下，仍能明顯提升寫程式能力。","thenewstack.io","https:\u002F\u002Fthenewstack.io\u002Fglm-5-3-post-training-coding\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786773771566-dghc.png","industry","zh","3ba7bf6c-88c7-4563-b5af-243565c5523c",[17,18,19,20,21,22,23],"GLM-5.3","post-training","coding model","code supervision","reasoning","agent workflows","model evaluation",[25,26,27],"GLM-5.3 的編碼提升主要來自後訓練，不是重做底模。","程式任務導向的監督資料，能明顯改善修 bug、測試與重構能力。","採購模型時要看訓練配方與 agent 評估，而不只看參數量或分數。",1,"2026-08-15T06:02:18.29616+00:00","2026-08-15T06:02:18.286+00:00",{"tags":32,"relatedLang":36,"relatedPosts":40},[33,34],{"name":21,"slug":21},{"name":19,"slug":35},"coding-model",{"id":15,"slug":37,"title":38,"language":39},"glm-5-3-coding-gains-post-training-en","GLM-5.3’s coding gains came from post-training","en",[41,47,53,59,65,71],{"id":42,"slug":43,"title":44,"cover_image":45,"image_url":45,"created_at":46,"category":13},"26c83d04-4b98-4e82-95f0-1e94e19d2c92","anthropic-data-center-push-big-capital-backing-zh","Anthropic 50 億美元數據中心背後的 5 個資本動作","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786752165620-j8sc.png","2026-08-15T00:02:21.073927+00:00",{"id":48,"slug":49,"title":50,"cover_image":51,"image_url":51,"created_at":52,"category":13},"057e9280-7f45-499f-85c7-bd678f90c812","dailyarxiv-arxiv-keyword-paper-feed-zh","DailyArxiv 把 arXiv 關鍵字變成每日論文流","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786716168076-7w4o.png","2026-08-14T14:02:22.664653+00:00",{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":13},"1ffd9308-dd6c-4aed-8cc2-91587d8e72a9","claude-vs-chatgpt-2026-claude-bi-jiao-qiang-ma-zh","Claude vs ChatGPT（2026）：Claude 比較強嗎？","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786710773369-l0tu.png","2026-08-14T12:32:29.306266+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":13},"5680d9ca-166f-4269-82fe-91146cf9ed13","cbdc-governance-standards-stakeholder-roles-zh","CBDC 治理看誰管、誰負責","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786672975574-0kxc.png","2026-08-14T02:02:29.052638+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":13},"46216108-15ba-40ae-811b-ecb63ecb5c7d","invisible-ai-watermarks-right-move-claude-zh","Claude 的無形 AI 水印是對的選擇","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786665770533-24wm.png","2026-08-14T00:02:27.653275+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":13},"8511b941-baf5-48d4-b941-675fa415e66d","pgvector-100k-vectors-not-default-zh","100K 向量以下，pgvector 就夠了，不該當預設","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1786656763977-1sw8.png","2026-08-13T21:32:18.3739+00:00",[78,83,88,93,98,103,108,113,118,123],{"id":79,"slug":80,"title":81,"created_at":82},"ee073da7-28b3-4752-a319-5a501459fb87","ai-in-2026-what-actually-matters-now-zh","2026 AI 真正重要的事","2026-03-26T07:09:12.008134+00:00",{"id":84,"slug":85,"title":86,"created_at":87},"83bd1795-8548-44c9-9a7e-de50a0923f71","trump-ai-framework-power-speech-state-preemption-zh","川普 AI 框架瞄準電力、言論與州權","2026-03-26T07:12:18.695466+00:00",{"id":89,"slug":90,"title":91,"created_at":92},"ea6be18b-c903-4e54-97b7-5f7447a612e0","nvidia-gtc-2026-big-ai-announcements-zh","NVIDIA GTC 2026 重點拆解","2026-03-26T07:14:26.62638+00:00",{"id":94,"slug":95,"title":96,"created_at":97},"4bcec76f-4c36-4daa-909f-54cd702f7c93","claude-users-spreading-out-and-getting-better-zh","Claude 用戶更分散，也更會用","2026-03-26T07:22:52.325888+00:00",{"id":99,"slug":100,"title":101,"created_at":102},"bd903b15-2473-4178-9789-b7557816e535","openclaw-raises-hard-question-for-ai-models-zh","OpenClaw 逼問 AI 模型價值","2026-03-26T07:24:54.707486+00:00",{"id":104,"slug":105,"title":106,"created_at":107},"eeac6b9e-ad9d-4831-8eec-8bba3f9bca6a","gap-google-gemini-checkout-fashion-search-zh","Gap 把結帳搬進 Gemini","2026-03-26T07:28:23.937768+00:00",{"id":109,"slug":110,"title":111,"created_at":112},"0740e53f-605d-4d57-8601-c10beb126f3c","google-pushes-gemini-transition-to-march-2026-zh","Google 把 Gemini 轉換延到 2026 年 3…","2026-03-26T07:30:12.825269+00:00",{"id":114,"slug":115,"title":116,"created_at":117},"e660d801-2421-4529-8fa9-86b82b066990","metas-llama-4-benchmark-scandal-gets-worse-zh","Meta Llama 4 分數風波又擴大","2026-03-26T07:34:21.156421+00:00",{"id":119,"slug":120,"title":121,"created_at":122},"183f9e7c-e143-40bb-a6d5-67ba84a3a8bc","accenture-mistral-ai-sovereign-enterprise-deal-zh","Accenture 攜手 Mistral AI 賣主權 AI","2026-03-26T07:38:14.818906+00:00",{"id":124,"slug":125,"title":126,"created_at":127},"191d9b1b-768a-478c-978c-dd7431a38149","mistral-ai-faces-its-hardest-year-yet-zh","Mistral AI 迎來最硬的一年","2026-03-26T07:40:23.716374+00:00"]