[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-mistral-leanstral-proof-engineering-open-model-zh":3,"article-related-mistral-leanstral-proof-engineering-open-model-zh":30,"series-model-release-5d3f053c-03c3-4bf9-9cbb-8fd307a076bc":73},{"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":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"5d3f053c-03c3-4bf9-9cbb-8fd307a076bc","mistral-leanstral-proof-engineering-open-model-zh","Mistral 押注 Leanstral 走向證明工程是對的","\u003Cp data-speakable=\"summary\">87% 的 FATE-H 分數說明，證明工程已經不是研究室展示，而是 \u003Ca href=\"\u002Ftag\u002Fai-工具\">AI 工具\u003C\u002Fa>真正能\u003Ca href=\"\u002Fnews\u002Fethereum-app-list-turns-discovery-into-map-zh\">變成\u003C\u002Fa>基礎設施的產品類別。\u003C\u002Fp>\u003Cp>我支持 Mistral 把 Leanstral 推向證明工程，因為 AI 真正有價值的下一步，不是更會說，而是更能被驗證。\u003C\u002Fp>\u003Ch2>第一個論點\u003C\u002Fh2>\u003Cp>在 Lean 4 這種形式化系統裡，模型不能靠話術混過去，證明要嘛編譯成功，要嘛直接失敗。Mistral 公布的數據很硬：Leanstral 1.5 在 miniF2F 上達到完全飽和，PutnamBench 解出 672 題中的 587 題，FATE-H 則拿到 87%。這些不是漂亮包裝，而是代表模型已經能在機器可檢查的規則下持續產出可接受的證明。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783323188187-clp8.png\" alt=\"Mistral 押注 Leanstral 走向證明工程是對的\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>更重要的是，這種能力不是只對數學有用。對工程團隊來說，最耗時的從來不是寫出一個看起來合理的說法，而是反覆修補引理、處理編譯器回饋、把錯誤一路修到檔案真的過關。能進入這個迴圈的模型，角色就不再是自動補字工具，而是驗證工具。這也是為什麼證明工程會比一般程式生成更接近企業級需求。\u003C\u002Fp>\u003Ch2>第二個論點\u003C\u002Fh2>\u003Cp>Leanstral 的價值不只在數學題，而是在真實程式碼裡找出問題。Mistral 在 \u003Ca href=\"\u002Ftag\u002Frust\">Rust\u003C\u002Fa> 驗證流程中表示，Leanstral 於 57 個儲存庫中標出 47 項違規，其中 11 個是確定的 bug，還有 5 個從未在 \u003Ca href=\"\u002Ftag\u002Fgithub\">GitHub\u003C\u002Fa> 上被回報。這種結果很關鍵，因為它把 AI 從「看起來懂」推進到「真的能幫你抓出錯誤」，而且錯誤是可以直接進入審查流程的。\u003C\u002Fp>\u003Cp>另一個值得注意的訊號，是 270 萬 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa> 的長篇證明嘗試，以及跨越 22 次上下文壓縮仍能完成 AVL 樹時間複雜度證明。這說明問題不只在單步推理，而在長程狀態管理。實務上的工程工作本來就不是一次答完，而是要在很多輪失敗、修正、再試之間維持脈絡。若模型能撐住這種長鏈條，它就有機會進入大型儲存庫、幫忙補輔助引理、追蹤錯誤來源，甚至在初版推理失敗後繼續推進。\u003C\u002Fp>\u003Ch2>反方可能怎麼說\u003C\u002Fh2>\u003Cp>反對者的說法其實很強：證明工程太小眾，工具鏈太專門，絕大多數團隊根本不寫 Lean 4。就算 Leanstral 很強，它也可能只是一個研究展示，而不是能真正擴散到主流開發流程的產品。再加上 Labs 頁面寫明這個模型預計在 \u003Ca href=\"\u002Fnews\u002Fai-weekly-2026-w28-zh\">2026\u003C\u002Fa> 年 9 月 30 日退役，這看起來更像短期實驗，而不是長期平台。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783323175192-dwpa.png\" alt=\"Mistral 押注 Leanstral 走向證明工程是對的\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這個質疑在採用面上成立，但它忽略了戰略重點。Mistral 不需要 Lean 4 變成全民語言，Leanstral 仍然有意義，因為形式化驗證本來就是高槓桿切入點，不是大眾\u003Ca href=\"\u002Fnews\u002Fdaily-huggingface-ai-papers-research-updates-zh\">功能\u003C\u002Fa>。能在 Lean 裡做好的事情，方法上可以外溢到程式修補、性質檢查、儲存庫層級推理。退役日期會削弱這個特定端點作為基礎設施的說服力，但不會削弱整個方向：未來最值錢的 AI 程式工具，會是能被機器驗證的那一類。\u003C\u002Fp>\u003Ch2>你能做什麼\u003C\u002Fh2>\u003Cp>如果你是工程師，不要只看 AI 會不會寫程式，要開始要求它證明自己改對了什麼。如果你是 PM，優先找能形成驗證閉環的流程，不是只產出草稿的流程。如果你是創辦人，應該往錯誤代價高、正確性可量化的領域找切入點，因為證明型代理人最先付費的，就是這種場景。\u003C\u002Fp>\u003Cp>Leanstral 1.5 的重點，不是又多了一個模型，而是它把 AI 從程式生成往程式保證推了一步。這一步很小，方向卻很大，因為真正能進企業、進基礎設施、進關鍵流程的 AI，最後都得學會被驗證。\u003C\u002Fp>","87% 的 FATE-H 分數說明，證明工程已經不是研究室展示，而是 AI 工具真正能變成基礎設施的產品類別。","www.testingcatalog.com","https:\u002F\u002Fwww.testingcatalog.com\u002Fmistral-releases-leanstral-1-5-open-model-for-proof-engineering\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783323188187-clp8.png","model-release","zh","557b8992-f5dd-4c62-8615-eede865e0d01",[17,18,19,20,21],"Mistral","Leanstral","證明工程","形式化驗證","AI 程式工具",[23,24,25],"證明工程把 AI 從「看起來對」變成「可被機器驗證」的工具。","Leanstral 的成績顯示，長程推理與錯誤修正已接近實用門檻。","真正先落地的，不是大眾寫碼助手，而是高正確性、高成本場景。",0,"2026-07-06T07:32:31.794396+00:00","2026-07-06T07:32:31.777+00:00","fa1a9e87-e815-49df-8195-3aa7909f05e5",{"tags":31,"relatedLang":32,"relatedPosts":36},[],{"id":15,"slug":33,"title":34,"language":35},"mistral-leanstral-proof-engineering-open-model-en","Mistral is right to push Leanstral into proof engineering","en",[37,43,49,55,61,67],{"id":38,"slug":39,"title":40,"cover_image":41,"image_url":41,"created_at":42,"category":13},"1b43c314-106c-44f2-a24c-42a82bde3398","google-june-2026-ai-updates-live-translation-zh","Google 2026 年 6 月 AI 更新先押直播翻譯","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783213387695-si5s.png","2026-07-05T01:02:36.092728+00:00",{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":13},"e2712f87-9ae9-4440-bef6-cd2c3d5f2343","mistral-small-2603-pricing-context-benchmarks-zh","Mistral Small 2603：256K 上下文很能打","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783018990238-oknm.png","2026-07-02T19:02:34.781303+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"a6c3f8d1-2698-4d18-af61-2c813605d1ab","doubao-seed-21-pro-agent-balanced-winner-zh","豆包 Seed 2.1 Pro 不是追赶者，而是 Agent 时代的均衡強者","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782910972725-cj17.png","2026-07-01T13:02:26.094239+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"7008aa9b-ddde-48eb-9702-7ab59a4278ef","ace-step-15-local-music-generation-product-zh","ACE-Step 1.5 證明本地音樂生成已經是產品，不是 demo","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782853373095-1hod.png","2026-06-30T21:02:21.571935+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"994e873c-330f-4f5d-9884-3d0aa4da6401","sora-30-seat-electric-aircraft-vtol-tests-zh","Sora 30座電動機首輪VTOL測試過關","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782840783244-971n.png","2026-06-30T17:32:31.996757+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"4c695cb3-da7c-494f-b36f-d184efae3cf0","anthropic-claude-tag-slack-ai-coworker-zh","Claude Tag 把 Slack 變成 AI 同事","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782802983792-5qek.png","2026-06-30T07:02:32.884446+00:00",[74,79,84,89,94,99,104,109,114,119],{"id":75,"slug":76,"title":77,"created_at":78},"58b64033-7eb6-49b9-9aab-01cf8ae1b2f2","nvidia-rubin-six-chips-one-ai-supercomputer-zh","NVIDIA Rubin 把六顆晶片塞進 AI 機櫃","2026-03-26T07:18:45.861277+00:00",{"id":80,"slug":81,"title":82,"created_at":83},"0dcc2c61-c2a6-480d-adb8-dd225fc68914","march-2026-ai-model-news-what-mattered-zh","2026 年 3 月 AI 模型新聞重點","2026-03-26T07:32:08.386348+00:00",{"id":85,"slug":86,"title":87,"created_at":88},"214ab08b-5ce5-4b5c-8b72-47619d8675dd","why-small-models-are-winning-on-device-ai-zh","小模型為何吃下裝置端 AI","2026-03-26T07:36:30.488966+00:00",{"id":90,"slug":91,"title":92,"created_at":93},"785624b2-0355-4b82-adc3-de5e45eecd88","midjourney-v8-faster-images-higher-costs-zh","Midjourney V8 變快了，也變貴了","2026-03-26T07:52:03.562971+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"9e1044b4-946d-47fe-9e2a-c2ee032e1164","xiaomi-mimo-v2-pro-1t-moe-agents-zh","小米 MiMo-V2-Pro 登場：1T MoE 模型","2026-03-28T03:06:19.002353+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"c4b6186f-bd84-4598-997e-c6e31d543c0d","cursor-composer-2-agentic-coding-model-zh","Cursor Composer 2 走向代理式寫碼","2026-03-28T03:13:06.422716+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"e112e76f-ec3b-408f-810e-e93ae21a888a","apple-siri-gemini-distilled-models-zh","Apple Siri 牽手 Gemini 的真相","2026-03-29T04:52:57.886544+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"c679b51f-194a-463b-87fc-7695256ff752","mimo-v2-pro-vs-omni-vs-flash-2026-zh","MiMo V2 Pro、Omni、Flash 怎麼選","2026-04-02T01:18:43.576128+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"3b988fd7-6749-4f01-ba25-c0ad7486dc31","z-ai-glm-5v-turbo-design2code-claude-zh","GLM-5V-Turbo 在 Design2Code 贏了…","2026-04-02T04:03:36.31741+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"975a7aef-030e-41a6-9401-1c6a342be68e","april-2026-ai-model-releases-zh","2026年4月 AI 模型更新追蹤","2026-04-02T08:45:33.308563+00:00"]