[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-skill-self-play-llm-co-evolving-skills-zh":3,"article-related-skill-self-play-llm-co-evolving-skills-zh":29,"series-research-5b14d405-fd26-4768-8ba8-06a62f61baab":76},{"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":11},"5b14d405-fd26-4768-8ba8-06a62f61baab","skill-self-play-llm-co-evolving-skills-zh","Skill Self-Play 讓 LLM 技能共演化","\u003Cp>你在訓練\u003Ca href=\"\u002Fnews\u002Fkpmg-openai-deal-turns-saas-into-agents-zh\">代理\u003C\u002Fa>時，常會卡在兩難：任務太封閉，模型學不到新東西；任務太開放，回饋又容易失真。Skill Self-Play 想處理的就是這個問題。\u003C\u002Fp>\u003Cp data-speakable=\"summary\">Skill Self-Play 用可驗證的技能執行來帶動自我生成，試著同時保留探索空間與訓練訊號。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>研究機構\u003C\u002Fstrong>：arXiv 摘要未明確標註\u003C\u002Fli>\u003Cli>\u003Cstrong>核心數據\u003C\u002Fstrong>：摘要無公開 benchmark 數字\u003C\u002Fli>\u003Cli>\u003Cstrong>突破點\u003C\u002Fstrong>：技能驅動的共演化迴圈\u003C\u002Fli>\u003C\u002Ful>\u003Cp>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.22529\">Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills\u003C\u002Fa> 不是在講單純的自我標註資料，也不是把模型丟進一個固定環境反覆刷分。它想做的是把「技能」變成中介層，讓模型在可驗證的前提下持續長出新任務、學新能力。\u003C\u002Fp>\u003Ch2>這篇在解哪個痛點\u003C\u002Fh2>\u003Cp>摘要先點出一個很現實的瓶頸。現在很多自我演化方法，會在兩種路線之間擺盪。一種是綁在環境裡，回饋很準，但學習範圍容易太窄。另一種是讓模型自由生成任務，空間變大了，卻可能把錯誤訊號也一起放大。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785133975659-tgmu.png\" alt=\"Skill Self-Play 讓 LLM 技能共演化\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這個問題對代理訓練特別傷。因為代理不是只要「答對題目」，而是要在互動中慢慢長出能力。如果訓練訊號不穩，模型可能越練越偏；如果環境太小，能力提升也很難外推到真實任務。\u003C\u002Fp>\u003Cp>作者的切入點是：不要直接在「任務」和「回饋」之間硬拉扯，而是先定義技能，再用技能去組織任務。這樣一來，系統既能在局部維持可驗證性，也能在整體上保持任務多樣性。\u003C\u002Fp>\u003Ch2>方法到底怎麼運作\u003C\u002Fh2>\u003Cp>這篇提出的 Skill-SP，是一個三角色的共演化框架：proposer、solver、dynamic skill controller。三者不是獨立運作，而是放進同一個強化學習迴圈裡互相推動。\u003C\u002Fp>\u003Cp>proposer 的工作，是根據動態抽樣出的技能去產生挑戰性任務。重點不是亂出難題，而是讓任務跟當下的技能空間對齊。這樣產生的題目，才有機會在訓練時被有效驗證。\u003C\u002Fp>\u003Cp>solver 負責解題。它不是只做標準答案式的輸出，而是要在候選解法中探索，逼近自己目前能力的邊界。換句話說，solver 的存在，是把「會不會做」變成可觀察的執行結果。\u003C\u002Fp>\u003Cp>dynamic skill controller 則是整個系統的關鍵。它會看執行回饋，更新並擴充技能庫。這代表技能不是靜態分類，而是會隨著訓練過程長大、重組，甚至改變下一輪任務生成的方向。\u003C\u002Fp>\u003Cp>這種設計的重點，在於把自我演化拆成三件事：生題、解題、更新技能。很多 \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> 訓練方法會把這三件事混在一起，最後很難看出訊號到底從哪來。Skill Self-Play 則是刻意把角色分開，讓訓練邏輯更像一個可管理的系統。\u003C\u002Fp>\u003Cp>摘要還明確說，這是一個 \u003Ca href=\"\u002Ftag\u002Freinforcement-learning\">reinforcement learning\u003C\u002Fa> loop。也就是說，它不是一次性產資料的管線，而是持續互動的訓練過程。任務生成、求解、技能擴張會互相餵資料，形成一個循環。\u003C\u002Fp>\u003Ch2>這篇實際證明了什麼\u003C\u002Fh2>\u003Cp>先講限制。摘要沒有公開完整 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 數字，所以不能從這裡直接讀出提升了多少、用了哪些具體評測集、或是樣本效率到底高多少。這篇能確認的是方向性結果，不是完整數表。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785133974805-31xl.png\" alt=\"Skill Self-Play 讓 LLM 技能共演化\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>作者聲稱，實驗顯示 Skill-SP 在工具使用與推理 benchmark 上，能持續推高強模型的表現上限。這句話的重點不是單次分數，而是「上限」這件事：它想證明這個迴圈\u003Ca href=\"\u002Fnews\u002Fethereum-price-network-demand-not-speculation-zh\">不只是\u003C\u002Fa>補洞，而是真的能把能力往上推。\u003C\u002Fp>\u003Cp>摘要也提到，這個框架對一開始「不對齊」的模型，還能帶來明顯翻轉。這是一個很強的說法，但摘要沒有交代模型名稱、任務細節或翻轉幅度，所以目前只能保守理解成：它不只對本來就強的 backbone 有效，也可能幫起點較差的模型拉回來。\u003C\u002Fp>\u003Cp>另一個值得注意的字眼是「robust evolution engine」。這暗示作者把 Skill-SP 看成一個可重複使用的訓練機制，而不是某個特定 benchmark 的小技巧。也就是說，貢獻不只是結果，而是這個讓技能持續演化的迴圈本身。\u003C\u002Fp>\u003Cul>\u003Cli>摘要提到的評測面向是工具使用與推理。\u003C\u002Fli>\u003Cli>摘要沒有公開完整 benchmark 數字。\u003C\u002Fli>\u003Cli>作者主張它能同時幫助強模型與初始不對齊模型。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>對開發者有什麼啟發\u003C\u002Fh2>\u003Cp>如果你在做 agent，最常見的痛點就是評估。你希望模型能探索，但你也需要知道它到底學到了什麼。Skill Self-Play 的想法，是把驗證放在技能層級，而不是把整個訓練空間壓成一個固定環境。\u003C\u002Fp>\u003Cp>這對工具型代理、推理型代理，或任何需要從自生成任務中學習的系統，都有參考價值。它提供的不是一個萬用解法，而是一個設計原則：局部可驗證，整體可擴張。\u003C\u002Fp>\u003Cp>從系統角度看，這也提醒我們一件事。好的自我提升流程，不一定是最會生新題目的流程，也不一定是最會刷分的流程。更可能是那種能同時做到三件事的系統：產生新任務、驗證執行、擴充下一輪 curriculum。\u003C\u002Fp>\u003Cp>對實作來說，這種架構的魅力在於可拆解。你可以把它理解成 task generator、executor、controller 三個模組，而不是一個混成黑盒的自我學習器。這讓訓練流程比較容易除錯，也比較容易分析哪一段出了問題。\u003C\u002Fp>\u003Ch2>限制與還沒回答的問題\u003C\u002Fh2>\u003Cp>這篇摘要還留了不少空白。它沒有說技能庫有多大、怎麼定義技能、多久會擴張一次，也沒有交代計算成本與穩定性。這些都會影響方法能不能真的\u003Ca href=\"\u002Fnews\u002Fwaic2026-ai-agent-embodied-landing-zh\">落地\u003C\u002Fa>。\u003C\u002Fp>\u003Cp>另外，controller 的品質會是關鍵。如果技能設計不好，或是回饋判斷不夠準，系統還是可能走回狹窄化，或是把噪聲當成有效訊號。也就是說，這個方法不是把問題消掉，而是把問題移到「技能如何被管理」這一層。\u003C\u002Fp>\u003Cp>所以，這篇最重要的訊息不是某個漂亮分數，而是一個訓練架構的方向：讓 \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> 的自我演化，不再只是靠開放式生成硬撐，而是用可驗證的技能執行來維持學習訊號。對正在做 agent 訓練的團隊來說，這是一個值得持續追的設計。\u003C\u002Fp>","Skill Self-Play 用可驗證的技能執行來帶動自我生成，試著同時保留探索空間與訓練訊號。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.22529",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785133975659-tgmu.png","research","zh","779ca356-4e9e-4741-a6ee-398941b44e0c",[17,18,19,20,21],"LLM","self-play","co-evolution","reinforcement learning","agent skills",[23,24,25],"把技能當成自我生成與驗證之間的中介層。","用 proposer、solver、controller 三角色形成共演化迴圈。","摘要未公開完整 benchmark 數字，限制了對成效幅度的判讀。",0,"2026-07-27T06:32:28.599852+00:00","2026-07-27T06:32:28.575+00:00",{"tags":30,"relatedLang":35,"relatedPosts":39},[31,33],{"name":17,"slug":32},"llm",{"name":20,"slug":34},"reinforcement-learning",{"id":15,"slug":36,"title":37,"language":38},"skill-self-play-llm-co-evolving-skills-en","Skill Self-Play lets LLMs co-evolve skills","en",[40,46,52,58,64,70],{"id":41,"slug":42,"title":43,"cover_image":44,"image_url":44,"created_at":45,"category":13},"ac1e5ec4-001e-4ecb-b8d0-0742f3b0287c","explainable-rl-air-traffic-control-zh","可解釋強化學習管空管路由","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785135782953-1dba.png","2026-07-27T07:02:29.948161+00:00",{"id":47,"slug":48,"title":49,"cover_image":50,"image_url":50,"created_at":51,"category":13},"35c33c61-6032-407b-8042-1997fa515ad9","sm4rt-structured-motion-4d-reconstruction-zh","SM4RT 把剛體運動帶進 4D 重建","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1785132182171-ph66.png","2026-07-27T06:02:27.718+00:00",{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":13},"cf300a40-a285-4a1c-a0fb-ddd8fb0c6cce","prompt-engineering-turns-codegen-into-repeatable-workflow-zh","Prompt 工程把 codegen 變成可重複流程","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784923395397-latp.png","2026-07-24T20:02:49.165518+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":13},"30e85daf-b3bd-47dc-a99c-f5fdbdf57a97","prompt-engineering-cheat-sheet-2026-zh","2026 Prompt Engineering 快速手冊","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784890982363-ipaj.png","2026-07-24T11:02:34.935593+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":13},"e37b2e5c-b360-4184-8223-005203aeb2f2","35-chatgpt-research-prompts-better-studies-zh","35 個 ChatGPT 研究提示詞實作指南","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784885594088-voer.png","2026-07-24T09:32:44.664883+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":13},"c6d14983-2dd2-457e-bc16-4122c07dd388","graphvid-interaction-graphs-video-generation-zh","GraphVid 用互動圖控影片生成","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784876580687-1hss.png","2026-07-24T07:02:27.302432+00:00",[77,82,87,92,97,102,107,112,117,122],{"id":78,"slug":79,"title":80,"created_at":81},"f18dbadb-8c59-4723-84a4-6ad22746c77a","deepmind-bets-on-continuous-learning-ai-2026-zh","DeepMind 押注 2026 連續學習 AI","2026-03-26T08:16:02.367355+00:00",{"id":83,"slug":84,"title":85,"created_at":86},"f4a106cb-02a6-4508-8f39-9720a0a93cee","ml-papers-of-the-week-github-research-desk-zh","每週 ML 論文清單，為何紅到 GitHub","2026-03-27T01:11:39.284175+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"c4f807ca-4e5f-47f1-a48c-961cf3fc44dc","ai-ml-conferences-to-watch-in-2026-zh","2026 AI 研討會投稿時程整理","2026-03-27T01:51:53.874432+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"cf046742-efb2-4753-aef9-caed5da5e32e","adaptive-block-scaled-data-types-zh","IF4：神經網路量化的聰明選擇","2026-03-31T06:00:36.990273+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"53a0dc54-0371-4e40-8d5e-74e94a73840c","geometry-aware-similarity-metrics-for-neural-representations-zh","超越距離測量：用微分幾何重新理解神經網路","2026-03-31T06:01:01.241968+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"fee7d472-a775-4b1d-bbc2-1e8bca1bbf8b","on-the-fly-repulsion-in-the-contextual-space-for-rich-divers-zh","讓AI繪圖更有創意：用排斥力提升生成多樣性","2026-03-31T06:01:25.439673+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"a9901203-d69b-447b-8854-15d14eab32b4","vision-aided-beam-prediction-cnn-eca-zh","影像輔助波束預測升級 CNN","2026-04-01T10:00:25.8073+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"b55e7dd4-0a24-4b3d-804d-b0309a03f498","triple-band-fss-mimo-antenna-sub-6-ghz-zh","三頻 FSS MIMO 天線瞄準 sub-6 GHz","2026-04-01T13:18:36.857305+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"f68290bd-e7f3-4b30-ba22-dcd4e0130a66","openclaw-1299-repos-eight-weeks-analysis-zh","OpenClaw 1299 個 Repo 的資料解讀","2026-04-02T05:03:45.208411+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"ed9f80eb-eb02-4d35-8ad4-0ddf428751dd","beam-coherence-aware-combining-mmwave-mimo-zh","毫米波 MIMO 的雙階合併法","2026-04-02T05:27:26.897188+00:00"]