[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-active-experiment-selection-cheaper-scaling-laws-zh":3,"tags-active-experiment-selection-cheaper-scaling-laws-zh":30,"related-lang-active-experiment-selection-cheaper-scaling-laws-zh":40,"related-posts-active-experiment-selection-cheaper-scaling-laws-zh":44,"series-research-8f6e04cc-3ebe-4cf7-a6a4-ab847fb6e355":81},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":18,"translated_content":10,"views":19,"is_premium":20,"created_at":21,"updated_at":21,"cover_image":11,"published_at":22,"rewrite_status":23,"rewrite_error":10,"rewritten_from_id":24,"slug":25,"category":26,"related_article_id":27,"status":28,"google_indexed_at":29,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":10,"topic_cluster_id":10,"embedding":10,"is_canonical_seed":20},"8f6e04cc-3ebe-4cf7-a6a4-ab847fb6e355","主動選樣讓 scaling law 更省錢","\u003Cp>模型訓練越來越貴，連「先跑哪些小實驗來估 scaling law」都開始燒預算。這篇論文直接把問題攤開來看：如果每個 pilot run 的成本不一樣，怎麼在有限 budget 內，把錢花在最值得的那些實驗上？作者提出一個逐步、帶不確定性判斷的選樣方法，目標是用更少的試跑成本，把 scaling-law fitting 做得更有效。\u003C\u002Fp>\u003Cp>對開發者和做訓練規劃的人來說，這不是抽象的學術題。你如果要排\u003Ca href=\"\u002Fnews\u002F-zh\">大模型\u003C\u002Fa>訓練，通常不會想把預算浪費在一堆對最後外推幫助很小的 pilot experiments。這篇工作的核心，就是把「選哪些 pilot run」本身當成一個最佳化問題，而不是先隨便做一批實驗再說。\u003C\u002Fp>\u003Cp>摘要裡最吸睛的說法是：在不少情況下，這套方法可以只用大約 10% 的總訓練預算，就逼近用完整實驗集做 fitting 的效果。這不是說所有情境都能直接複製，但它至少指出一件事：scaling law 的價值，不一定要靠把所有候選實驗都跑完才能拿到。\u003C\u002Fp>\u003Ch2>這篇論文想解什麼痛點\u003C\u002Fh2>\u003Cp>Scaling laws 的用途很直白：在你還沒真的砸下大筆 compute 之前，先估計模型表現會怎麼隨著資料量、算力或訓練時間變化。它們常被拿來做容量規劃、成本評估，甚至幫團隊決定下一階段值不值得繼續加碼。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1777269833885-q0cj.png\" alt=\"主動選樣讓 scaling law 更省錢\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但問題也很現實：要把這些 law fit 出來，本身就要先做一批 pilot experiments。這些試跑不是免費的。當實驗成本不一樣，且你手上的預算有限時，光是決定「先跑哪幾個」就變成一個需要認真算的問題。\u003C\u002Fp>\u003Cp>論文特別指出，不同實驗的成本和價值不一定成正比。有些 run 很便宜，但對你真正關心的 target region 幫助有限；有些 run 很貴，但如果選對位置，對外推會很有用。若選錯，預算可能花掉了，卻沒有換到\u003Ca href=\"\u002Fnews\u002Fqwen36-27b-open-source-coding-model-zh\">更準的\u003C\u002Fa> extrapolation。\u003C\u002Fp>\u003Cp>所以這篇不是在討論「怎麼把 scaling law 擬合得更漂亮」而已，而是在問更務實的事：在有限 budget 下，怎麼挑出最有資訊量的試跑，讓最後的 scaling-law fit 對你真正要決策的區域最有用。\u003C\u002Fp>\u003Ch2>方法到底怎麼運作\u003C\u002Fh2>\u003Cp>這套方法的關鍵字有兩個：sequential 和 uncertainty-aware。白話說，它不是一次把候選實驗全選完，而是一步一步選。每一步都會看目前的不確定性，判斷下一個實驗最值得補哪裡。\u003C\u002Fp>\u003Cp>這個\u003Ca href=\"\u002Fnews\u002Fai-zh\">設計\u003C\u002Fa>很像 active learning 或 adaptive sampling 的思路，但這裡的目標不是分類準不準，而是讓 scaling-law fitting 的外推更可靠。重點不在於多收集資料，而在於每一筆預算支出能不能有效降低你對目標區域的預測誤差。\u003C\u002Fp>\u003Cp>論文把問題設定成一個有限候選池的選擇問題，而且每個候選實驗有自己的成本。這很重要，因為如果所有 run 成本都一樣，問題會單純很多；但真實世界通常不是這樣。某些設定就是比別的設定貴，資源分配也就不能用同一把尺來看。\u003C\u002Fp>\u003Cp>在這個框架下，方法會根據剩餘預算，逐步挑選最能改善 target region fit 的實驗。也就是說，它不是在追求最大資料量，而是在追求「每花一塊錢，能多拿多少對外推有用的資訊」。這種思路對 compute 是瓶頸的團隊特別實用。\u003C\u002Fp>\u003Cp>如果用工程語言來講，這篇論文的重點不是把訓練流程再複雜化，而是把 pilot selection 變成一個可控的資源配置問題。你要的不是更多 run，而是更會挑的 run。\u003C\u002Fp>\u003Ch2>論文實際證明了什麼\u003C\u002Fh2>\u003Cp>摘要提到，作者在一個多樣化的 scaling-law 任務基準上評估這個方法，而且結果優於傳統的 design-based baselines。這表示它不是只在單一情境下看起來漂亮，而是至少在多個任務上都能打贏一般的設計式選樣方法。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1777269832981-o5t8.png\" alt=\"主動選樣讓 scaling law 更省錢\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>更直接的結果是：這套策略常常能用大約 10% 的總訓練預算，達到接近全量實驗集 fitting 的表現。這句話很值得注意，因為它把「省預算」和「維持效果」兩件事綁在一起了。對很多團隊來說，只要 pilot 階段能少燒一個數量級的成本，就足以改變整個實驗規劃方式。\u003C\u002Fp>\u003Cp>不過，這裡也要老實說：摘要沒有公開完整 benchmark 細節。它沒有列出每個任務名稱、每個 benchmark 的數字、誤差區間，或是和 baseline 的逐項比較。因此，從目前可見的資訊來看，我們能確定的是「整體趨勢有利」，但不能從摘要直接讀出每一個任務到底贏多少。\u003C\u002Fp>\u003Cp>也就是說，這篇論文的主張是有說服力的，但摘要層級還不足以讓你直接判斷它在你自己的工作負載上會不會同樣有效。要做那個判斷，還是得看完整方法與實驗設定。\u003C\u002Fp>\u003Ch2>對開發者有什麼影響\u003C\u002Fh2>\u003Cp>如果你在做模型訓練、容量規劃，或是要決定下一輪要不要加大算力，這篇論文的價值很直接：它提醒你，連「前期試跑」本身都應該是 budget-aware 的。很多團隊會把 pilot experiments 當成固定成本，但這篇工作把它拉回到資源配置的問題上。\u003C\u002Fp>\u003Cp>這對實務很重要，因為 scaling laws 的用途通常不是為了寫報告，而是為了做決策。你想知道再多花多少 compute 值不值得，想知道往哪個區間加碼最有效。若 pilot run 的選擇不夠聰明，後面的外推再漂亮也可能建立在不夠好的資料上。\u003C\u002Fp>\u003Cp>這篇論文也傳達一個很實際的訊號：在實驗成本不均的環境裡，主動選樣可能比固定抽樣更划算。換句話說，不是每個候選 run 都值得同等對待。你真正要做的是，優先買到對目標區域最有用的資訊。\u003C\u002Fp>\u003Cul>\u003Cli>適合 pilot run 成本差異很大的情境。\u003C\u002Fli>\u003Cli>適合你已經有明確 target region，要做外推判斷的情境。\u003C\u002Fli>\u003Cli>適合 full fitting 太貴、不能直接把所有候選實驗都跑完的情境。\u003C\u002Fli>\u003Cli>不適合只想要一次性、固定流程的人，因為這方法是逐步決策。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>限制與還沒回答的問題\u003C\u002Fh2>\u003Cp>這篇論文最大的限制，不是方法本身，而是目前公開資訊只有摘要。摘要沒有交代完整的 uncertainty model，也沒有說明實際的 selection rule 細節。這代表你可以知道它是「逐步、看不確定性、成本感知」的方法，但還不能從摘要看出它到底怎麼實作。\u003C\u002Fp>\u003Cp>另一個限制是 benchmark 的資訊不完整。摘要確實說了有 diverse benchmark，也說方法表現優於 classical design-based baselines，但沒有提供任務清單與具體數字。對研究讀者來說，這會讓你很難直接評估它的泛化範圍。\u003C\u002Fp>\u003Cp>還有一個實務上的問題，是 sequential design 本來就會帶來流程成本。你不是只做一次實驗就結束，而是要選、跑、再評估、再選。這種流程在大型、昂貴的訓練計畫裡很合理，但在較小的專案裡，額外的決策開銷可能就沒那麼划算。\u003C\u002Fp>\u003Cp>即便如此，這篇論文的方向還是很清楚：如果 scaling laws 是你做訓練決策的重要工具，那麼 fitting 它們的過程也應該被當成一個要省錢、要講效率的問題。這個觀點很務實，也很符合現在大型模型訓練的現場。\u003C\u002Fp>\u003Cp>總結來說，這篇工作不是在說「少做實驗也沒差」，而是在說「實驗要做得更精準」。如果摘要中的結果在你的場景也成立，那麼主動選樣有機會讓你用更少的 pilot budget，拿到接近全量 fitting 的價值。\u003C\u002Fp>\u003Cp>論文網址是 \u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.22753\">https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.22753\u003C\u002Fa>。從這份 raw 資料能確定的結論很簡單：它把 scaling-law fitting 從固定流程，推向了更像資源管理的問題；而這件事，對任何要花大錢訓練模型的團隊，都很有現實意義。\u003C\u002Fp>","這篇論文把 pilot experiments 也當成最佳化問題，透過逐步、看不確定性的選樣，在有限預算下挑最有價值的跑法；摘要聲稱常能用約 10% 訓練預算，逼近全量 fitting 的效果。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.22753",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1777269833885-q0cj.png",[13,14,15,16,17],"scaling laws","active experiment selection","sequential experimental design","uncertainty-aware","budget-efficient training","zh",0,false,"2026-04-27T06:03:36.246487+00:00","2026-04-27T06:03:36.12+00:00","done","0a18eb6f-5377-40aa-ba99-336ea4f9ca2c","active-experiment-selection-cheaper-scaling-laws-zh","research","7cebbb83-795e-4b40-9878-51bca1a56c08","published","2026-04-27T09:00:07.534+00:00",[31,33,35,37,39],{"name":13,"slug":32},"scaling-laws",{"name":15,"slug":34},"sequential-experimental-design",{"name":14,"slug":36},"active-experiment-selection",{"name":17,"slug":38},"budget-efficient-training",{"name":16,"slug":16},{"id":27,"slug":41,"title":42,"language":43},"active-experiment-selection-cheaper-scaling-laws-en","Active Experiment Selection for Cheaper Scaling Laws","en",[45,51,57,63,69,75],{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":26},"667b72b6-e821-4d68-80a1-e03340bc85f1","turboquant-seo-shift-small-sites-zh","TurboQuant 與小站 SEO 變化","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778840440690-kcw9.png","2026-05-15T10:20:27.319472+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":26},"381fb6c6-6da7-4444-831f-8c5eed8d685c","turboquant-vllm-comparison-fp8-kv-cache-zh","TurboQuant 與 FP8 實測結果","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778839867551-4v9g.png","2026-05-15T10:10:36.034569+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":26},"c15f45ee-a548-4dbf-8152-91de159c1a11","llmbda-calculus-agent-safety-rules-zh","LLMbda 演算替 AI 代理人立安全規則","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778825503412-mlbf.png","2026-05-15T06:10:34.832664+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":26},"0c02225c-d6ff-44f8-bc92-884c8921c4a3","low-complexity-beamspace-denoiser-mmwave-mimo-zh","更簡單的毫米波波束域去噪器","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778814650361-xtc2.png","2026-05-15T03:10:30.06639+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":26},"9d27f967-62cc-433f-8cdb-9300937ade13","ai-benchmark-wins-cyber-scare-defenders-zh","為什麼 AI 基準賽在資安領域的勝利，應該讓防守方警醒","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778807450006-nofx.png","2026-05-15T01:10:29.379041+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":26},"bc402dc6-5da6-46fc-9d66-d09cb215f72b","why-linux-security-needs-patch-wave-mindset-zh","為什麼 Linux 安全需要「補丁浪潮」思維","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778741449813-s2wn.png","2026-05-14T06:50:24.052583+00:00",[82,87,92,97,102,107,112,117,122,127],{"id":83,"slug":84,"title":85,"created_at":86},"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":88,"slug":89,"title":90,"created_at":91},"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":93,"slug":94,"title":95,"created_at":96},"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":98,"slug":99,"title":100,"created_at":101},"9f50561b-aebd-46ba-94a8-363198aa7091","openclaw-agents-manipulated-self-sabotage-zh","OpenClaw Agent 會自己搞砸自己","2026-03-28T03:03:18.786425+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"11f22e92-7066-4978-a544-31f5f2156ec6","vega-learning-to-drive-with-natural-language-instructions-zh","Vega：使用自然語言指示進行自駕車控制","2026-03-28T14:54:04.847912+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"a4c7cfec-8d0e-4fec-93cf-1b9699a530b8","drive-my-way-en-zh","Drive My Way：個性化自駕車風格的實現","2026-03-28T14:54:26.207495+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"dec02f89-fd39-41ba-8e4d-11ede93a536d","training-knowledge-bases-with-writeback-rag-zh","用 WriteBack-RAG 強化知識庫提升檢索效能","2026-03-28T14:54:45.775606+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"3886be5c-a137-40cc-b9e2-0bf18430c002","packforcing-efficient-long-video-generation-method-zh","PackForcing：短影片訓練也能生成長影片","2026-03-28T14:55:02.688141+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"72b90667-d930-4cc9-8ced-aaa0f8968d44","pixelsmile-toward-fine-grained-facial-expression-editing-zh","PixelSmile：提升精細臉部表情編輯的新方法","2026-03-28T14:55:20.678181+00:00",{"id":128,"slug":129,"title":130,"created_at":131},"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"]