[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-marlin-greener-llm-inference-datacenters-zh":3,"article-related-marlin-greener-llm-inference-datacenters-zh":36,"series-research-9580adce-69ec-4880-ad8b-227c384cb377":87},{"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":30,"topic_cluster_id":34,"embedding":35,"is_canonical_seed":20},"9580adce-69ec-4880-ad8b-227c384cb377","MARLIN 用多代理 RL 省雲端推理資源","\u003Cp data-speakable=\"summary\">MARLIN 把雲端 LLM 推理視為多代理協調問題，用遊戲理論式強化學習來追求更永續的\u003Ca href=\"\u002Ftag\u002F資料中心\">資料中心\u003C\u002Fa>運作。\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>：多代理遊戲理論 RL\u003C\u002Fli>\u003C\u002Ful>\u003Cp>這篇論文想處理的，\u003Ca href=\"\u002Fnews\u002Fwei-shi-mo-minimax-geng-xiang-xiao-fei-ji-ai-gong-si-er-bu-s-zh\">不是模型\u003C\u002Fa>本身多強，而是模型上線後怎麼用得更省。\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.13496\">MARLIN: Multi-Agent Game-Theoretic Reinforcement Learning for Sustainable LLM Inference in Cloud Datacenters\u003C\u002Fa> 盯上的，是雲端資料中心裡的 LLM 推理負載。作者把它描述成一個需要協調、需要適應的系統問題，而不是單純把請求丟給固定規則就好。\u003C\u002Fp>\u003Cp>對開發者來說，這個切法很實際。LLM \u003Ca href=\"\u002Fnews\u002Fwhy-minimax-m27-self-evolution-matters-zh\">in\u003C\u002Fa>ference 已經不是邊角料工作負載，而是雲服務的一部分。它會跟其他服務搶 GPU、記憶體、電力與排程注意力。只要流量一變，原本看起來合理的配置就可能\u003Ca href=\"\u002Fnews\u002Futah-jazz-2026-roster-injury-report-stats-zh\">開始\u003C\u002Fa>浪費資源，或反過來讓延遲和吞吐出問題。\u003C\u002Fp>\u003Ch2>MARLIN 在解什麼痛點\u003C\u002Fh2>\u003Cp>摘要明確說，LLM 已經在雲端平台裡變得越來越普遍，背後是 AI 消費與企業服務需求在推動。意思很直接：推理不再是偶發任務，而是雲端基礎設施的常態工作之一。既然它變成常態，系統就不能只看「能不能跑」，還要看「跑得是否浪費」。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779084247021-qzhd.png\" alt=\"MARLIN 用多代理 RL 省雲端推理資源\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這裡的核心矛盾是服務品質和永續性之間的拉扯。資源配太多，會浪費算力與能源。資源配太少，又可能造成延遲飆高、吞吐不穩，或使用者體感變差。MARLIN 的出發點，就是把這個拉扯當成一個需要被控制的系統問題。\u003C\u002Fp>\u003Cp>摘要沒有公開完整的系統指標細節，所以我們無法從這份 raw 資料判定它到底優化的是哪個具體數字。它沒有交代是以能源、碳排、利用率，還是其他指標為主，也沒有列出部署環境。能確定的是，作者把資料中心視為多角色互動場景，而不是單一控制器就能解完的最佳化題目。\u003C\u002Fp>\u003Ch2>方法重點：把推理當協調遊戲\u003C\u002Fh2>\u003Cp>MARLIN 這個名字本身就透露方法骨架：Multi-\u003Ca href=\"\u002Ftag\u002Fagent\">Agent\u003C\u002Fa> Game-Theoretic \u003Ca href=\"\u002Ftag\u002Freinforcement-learning\">Reinforcement Learning\u003C\u002Fa>。三個關鍵詞分別代表三層意思。強化學習是讓系統透過互動與回饋慢慢學決策。多代理表示不是只有一個決策者。遊戲理論則是把這些決策者之間的互動也算進去，而不是假設彼此獨立。\u003C\u002Fp>\u003Cp>白話一點，作者是在說：資料中心裡的 LLM 推理，不是單一控制器對單一目標做最佳化，而是多個決策面一起作用的協調問題。某個代理做出的選擇，會改變其他代理看到的狀態、獎勵或限制。這種互相影響，正是遊戲理論派上用場的地方。\u003C\u002Fp>\u003Cp>這個框架對雲端系統很有吸引力。因為真實資料中心通常不是一條直線的控制流程，而是多個元件同時決策。像是請求路由、資源配置、排程，甚至是不同層級的策略調整，都可能在變動負載下彼此牽動。單點最佳化常常會在局部看起來很好，放到整體卻變糟。\u003C\u002Fp>\u003Cp>因此，MARLIN 的重點不是提出一個靜態規則，而是讓控制策略能在互動中學習。這也是它和傳統手工閾值、固定 policy 的差別。從摘要能讀到的訊息是，作者想讓系統對負載變化與資源壓力更有適應性。\u003C\u002Fp>\u003Ch2>論文實際證明了什麼\u003C\u002Fh2>\u003Cp>這裡要先講清楚：目前提供的摘要沒有 benchmark 數字，也沒有比較表。沒有公開的百分比提升、延遲下降、能源節省，或吞吐量變化。也就是說，這份 raw 資料不足以支持任何具體性能宣稱。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779084243448-qjbo.png\" alt=\"MARLIN 用多代理 RL 省雲端推理資源\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>所以，如果你想問 MARLIN 比哪個 baseline 好多少、在哪些 workload 上有效、是否有明確的部署場景，這份摘要都沒有回答。至少在我們拿到的內容裡，沒有實驗設定、沒有數據，也沒有完整評估細節。\u003C\u002Fp>\u003Cp>但這不代表它沒有貢獻。它真正展示的是一種系統建模方式：把永續推理視為學習與協調問題，而不是固定政策問題。對做 inference infrastructure 的人來說，這個視角本身就有價值，因為雲端流量與資源條件本來就不是靜態的。\u003C\u002Fp>\u003Cp>換句話說，這篇摘要證明的是「問題怎麼被重新定義」。它不是在摘要裡證明一個已經量化的 SOTA 結果，而是在提出一個更適合動態資料中心的控制框架。這類研究常見的下一步，通常才會是完整實驗、消融、與不同 baseline 的對照。\u003C\u002Fp>\u003Ch2>對開發者的實際影響\u003C\u002Fh2>\u003Cp>如果你在做推理平台、排程器、autoscaler、或資源管理，這篇 paper 的訊號很明確：永續性正在變成第一級系統目標。LLM serving 的成本不只在算力，還會放大到電力與基礎設施使用效率。只要需求波動大，這些成本就更難靠手工規則穩定壓住。\u003C\u002Fp>\u003Cp>多代理 RL 的吸引力，在於它理論上能讓策略跟著環境變化動態調整，而不是一直靠人工調 threshold。這對多服務共用資料中心的場景特別有感，因為局部決策常常不是獨立的。某個節點的調整，可能會影響整體排程、佇列壓力，甚至其他服務的可用資源。\u003C\u002Fp>\u003Cp>不過，摘要也暴露出一個很現實的限制：我們還不知道它有多好部署。多代理 RL 的常見問題是難穩定、難除錯，獎勵設計也容易出事。摘要沒有說明它怎麼避免這些坑，也沒有交代是否需要線上訓練、離線訓練，或模擬環境。\u003C\u002Fp>\u003Cp>所以，對工程團隊來說，MARLIN 比較像一個值得關注的研究方向，而不是可以直接搬進 production 的方案。它提醒大家，LLM inference 的最佳化不能只看單次請求，而要看整個資料中心裡各個決策者怎麼互相影響。\u003C\u002Fp>\u003Ch2>摘要沒講的關鍵問題\u003C\u002Fh2>\u003Cul>\u003Cli>代理到底在控制資料中心哪一層？\u003C\u002Fli>\u003Cli>優化目標是能源、碳排、利用率，還是別的指標？\u003C\u002Fli>\u003Cli>和簡單 heuristics 或非 RL baseline 相比如何？\u003C\u002Fli>\u003Cli>需要線上學習，還是可用預訓練 policy 部署？\u003C\u002Fli>\u003C\u002Ful>\u003Cp>這些問題很重要，因為它們直接決定方法能不能落地。沒有這些資訊，就很難判斷 MARLIN 是一個概念漂亮但成本高的控制框架，還是真的能在雲端推理場景裡穩定運作。\u003C\u002Fp>\u003Cp>但方向本身是對的。LLM 越往雲端基礎設施深處走，問題就越不是「怎麼更快吐 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa>」，而是「怎麼在電力、容量、排程限制下持續吐 token」。MARLIN 想做的，就是把這件事變成可學習、可協調的系統問題。\u003C\u002Fp>\u003Cp>對\u003Ca href=\"\u002Ftag\u002F台灣開發者\">台灣開發者\u003C\u002Fa>來說，這類研究的價值在於它很接近真實營運現場。當模型服務變成雲端核心負載，效率和永續就不再是附加題，而是架構設計的一部分。MARLIN 提供的是一個值得後續追蹤的控制思路，但目前這份摘要還不足以支持任何性能結論。\u003C\u002Fp>","MARLIN 把雲端 LLM 推理視為多代理協調問題，用遊戲理論式強化學習來追求更永續的資料中心運作。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.13496",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779084247021-qzhd.png",[13,14,15,16,17],"LLM inference","multi-agent RL","game theory","datacenter","sustainability","zh",2,false,"2026-05-18T06:03:35.259834+00:00","2026-05-18T06:03:35.243+00:00","done","b17b03b2-2610-4590-99c1-bfed5f4ae928","marlin-greener-llm-inference-datacenters-zh","research","8b3832ee-9b1b-4684-9d11-919559a92b28","published","2026-05-18T09:00:28.717+00:00",[31,32,33],"把 LLM 推理視為多代理協調問題，而不是單一控制器最佳化。","摘要沒有公開 benchmark 數字，因此無法判定實際提升幅度。","對開發者的啟發是：雲端推理的永續性開始需要動態控制思維。","0c35a120-52fc-41fc-afa3-d404eb934158","[-0.022845136,0.010994284,0.011787959,-0.10262056,-0.02276102,-0.024282273,-0.028462147,0.022431146,0.011113483,0.01021918,-0.0064676967,-0.023075338,0.01925501,0.0007088568,0.11167539,0.021967417,0.0035420898,0.008854091,0.021386089,-0.00870284,0.011371223,-0.012177315,-0.018519253,0.0072028143,-0.0072301915,0.024916772,0.0020487057,0.0071540577,0.03279675,0.0035075943,0.019988412,0.01576668,0.0011234976,0.032161828,0.008189457,0.0013483331,0.006103216,-0.00889486,0.00911558,0.01616692,-0.018938623,-4.523384e-05,0.01173174,0.006834374,-0.01453034,0.013163875,-0.014267502,-0.036858205,-0.034963325,-0.00059716013,-0.012713706,0.018251557,0.007467,-0.15977962,0.009525482,0.013223053,-0.016274925,-0.005331183,0.0023210356,-0.0035596772,-0.021813398,0.01437579,-0.012974836,0.0073883045,-0.02126452,-0.023889419,0.029851753,0.009694815,-0.0065837717,-0.0050011147,-0.03328615,-0.016595604,0.0146361645,-0.02380416,0.02537637,-0.0271131,-0.014892654,0.025976324,0.0142101385,0.038032085,0.035707016,-0.026825212,-0.018374283,-0.012495832,-0.013818104,0.011026323,0.031690862,-0.0076306406,0.015735397,-0.0018362957,0.015025102,-0.0075117443,0.02310913,0.030242063,-0.020043531,0.0042574177,0.005251104,-0.013177461,-0.007527641,-0.027392365,-0.008635841,0.008263029,0.0037810511,0.016235266,-0.0006195321,-0.020288307,0.015516358,0.012314621,-0.008056592,-0.0059966454,0.00030026323,-0.01018002,-0.041630372,0.016955653,-0.014813932,-0.1087843,0.009983939,0.010383536,-0.00821896,-0.0045981617,-0.012384478,0.0049226414,0.010205926,0.030408666,0.009967674,-0.009111449,0.009412867,0.010508589,0.008465758,-0.010530938,0.008530266,0.0036500173,0.007594764,0.018896068,-0.010227423,0.033367544,0.012530001,-0.021044642,-0.044403043,-0.013864118,-0.046379566,0.021370038,0.024277434,0.002966799,-0.0007585374,-0.027945846,-0.045474425,0.013294221,-0.0064801215,-0.0072377,0.021632295,-0.0053412803,-0.0015061796,0.005768472,-0.0016462744,-0.013077369,-0.006705956,-0.005521854,-0.013448251,0.03224281,0.008228289,-0.022369904,-0.001434194,-0.0017847611,-0.02290321,-0.008772734,-0.006195968,-0.0056039514,0.0060536196,0.027729103,-0.0046672253,-0.009121225,-0.014317105,0.023083968,0.010726269,0.0025618693,-0.018728193,0.0059907027,0.0055001085,-0.03570456,-0.002915549,0.00063081965,-0.026974933,0.030612359,-0.017460473,0.0014785546,-0.019178119,-0.0068179565,0.025031522,0.002143129,0.0014055748,-0.011698408,-0.0028041287,-0.011582029,-0.006476065,-0.0383902,0.007861714,0.005373893,-0.015118461,0.009080547,-0.036444448,0.012778616,0.042258106,-0.03650756,-0.007538395,-0.029483967,-0.015888488,-0.007634326,0.023902465,-0.0055424804,0.002445948,0.008380261,0.0032655362,0.0035979878,0.021010105,0.008597036,-0.009206167,-0.00872077,-0.0016681852,-0.01568861,0.030554257,-0.0055873143,-0.0057149576,-0.004066486,-0.006736431,0.009233375,-0.020806856,-0.010659996,-0.026235674,-0.010668006,0.008842407,0.017389935,0.011642682,0.01382686,0.0290958,-0.010326878,-0.029296273,0.023754163,0.010389859,0.025029471,-0.015021485,0.034044284,0.014582224,0.0052974266,0.022186937,-0.007621714,0.013264919,-0.008904644,-0.021895844,-0.0005738739,0.015132865,-0.008943095,0.0041497084,0.011987815,0.022624848,-0.0051586158,-0.02218302,-0.00835944,-0.009864672,0.0053618317,0.00034212487,-0.022291929,-0.00037748035,-0.020080343,0.015200328,0.022862513,0.03999886,-0.0070914896,-0.039692733,-0.0019854677,-0.007760861,0.039083663,-0.009001407,0.012198208,0.019273745,-0.024084972,-0.035133433,0.0324772,0.0029313066,-0.012223835,0.010960274,0.022045394,0.023660848,0.018285427,-0.0042729694,-0.004409501,-0.015822254,0.003958836,-0.03186306,-0.010283203,0.003558979,0.024421506,-0.022787971,0.00999604,0.019850984,-0.01110041,-0.01604424,0.0122060655,-0.0022745728,0.027149828,-0.0089424355,-0.010431021,0.012551924,0.03710035,-0.024112208,0.0012869921,0.012952982,0.038493212,-0.014915377,-0.0098738475,-0.021201046,-0.015162254,0.011863508,-0.018587934,-0.0041859117,-0.011260525,-0.022475125,-0.0077873277,0.01429941,-0.012535459,0.0018922728,-0.019386776,-0.010837401,0.003723327,0.0029865114,0.0040837578,-0.0038347784,0.004213413,0.013639527,-0.015310507,-0.022440096,0.0022262195,0.03523196,-0.021582806,-0.016901879,-0.017940605,-0.012728806,-0.015744632,-0.019695103,0.005058603,0.0038597612,0.010746815,-0.0029569706,0.005260433,-0.010451296,0.01400418,0.021499742,0.024730567,0.00045250624,-0.03279627,0.037187327,0.00133704,0.007690018,-0.019527903,-0.021921193,0.014639399,-0.011764306,0.010868828,0.019203665,-0.011699683,0.012884155,-0.0054475395,-0.008721528,-0.027929226,0.006695135,-0.049362756,0.02927931,0.023795642,-0.0067591565,0.021790342,0.011420122,0.010995238,-0.0025972188,-0.00026787206,0.0073593166,-0.0037899152,-0.009110095,-0.002460535,-0.019902468,-0.005045598,-0.013507961,0.020001652,-0.0001548664,-0.012915752,-0.027366003,0.007200967,0.03058201,0.002735056,0.0005277532,0.008754597,0.009315883,0.006857488,-0.008751151,0.022687763,-0.012778778,-0.008621263,-0.0063247215,0.022304071,-0.017116105,-0.030053155,0.009054912,-0.00093620806,0.0018424726,-0.02218489,-0.0016329138,-0.030643644,-0.014865739,0.01611549,0.042414747,-0.019290453,-0.013101619,0.034778133,0.0023550559,0.015412524,0.00039256603,-0.0020882278,-0.0019780833,0.004908702,0.0068827053,0.0027200368,0.003907738,-0.016109124,0.011115887,-0.005706927,-0.025285669,0.0015029091,-0.01401055,-0.0073720706,-0.019742634,-0.019507872,-0.009527491,-0.017517334,-0.007008646,-0.023684239,-0.00025779707,-0.0053116684,-0.007508139,-0.027895343,-0.026230674,-0.02109258,-0.0029569117,-0.031710282,9.182705e-05,0.011887892,-0.0071114744,0.0025752843,0.0005254491,0.011143832,-0.017448058,0.00893742,0.020527845,0.008033451,0.0072301948,-0.009226077,-0.015073679,0.020735351,-0.012320974,0.021829966,0.006595238,0.0024834003,-0.004066042,-0.007267112,-0.009186234,-0.020791108,0.01684296,-0.013149686,-0.021778477,-0.003483599,0.02321638,0.0019632485,-0.0063128485,-0.007507826,0.046724193,-0.016030755,-0.0038171182,-0.01880758,-0.0017149177,0.019079594,0.021783277,0.003507044,-0.0067593646,-0.017872741,0.02304155,-0.0068122824,-0.0022530253,0.01046072,-0.015813647,0.008094855,-0.0032152606,0.017317556,0.011235408,0.014111409,-0.006284304,0.0003897812,0.0001647953,0.0033181754,0.007140295,-0.010711101,0.00022458014,-0.008605195,-0.010844965,-0.0008948418,0.023119904,-0.004086965,0.017927159,-0.030526059,-0.0010199995,0.008133668,0.0339789,-0.022988215,-0.028762722,0.0061692116,-0.007153581,0.009765699,-0.008153946,-0.016428301,-0.0010972657,0.0065607764,0.011656104,-0.01994028,-0.0032282693,-0.0043417118,-0.011246819,0.025310844,-0.024051988,-0.020547971,-0.0074073845,0.03816762,0.032921135,0.02656648,-0.024431376,0.028079882,0.0052559995,0.0070577166,0.0027610566,-0.0011760149,0.0074875285,0.014267829,-0.00068862754,0.014463042,-0.024875695,0.005550119,0.01942779,-0.013794921,0.040525205,-0.09761633,-0.0057273326,0.0048069274,-0.0017606145,-0.0023498524,-0.0145039875,-0.007712888,-0.009912704,-0.0002724338,0.0052954233,-0.0095610805,0.0025756964,0.018168412,0.025121538,-0.0254808,-0.02185842,-0.01837264,-0.008875926,0.0071775857,0.01400988,0.03316371,-0.0024159309,-0.012799213,0.02116648,0.0075273295,-0.0235168,0.006974655,0.033881832,0.042562537,-0.017659629,-0.021096807,-0.0060248575,0.014824832,0.039900374,0.009692664,5.921409e-05,0.031441987,-0.025375959,0.031431995,0.037149675,-0.002346082,-0.0057464032,-0.053299416,-0.010225985,-0.0051348414,-0.01641604,-0.026230846,0.00087040546,-0.015287478,-0.011376171,-0.0487602,-0.022570265,-0.033867434,-0.022649026,-0.0024506599,0.008901766,-0.009264406,-0.024666205,-0.019613335,0.02433304,-0.026558463,-0.006488752,-0.013184049,0.037609078,-0.011176991,-0.0125335725,0.012078761,0.044777058,0.017588295,-0.0013198461,-0.025821697,-0.02466212,0.00896355,-0.016756183,-0.017545091,0.024585392,0.018804302,0.024037728,0.04125103,0.017804312,-0.029800165,-0.021336976,-0.09433249,-0.015230011,-0.033518,-0.027935969,0.018790124,0.0022618421,0.004737744,-0.028309131,-0.014152091,-0.033141438,0.009388369,-0.0130138155,0.009303019,-0.034719884,0.0069498057,0.008650792,-0.025963312,0.030727081,0.018267823,-0.049332183,0.035774212,-0.012291466,-0.0062441416,-0.015682185,-0.01419634,-0.0090620145,-0.029384686,-0.010762511,0.0074022524,-0.020440623,-0.008289656,-0.13305503,-0.0069788652,0.017564982,-0.0008826299,0.0272003,0.008156454,0.016105594,-0.0041545327,0.0068889773,-0.011730973,0.0035336232,-0.020338755,-0.022299588,-0.01866323,-0.016384304,0.112390935,0.012520969,0.0072352337,-0.0037478276,-0.016138254,0.018975452,0.0022856435,-0.003517269,0.035590984,-0.023750642,0.00700438,0.03407554,-0.019731117,0.0018893118,0.014409623,0.018984707,0.0057359524,0.0053190044,-0.01299078,0.004822556,-0.0018392527,0.008837132,-0.037348215,-0.030813176,0.008919208,0.02076128,4.460808e-05,0.018040612,0.016737197,0.011789824,-0.0061545363,-0.027280137,-0.01752863,0.010220877,0.009606309,-0.03531391,-0.08556261,-0.019688915,0.003524272,0.0007443369,0.005376618,-0.0228894,0.03158846,0.037691683,0.015326504,0.032897323,0.010878892,-0.008329896,0.018460814,-0.0025121449,-0.0016504969,0.0040674238,0.0019526413,0.01084012,0.009594055,0.003990287,-0.00703535,0.01026019,-0.014273908,-0.0043605124,-0.010588023,0.038320474,0.008471482,0.014623638,-0.020902805,0.009550565,-0.0019149751,0.022766117,-0.01652171,0.014773746,-0.01439468,0.018540412,0.0041887295,-0.013143827,-0.00011603266,-0.013554035,0.018931199,-0.008390761,-0.0025478099,0.027975323,0.01964573,-0.040452138,0.024226295,7.088239e-05,0.001815856,0.0032751798,-0.04100816,0.025552224,-0.012685503,-0.018858286,0.01959917,0.018893113,0.015590999,0.027404616,0.0060814912]",{"tags":37,"relatedLang":46,"relatedPosts":50},[38,39,40,42,44],{"name":17,"slug":17},{"name":16,"slug":16},{"name":13,"slug":41},"llm-inference",{"name":14,"slug":43},"multi-agent-rl",{"name":15,"slug":45},"game-theory",{"id":27,"slug":47,"title":48,"language":49},"marlin-greener-llm-inference-datacenters-en","MARLIN tackles greener LLM inference in datacenters","en",[51,57,63,69,75,81],{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":26},"7c89c3bd-48cb-4b4e-942d-bbf0409fc392","cattle-trade-llm-bluffing-bargaining-benchmark-zh","Cattle Trade 要測 LLM 談判 bluffing","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779085437419-b0zw.png","2026-05-18T06:23:27.885037+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":26},"492aa1ec-02ce-491e-ad03-ae804f261f87","weak-rewards-persistent-llm-user-models-zh","弱回饋讓 LLM 記住偏好","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779084838002-5od2.png","2026-05-18T06:13:32.906335+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":26},"e3f8d32d-9094-4717-b9fd-d799de0e521b","weishenme-fensanshi-xitong-yanjiang-bi-buluoge-wenzhang-geng-zh","為什麼分散式系統演講比部落格文章更值得學","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779075234067-fff9.png","2026-05-18T03:33:21.6849+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":26},"0b28782b-fc24-49fc-bc5c-ec9c07c8ad46","wei-shen-me-sora-zheng-ming-ying-pian-ai-hai-mei-zhun-bei-ha-zh","為什麼 Sora 證明影片 AI 還沒準備好走向主流","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779059031003-tsg7.png","2026-05-17T23:03:22.155232+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":26},"aefdd28e-fccb-46ca-a78b-ad6ad718058d","microsoft-mdash-finds-16-windows-flaws-zh","Microsoft MDASH 找出 16 個 Windows 漏洞","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779041037625-66oq.png","2026-05-17T18:03:35.214691+00:00",{"id":82,"slug":83,"title":84,"cover_image":85,"image_url":85,"created_at":86,"category":26},"902b314d-316c-48aa-9a2a-e4d16f32d2ac","browser-exploit-benchmarks-prove-ai-security-here-zh","為什麼瀏覽器 exploit 基準已證明 AI 安全威脅就在眼前","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779019382261-mfmw.png","2026-05-17T08:03:21.360298+00:00",[88,93,98,103,108,113,118,123,128,133],{"id":89,"slug":90,"title":91,"created_at":92},"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":94,"slug":95,"title":96,"created_at":97},"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":99,"slug":100,"title":101,"created_at":102},"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":104,"slug":105,"title":106,"created_at":107},"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":109,"slug":110,"title":111,"created_at":112},"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":114,"slug":115,"title":116,"created_at":117},"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":119,"slug":120,"title":121,"created_at":122},"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":124,"slug":125,"title":126,"created_at":127},"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":129,"slug":130,"title":131,"created_at":132},"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":134,"slug":135,"title":136,"created_at":137},"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"]