[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-llmbda-calculus-agent-safety-rules-zh":3,"tags-llmbda-calculus-agent-safety-rules-zh":36,"related-lang-llmbda-calculus-agent-safety-rules-zh":46,"related-posts-llmbda-calculus-agent-safety-rules-zh":50,"series-research-c15f45ee-a548-4dbf-8152-91de159c1a11":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},"c15f45ee-a548-4dbf-8152-91de159c1a11","LLMbda 演算替 AI 代理人立安全規則","\u003Cp data-speakable=\"summary\">這篇論文用形式化演算描述 LLM 代理人的對話與資訊流，目標是把安全規則變成可證明的約束。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2602.20064\">LLMbda Calculus: AI Agents, Conversations, and Information Flow\u003C\u002Fa> 想處理的，不是模型答得準不準，而是代理人系統怎麼在多輪對話、工具呼叫、產生程式碼之間維持安全邊界。它把對話本身納入語意模型，讓系統可以開始談「哪些資訊可以影響這次 LLM 呼叫」，而不是只靠提示詞習慣或工程經驗去猜。\u003C\u002Fp>\u003Cp>這件事很實際。當一個 LLM 代理人能讀歷史訊息、呼叫工具、保存狀態，甚至產生下一步要執行的程式碼時，整個流程就不再只是單次推理。任何一段不受信任的輸入，都可能沿著對話鏈、工具輸出或子流程一路滲透到敏感決策裡。這篇論文就是要把這種風險，變成可以形式化描述、也可以證明的問題。\u003C\u002Fp>\u003Ch2>這篇在解什麼痛點\u003C\u002Fh2>\u003Cp>代理人系統最大的麻煩，不是「模型會不會胡說」，而是「系統會不會把不該混在一起的東西混在一起」。一個 LLM 呼叫可能同時依賴前文、工具結果、生成程式碼，還有其他子對話。只要其中一個來源是不可信的，整個流程就可能被帶偏。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778825503412-mlbf.png\" alt=\"LLMbda 演算替 AI 代理人立安全規則\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>摘要點出三個核心關切：被隔離的子對話、生成程式碼的隔離，以及資訊流限制。白話一點說，作者想讓你能把代理工作流切成可信與不可信的部分，然後證明這些邊界真的有守住，而不是只在文件上寫得很好看。\u003C\u002Fp>\u003Cp>這篇不是在追求準確率，也不是在比延遲。它關心的是安全性質，尤其是當 LLM 被放進更大系統後，怎麼避免一段輸入悄悄影響到不該碰的輸出或執行路徑。\u003C\u002Fp>\u003Ch2>方法到底怎麼運作\u003C\u002Fh2>\u003Cp>作者的核心做法，是定義一個 calcul\u003Ca href=\"\u002Fnews\u002Fwhy-nebius-ai-pivot-is-more-real-than-hype-zh\">us\u003C\u002Fa>，也就是一套形式系統，直接把「對話」當成語意的一部分來建模。這點很關鍵。它不是把聊天看成一串字串，也不是把 prompt、tool call、state 當成分散的應用邏輯，而是把對話結構本身視為程式語意的一部分。\u003C\u002Fp>\u003Cp>一旦對話結構進到語意裡，系統就能表達哪些子對話是 quarantined、哪些生成程式碼要被隔離、哪些輸入可以影響某次 LLM 呼叫。也就是說，安全規則不是後面再補一層防火牆，而是直接長在語言模型裡。\u003C\u002Fp>\u003Cp>這篇摘要也明確提到，它支援對資訊流限制做推理。這代表形式系統可以用來談「某些敏感資料不能跨過某條邊界」，或「未受信任的資料不能以錯誤方式影響受保護的計算」。對做代理人的工程師來說，概念上很像 policy-aware 的執行模型，只是這次是用形式化語意來保證。\u003C\u002Fp>\u003Cp>如果要用工程語言翻譯，就是把 \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> workflow 做成一個有規則的執行圖。不同對話段落可以被隔離，資料的影響範圍也可以被限制。這不是靠約定俗成，而是靠模型本身來定義哪些流動是允許的。\u003C\u002Fp>\u003Ch2>論文實際證明了什麼\u003C\u002Fh2>\u003Cp>摘要裡最重要的結果，是一個 termination-insensitive no\u003Ca href=\"\u002Fnews\u002Fnvidia-backs-corning-factories-with-billions-zh\">nin\u003C\u002Fa>terference theorem。這個定理對應到完整性與保密性保證。\u003Ca href=\"\u002Fnews\u002Flow-complexity-beamspace-denoiser-mmwave-mimo-zh\">簡單\u003C\u002Fa>講，noninterference 關心的是：秘密資料或不可信輸入，不應該以違反政策的方式改變受保護的輸出。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778825459324-jgzm.png\" alt=\"LLMbda 演算替 AI 代理人立安全規則\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>但這裡有個重要限定詞：termination-insensitive。這表示定理沒有把「是否終止」這種通道算進去。換句話說，它提供的是一種強但不是全包的保證。對某些資訊流問題，它能給出形式化保證；但終止通道仍然不在這個結果的範圍內。\u003C\u002Fp>\u003Cp>摘要沒有公開完整 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 細節，也沒有任何數字。沒有吞吐量、沒有 token 成本、沒有延遲、也沒有實驗比較。所以這篇不是在展示某個系統跑得更快，而是在展示一個形式模型可以成立，還能證明安全性質。\u003C\u002Fp>\u003Cp>這也讓它更像基礎研究，而不是系統優化論文。它的價值在於定義了一個可以推理的語意框架，並且證明這個框架能支撐資訊流安全，而不是證明某個產品或框架立刻可部署。\u003C\u002Fp>\u003Ch2>對開發者有什麼影響\u003C\u002Fh2>\u003Cp>如果你在做 \u003Ca href=\"\u002Ftag\u002Fai-agent\">AI agent\u003C\u002Fa>，真正難的常常不是讓模型回答，而是讓整個外圍系統在多輪互動下仍然可控。這篇論文的訊號很清楚：conversation structure 本身應該被當成一等公民，而不是把它當成一堆附加字串處理。\u003C\u002Fp>\u003Cp>這對實作的啟發是，未來你可能會更常看到這種思路：\u003C\u002Fp>\u003Cul>\u003Cli>把代理人子流程當成可隔離的單位，而不是全域共享上下文。\u003C\u002Fli>\u003Cli>把生成程式碼視為需要隔離與約束的輸出，不是預設可信。\u003C\u002Fli>\u003Cli>把「哪些資料能影響這次 LLM 呼叫」寫成正式政策，而不是只靠團隊共識。\u003C\u002Fli>\u003Cli>讓安全審查從「看起來應該沒問題」進一步變成「模型裡可證明沒問題」。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>對做 \u003Ca href=\"\u002Ftag\u002Fcopilot\">copilot\u003C\u002Fa>、自治代理、工具型助理的團隊來說，這種形式化特別有價值。因為一旦系統要碰到敏感資料、跨服務工具鏈，或多個子對話並行，單靠 prompt hygiene 很難讓人放心。\u003C\u002Fp>\u003Cp>不過，這篇摘要沒有說它怎麼對應到實際的程式語言、runtime，或現成的 agent framework。也沒有說在真實系統裡要怎麼落地 enforcement。這表示它先解的是語意與證明問題，離工程部署還有一段距離。\u003C\u002Fp>\u003Ch2>限制與未解問題\u003C\u002Fh2>\u003Cp>從摘要能看出的限制也很明確。首先，這是一個形式系統，不是實測系統，所以你看不到它在真實工作負載下的成本。其次，termination-insensitive noninterference 雖然有用，但它不涵蓋所有 side-channel。尤其是終止行為本身可能還是資訊洩漏的一部分。\u003C\u002Fp>\u003Cp>另外，這篇只承諾在它定義的形式模型裡提供完整性與保密性。它沒有直接宣稱可以解決 prompt injection、tool misuse、sandbox escape，或所有實務上的攻擊面。這些問題在真實 agent 系統裡仍然存在，而且通常比形式證明更難處理。\u003C\u002Fp>\u003Cp>所以比較務實的看法是：這篇不是在說 LLM agent 已經安全了，而是在說，如果你真的想把 agent 當成嚴肅軟體元件，那你需要比「一個聊天視窗加幾個工具」更好的語意基礎。LLMbda calculus 提供的，就是這種基礎的雛形。\u003C\u002Fp>\u003Cp>對\u003Ca href=\"\u002Ftag\u002F台灣開發者\">台灣開發者\u003C\u002Fa>來說，這類研究的價值不一定是馬上拿來上線，而是幫你重新定義問題。當系統開始處理敏感資料、內部知識、或可執行動作時，安全邊界不能只靠 prompt 寫法。你需要能描述、能隔離、最好還能證明的資訊流規則。這篇論文就是朝那個方向走的一步。\u003C\u002Fp>","這篇論文用形式化演算描述 LLM 代理人的對話與資訊流，目標是把隔離、保密與完整性變成可證明的安全規則。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2602.20064",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778825503412-mlbf.png",[13,14,15,16,17],"LLM agent","information flow","noninterference","formal calculus","conversation semantics","zh",0,false,"2026-05-15T06:10:34.832664+00:00","2026-05-15T06:10:34.608+00:00","done","09969775-4461-4629-9638-a8c0388d5c86","llmbda-calculus-agent-safety-rules-zh","research","5aef1c57-961f-49f7-8277-f83f7336799a","published","2026-05-15T09:00:14.576+00:00",[31,32,33],"這篇把 LLM 代理人的對話與工具流程做成形式化演算，重點是資訊流安全，不是模型準確率。","論文主張可把子對話、生成程式碼與 LLM 呼叫的影響範圍納入同一套語意模型。","摘要只公開形式結果，沒有 benchmark 數字；已知限制是 termination-insensitive，且未涵蓋所有實務攻擊面。","0c35a120-52fc-41fc-afa3-d404eb934158","[-0.025262767,0.006909106,0.018274235,-0.067886755,0.0014029013,-0.008207455,-0.005738952,0.017292388,0.03713665,0.0074439356,0.0071702064,-0.0041317716,-0.0007898715,-0.00032138496,0.12534119,0.027464705,0.005780468,0.050627973,0.01499755,-0.021711823,0.017471684,0.003683197,-0.030857652,-0.002611954,-0.0183127,0.010030404,0.012403155,0.016834496,0.03722334,0.016012201,-0.002651555,0.025691008,-0.0072345147,0.030021094,0.015333777,0.033304337,0.015508003,-0.015020312,0.019323595,0.02701672,0.008403236,-2.4973075e-05,0.0036050952,-0.016638063,0.024059065,0.010665179,-0.0019398074,-0.03291026,-0.022848161,0.04830995,-0.01733625,0.0075361924,0.00074719527,-0.15805767,-0.012588495,-0.007711462,0.011172085,-0.00035859775,0.0035510298,0.021888822,-0.017409187,0.027085971,-0.0204665,0.010679028,-0.0020866047,-0.003940616,0.024207076,0.010683956,-0.037030187,-0.019550346,-0.030057356,0.02299344,0.021220954,-0.03755758,0.0013863172,-0.026943136,-0.015340704,-0.019537676,0.01345484,0.013958437,-0.019006366,-0.034265384,0.014706385,-0.011259844,-0.010050674,-0.015322096,0.03708606,0.019497726,0.0134637635,0.025609251,-0.02664231,-0.0009897504,0.02667894,0.030681107,-0.0006641618,0.0024914024,-0.02181087,0.0070577823,0.0006903619,-0.010889518,-0.033259783,-0.005675881,-0.0057189916,0.0033681376,-0.004730735,0.0046321778,-0.009997878,-0.008544975,-0.0016993954,0.03705121,0.025675444,-0.024985626,-0.018372318,-0.019784262,0.016368095,-0.12472064,0.010818352,0.002693522,-0.0021555023,0.014286826,-0.0049012625,-0.009990853,0.010023396,0.06315913,0.020229967,0.0014547816,0.008461793,0.010353342,-0.020366844,0.013436108,-0.0130701065,0.0002472978,0.01363371,-0.007544927,0.007788265,0.029538369,-0.0031367498,-0.0029138578,-0.035598665,-0.019843018,-0.017716909,0.03224251,0.012330853,0.017063122,0.0011503687,-0.018759446,-0.056419034,-0.017210228,-0.0013080544,-0.017907754,0.027081463,-0.0021943392,-0.020737523,-0.009405079,0.010945986,-0.033908226,0.0062355725,0.0055398964,0.011625181,0.017782973,-0.008602748,0.010536943,-0.01434205,0.008181846,0.0028191002,0.023319153,-0.00050846854,-0.026631724,-0.012604766,0.027269265,-0.0017954761,-0.034017935,0.0003191257,0.0070588198,0.0342208,0.019130073,-0.014636054,-0.01012885,-0.0023985594,-0.03831744,-0.0066292468,-0.007845318,-0.02019531,0.04783254,0.0037682245,0.019354777,0.0090362225,0.009694827,0.015874993,0.011542805,-0.03675924,-0.007687161,0.0064470754,-0.0056752674,0.0010336835,-0.017874498,0.009381882,0.008655824,-0.021123512,0.027185274,-0.0007176575,-0.012957163,0.020387748,-0.017699549,0.011860798,-0.018005595,0.007907985,-0.015553873,-0.003381598,0.0027368339,-0.013272548,-0.011255205,0.015471005,-0.008076302,0.027660312,-0.0066724615,0.009651932,0.0038003516,-0.0065680016,-0.023942614,0.025688699,-0.009500973,-0.0041795834,0.010097969,-0.004428734,-0.02074044,0.008156705,-0.0005374022,-0.013911263,0.0058360253,-0.009134609,0.03263531,0.015460767,0.011928327,0.017019177,-0.00065979536,0.017652785,0.012223256,0.007921142,0.01260251,-0.0019591195,0.013805668,-0.008506325,0.024586106,0.021607202,-0.013883473,-0.0006419242,-0.01129206,-0.02429504,0.021836301,-0.0019624624,-0.0025590046,0.0043476783,-0.023188068,0.012647784,-0.004840558,-0.0046180873,0.013356099,-0.021813508,0.02001165,-0.0061368276,0.005045428,-0.040907823,0.0025334617,0.005009825,-0.0064585605,0.0019175818,0.004541652,-0.016519403,-0.0046454403,-0.021548593,0.037099417,0.01654949,-0.0038003265,0.030729633,-0.009604484,-0.05831764,0.048925344,0.033232484,0.017044444,0.016612815,-0.00087595603,-0.004591443,0.020039545,-0.0025591918,-0.002754808,-0.01811728,-7.9056235e-06,-0.014852552,0.00020223921,0.0089754015,0.018267376,-0.013888476,0.011754031,-0.021361591,-0.018229095,0.011078792,0.032649152,0.0027984614,-0.021156441,0.013118266,-0.016399832,0.0046465304,0.047169298,-0.026460068,-0.013280591,-0.025828332,0.023872912,-0.017221468,-0.027338242,0.008653174,-0.015792904,0.03048368,-0.009629792,-0.015566457,0.0038171315,0.008458421,-0.01656048,0.027896844,-0.025351219,0.0038139503,-0.025937973,0.0049480773,0.016844414,-0.0034834452,0.004279668,0.005715192,0.00055663014,0.031442657,-0.0052056727,-0.022886006,0.015324737,0.022650905,-0.005692044,0.02161591,-0.0060073864,-0.020079752,-0.007064308,-0.017399227,0.0072028236,0.0034918534,0.017663527,0.01130233,0.006549982,-0.011488871,0.015729656,0.033407915,0.011146735,0.0033955702,-0.025829824,0.03922161,0.012054225,-0.004627597,-0.014342653,-0.043911215,0.00039957452,0.0039774934,-0.001265702,0.022064997,-0.0058937673,0.004778716,0.008862081,-0.02010816,0.0017003716,0.021847153,-0.037502468,0.0052115787,0.022114936,-0.008465665,0.0050493535,0.014759481,-0.019826071,0.00069980195,0.0016378658,-0.006717453,-0.016130483,-0.033466402,-0.015001847,-0.018806756,0.01824952,0.0029027115,0.039730582,-0.020230472,-0.010567015,-0.03023343,0.033740677,0.011967282,0.00736134,-0.013190503,0.013057172,-0.00738045,0.019511774,0.002249893,0.0018006288,0.007899124,-0.009992233,0.014368883,0.03150726,-0.00683234,0.004201305,0.006306348,0.026011685,0.015148589,-0.020162342,0.006004449,-0.0039698984,0.0068564387,-0.013816106,0.0383826,-0.0018559225,-0.0025048915,0.024762418,0.026329223,-0.007242268,0.0070147756,0.013414933,-0.014558404,-0.01302543,0.006215942,0.0042185234,0.004973648,-0.010768552,-0.0030417405,-0.010508724,0.0045434274,-0.021073956,-0.024247203,-0.030559523,-0.018418312,-0.024657695,-0.010225466,-0.0111516155,-0.004582792,-0.020099347,0.022892524,-0.017715864,-0.0144666415,-0.03485428,-0.010404787,-0.020309862,-0.014388333,0.002452504,-0.0018339043,0.022647083,0.03407105,-0.0090676285,0.0065696333,-0.024976961,-0.019129414,0.022261359,0.019332776,0.00051975995,0.010022831,-0.003697452,-0.0030025314,-0.0025497407,-0.01108381,0.008174582,0.0017341684,-0.012358379,0.017147895,-0.003644475,0.016862234,0.003368797,0.0015606971,0.0070548104,-0.025419608,6.630157e-06,0.008203003,0.03559773,-0.034118116,-0.011481155,0.016475981,-0.0022915192,0.016547346,-0.006320954,-0.0248051,0.031249836,-0.011061773,0.0002766167,0.0067870324,-0.014325344,-0.002877985,-0.016795188,-0.017703023,0.029981436,-0.026798239,-0.001099214,-0.030102072,0.0034738742,0.015457101,0.0044813463,0.012008532,0.00052800955,-0.014525206,-0.014345938,0.014182543,-0.01425994,0.010451256,-0.013905089,-0.0014550482,-0.030685425,0.0023427082,-0.014544669,-0.0020806063,-0.011607563,-0.026761202,0.004386,0.00019629701,-0.025870932,-0.029833406,-0.006765408,0.013194449,0.0074317018,0.0023547395,-0.014301653,-0.00098271,0.0056389025,-0.012267824,-0.0005759276,0.012309894,-0.010722023,0.01618308,0.0321299,-0.004806719,-0.02865044,0.016691597,0.018119546,0.017790658,0.0315778,-0.025885038,0.018860497,0.022711225,-0.019111492,0.030308403,-0.003612843,0.02271892,0.0026506283,0.0006966023,0.010673274,-0.03297019,-0.01113076,-0.0043622716,-0.026592238,0.03176791,-0.08990221,-6.957141e-05,0.0111117335,0.00440326,-0.010215087,-0.023384416,0.019288054,0.0005078045,-0.008231766,0.0040984266,0.0088291345,-0.01165458,0.05328682,0.002421972,-0.0016225729,-0.030287724,-0.02688378,0.005941745,0.0043834317,-0.026809888,0.02929347,-0.0032823016,-0.0011760925,0.012889402,0.00082824816,-0.014059093,0.0021121695,0.012395762,0.012627156,0.00036068604,-0.008512918,-0.0119215995,-0.016246626,0.00899256,-0.012903575,0.0124204485,0.013253051,-0.029921638,0.013714261,0.023488091,-0.01396147,0.018144231,-0.0021836571,-0.039963048,0.012418134,-0.0029162606,-0.009411323,0.021190008,0.021981474,-0.008958206,-0.044049915,-0.010395937,-0.023956511,-0.022787647,-0.005880367,-0.013224492,-0.021289276,-0.0044170273,-0.004886671,0.012505206,-0.018675411,0.011779449,-0.030687299,0.024905195,0.0171099,0.009200232,-0.015102101,0.021794675,-0.0116828615,0.0003090372,-0.010962044,-0.019123988,0.023270287,0.0060740164,-0.009270828,0.021037152,-0.017706957,0.03210536,-0.0006216772,0.0011124698,-0.0041178893,-0.0121660195,-0.088271625,-0.008999345,-0.01422731,-0.019051962,-0.0015941807,-0.007491298,0.007094648,-0.05252988,-0.023926694,-0.014981143,0.016956542,-0.005095833,-0.0332655,-0.036274042,-0.0012021054,-0.006783209,-0.032167982,0.043646686,0.012604345,-0.005206292,-0.005968315,0.015138369,0.01706405,-0.031222563,-0.018130755,-0.0016941312,-0.014492666,-0.011263299,-0.011760991,-0.012688344,-0.029154303,-0.12754297,0.008202326,-0.020694751,-0.030360242,0.0077234884,-0.015640575,-0.0038804223,-0.0038497234,0.020296218,0.0014100729,-0.00075401895,-0.03258383,0.008898148,-0.03216262,-0.014291251,0.11522401,-0.011066156,0.00070341444,-0.040589508,-0.04380707,0.030371211,-0.034253336,-0.014804196,-0.0020873595,0.0041300845,-0.008291379,0.019599559,-0.024303123,0.012535189,0.016399793,-0.002256816,0.00097074994,0.0011687032,-0.013903565,0.017665332,-0.014030218,0.0035286658,-0.021158006,-0.025746165,0.009073009,0.029417418,0.021972975,0.021330792,-0.0036601515,0.002370503,-0.0045803823,-0.0061165206,-0.029047018,0.01083082,0.011895097,0.012642597,-0.06314039,-0.013328432,-0.01832227,0.017944664,-0.009697085,-0.00088733324,0.022895146,0.0063914396,0.002816651,0.022190882,-0.017724209,-0.019703446,0.017215148,-0.013081038,-0.0013021078,0.01923319,0.040718906,0.020607762,0.019949269,0.0017452248,-0.006891242,-0.0075374595,-0.025064876,0.0064996285,-0.025877517,0.0033249161,0.030280085,0.0020565675,0.018992553,-0.002894991,0.015047784,0.00249435,0.0074074,-0.019700106,0.0034141839,0.014831023,0.0106211975,-0.02023529,-0.028036173,-0.019358229,-0.0015090234,-0.00521743,-0.002640208,0.030816017,0.015197485,-0.01924915,-0.012431519,-0.0036400596,0.0055742273,1.3287746e-05,-0.0096638,-0.0059073875,-0.021834396,-0.021219747,0.04048015,0.0011621043,0.018002331,0.0069196355,-0.008487234]",[37,39,41,42,44],{"name":16,"slug":38},"formal-calculus",{"name":13,"slug":40},"llm-agent",{"name":15,"slug":15},{"name":17,"slug":43},"conversation-semantics",{"name":14,"slug":45},"information-flow",{"id":27,"slug":47,"title":48,"language":49},"llmbda-calculus-agent-safety-rules-en","LLMbda calculus gives agents safety rules","en",[51,57,63,69,75,81],{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"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":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"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":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"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",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":26},"d75b5708-d4ec-4c46-9592-fa0a68d4bc26","judge-reliability-harness-stress-tests-llm-judges-zh","LLM 評審也會不穩","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778740856189-g1zr.png","2026-05-14T06:40:32.198872+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":26},"941f698a-1dcf-4807-bd56-5295c07d2dee","taming-black-box-llm-inference-scheduling-zh","黑箱 LLM 排程更聰明了","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778740253221-wgy6.png","2026-05-14T06:30:31.546746+00:00",{"id":82,"slug":83,"title":84,"cover_image":85,"image_url":85,"created_at":86,"category":26},"0eb3d74f-c737-41a4-8b9b-fc30b2b3b0ac","aisafetybenchexplorer-ai-safety-benchmarks-zh","AISafetyBenchExplorer：AI 安全基準地圖","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778739657024-fgxt.png","2026-05-14T06:20:28.07969+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"]