[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-entitybench-long-range-video-consistency-zh":3,"article-related-entitybench-long-range-video-consistency-zh":36,"series-research-bfd03801-a200-4222-9370-8b441be41483":88},{"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},"bfd03801-a200-4222-9370-8b441be41483","EntityBench 盯住長片一致性","\u003Cp data-speakable=\"summary\">EntityBench 用長篇多鏡頭影片測試模型能否跨鏡頭維持角色、物件與場景一致。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.15199\">EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation\u003C\u002Fa> 不是在追求單段影片有多漂亮，而是在問一個更實際的問題：當同一個角色、物件或地點隔了很多鏡頭再次出現時，生成模型還記不記得它是誰。對做敘事型影片、分鏡式內容或長流程影片系統的開發者來說，這件事不是加分項，而是基本門檻。\u003C\u002Fp>\u003Cp>這篇論文的核心主張很直接：現有影片評估，對「長距離一致性」這件事測得不夠。很多方法只看單獨生成的片段，或是用比較簡單的連貫性指標，結果很容易高估模型在長篇故事裡的實際能力。EntityBench 想補的，就是這個落差。\u003C\u002Fp>\u003Ch2>這篇在解什麼痛點\u003C\u002Fh2>\u003Cp>長篇、多鏡頭的影片生成，難點從來不只是畫面好看。真正麻煩的是，模型要在不同鏡頭之間維持同一個故事世界：角色長得要像同一個角色，物件要像同一個物件，場景回來時也不能整個走鐘。只要其中一項失真，觀眾就會感覺「這不是同一部片」。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778911845686-4mc8.png\" alt=\"EntityBench 盯住長片一致性\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>論文指出，現有評估方式常常沒有把這種問題逼出來。若 prompt 是彼此獨立生成的，就不一定會要求模型記住前面的實體；若評分方式太粗，就可能只看到影片流暢，卻沒看出角色身份早就漂掉。EntityBench 的目的，就是把這種失敗模式明確攤開。\u003C\u002Fp>\u003Cp>對開發者而言，這個差異很重要。測單一 prompt 的模型，和測能不能撐住一整段敘事流程，是兩種完全不同的工作。前者像 demo，後者才像產品。\u003C\u002Fp>\u003Ch2>EntityBench 怎麼設計\u003C\u002Fh2>\u003Cp>EntityBench 是從真實敘事媒體整理出來的 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa>，不是把一堆無關 prompt 湊在一起。資料規模包含 140 集與 2,491 個 shots，整體結構更接近故事內容，而不是隨機片段集合。它還使用明確的 per-shot entity schedule，追蹤角色、物件與場景在整段序列中的出現位置。\u003C\u002Fp>\u003Cp>這個設計讓「請記住這個實體」變成可評估的任務，而不是一句模糊要求。論文把 benchmark 分成 easy、medium、hard 三種層級，最長可到 50 個 shots；跨 shot 的角色最多 13 個，跨 shot 的場景最多 8 個，跨 shot 的物件最多 22 個。它也包含最長 48 個 shots 的 recurrence gap，也就是同一個實體兩次出現之間可以隔很遠。\u003C\u002Fp>\u003Cp>這點很關鍵，因為一致性問題通常不是在角色下一秒就壞掉，而是隔得越久越容易失憶。EntityBench 把這種長距離回歸的壓力，做成明確的測試條件。\u003C\u002Fp>\u003Ch2>評估不是只看一個分數\u003C\u002Fh2>\u003Cp>EntityBench 不只是資料集，還搭配一套三層評估框架。這套框架把不同問題拆開，不讓它們混成一團。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778911857671-d2zl.png\" alt=\"EntityBench 盯住長片一致性\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cul>\u003Cli>\u003Cstrong>Intra-shot quality\u003C\u002Fstrong>：看單一 shot 自己好不好。\u003C\u002Fli>\u003Cli>\u003Cstrong>Prompt-following alignment\u003C\u002Fstrong>：看模型有沒有照要求生成。\u003C\u002Fli>\u003Cli>\u003Cstrong>Cross-shot consistency\u003C\u002Fstrong>：看跨 shot 的實體能不能維持穩定。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>另外還有一個 fidelity gate。只有實際渲染正確的實體，才會進入 cross-shot scor\u003Ca href=\"\u002Fnews\u002Flovable-backs-atech-vibe-coding-hardware-zh\">ing\u003C\u002Fa>。這個設計很實際，因為它避免模型在第一個 shot 就畫錯人，卻還被算成「跨 shot 很一致」。\u003C\u002Fp>\u003Cp>對工程團隊來說，這樣的拆法很有價值。當結果不好時，你至少能知道問題是在畫面品質、prompt 對齊，還是長距離身份維持。沒有這種分層，除錯只會變成猜謎。\u003C\u002Fp>\u003Ch2>基線方法 EntityMem 做了什麼\u003C\u002Fh2>\u003Cp>為了展示 benchmark 的效果，作者提出一個基線系統 EntityMem。它的核心概念是記憶模組：在生成開始前，先把每個實體經過驗證的視覺參考存進持久記憶庫。這樣一來，模型不必每次角色或物件重現時都重新猜它長\u003Ca href=\"\u002Fnews\u002Faws-repository-wide-security-scanner-matters-zh\">什麼\u003C\u002Fa>樣。\u003C\u002Fp>\u003Cp>這個想法其實很務實。長篇影片的一致性，很多時候不只是生成問題，也是記憶問題。如果系統能取回可信的角色參考圖，後面再生成時就比較有機會保住同一個視覺身份。論文把 EntityMem 當作 baseline，而不是最終答案，但它清楚示範了「顯式記憶」可以怎麼幫助長篇影片生成。\u003C\u002Fp>\u003Cp>換句話說，這篇不是在說把模型再堆大一點就好，而是在暗示：你可能需要一個能記住故事世界的機制。\u003C\u002Fp>\u003Ch2>論文實際證明了什麼\u003C\u002Fh2>\u003Cp>作者最主要的發現，是現有方法在跨 shot 一致性上，會隨著 recurrence distance 增加而明顯退化。這就是 EntityBench 要抓出的核心失敗：兩次出現隔得越遠，模型越難維持同一個角色、物件或場景。\u003C\u002Fp>\u003Cp>論文也指出，在評估過的方法裡，顯式的 per-entity memory 帶來最高的 character fidelity 與 pr\u003Ca href=\"\u002Fnews\u002Fvibe-research-ai-tools-workflows-zh\">ese\u003C\u002Fa>nce。摘要中給出的角色 fidelity 效果量是 Cohen’s d = +2.33。除此之外，摘要沒有公開完整 benchmark 細節，所以沒有更多數字可以補充。\u003C\u002Fp>\u003Cp>這個結果值得注意，因為它暗示長距離一致性不只是「畫得更好」就能解決。它可能需要明確的實體儲存與重用機制。對做多鏡頭影片系統的人來說，這代表架構設計可能要往 memory-aware 的方向走，而不是只靠 prompt 逐段接力。\u003C\u002Fp>\u003Ch2>對開發者的實際影響\u003C\u002Fh2>\u003Cp>如果你在做影片生成、敘事剪輯、分鏡工具，或任何需要角色反覆出場的系統，EntityBench 提供的是更接近真實使用情境的壓力測試。它測的不只是「能不能生成影片」，而是「能不能在長篇故事裡保持世界觀一致」。\u003C\u002Fp>\u003Cp>它也讓失敗分析變得更清楚。模型可能因為單 shot 畫面品質差而失敗，也可能因為 prompt 沒跟上而失敗，還可能因為跨 shot 的身份記憶斷掉而失敗。EntityBench 的價值，就是把這三種問題分開看。\u003C\u002Fp>\u003Cp>這對產品團隊特別重要。因為當你要把生成模型放進實際工作流時，使用者在意的不只是某一幀好不好看，而是前後鏡頭能不能對得起來。角色錯位、物件換臉、場景漂移，這些都會直接破壞敘事可信度。\u003C\u002Fp>\u003Cp>不過，這篇摘要也有明確限制。它沒有把完整評估流程、所有指標細節，或 EntityMem 的泛化範圍講完整。摘要也沒有宣稱 memory 就是終極解法。它真正證明的是：長距離一致性確實會掉，而且顯式的實體記憶是一條值得走的方向。\u003C\u002Fp>\u003Cp>如果把這件事放到更大的脈絡來看，影片模型正在從短 clip 走向長故事。當長度拉高後，難題就不再只是「會不會動」，而是「能不能記得前面發生過什麼」。EntityBench 做的，就是把這個問題變成可以量測、可以比較、也可以繼續改進的 benchmark。\u003C\u002Fp>\u003Cp>對\u003Ca href=\"\u002Ftag\u002F台灣開發者\">台灣開發者\u003C\u002Fa>來說，這類研究的價值很實際。它提醒你，評估一個影片模型時，不能只看單段效果。只要產品有連續劇、教學流程、角色敘事、或任何需要重複出現實體的場景，就要把一致性當成核心指標，而不是事後補救。\u003C\u002Fp>\u003Cp>也因為如此，EntityBench 的意義不只是一個新 benchmark。它是在幫整個領域重新定義問題：不是生成一段好看的影片，而是生成一個記得住自己的故事世界。\u003C\u002Fp>","EntityBench 用長篇多鏡頭影片做一致性測試，檢查角色、物件與場景能不能跨鏡頭維持同一性，也提出帶記憶的基線方法 EntityMem。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.15199",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778911845686-4mc8.png",[13,14,15,16,17],"video generation","long-range consistency","multi-shot video","entity memory","benchmark","zh",2,false,"2026-05-16T06:10:27.85068+00:00","2026-05-16T06:10:27.743+00:00","done","f275c9a6-0b6e-434f-bb3d-a8611716c3d0","entitybench-long-range-video-consistency-zh","research","d60602fc-ed44-4c5e-8aa1-b0285672b8ba","published","2026-05-16T09:00:17.027+00:00",[31,32,33],"EntityBench 把長篇多鏡頭影片的一致性問題做成可評估 benchmark。","它用實體排程、recurrence gap 與三層評估，把畫面品質、prompt 對齊和跨鏡頭一致性拆開。","摘要顯示一致性會隨距離增加而退化，而顯式 per-entity memory 是有前景的方向。","0c35a120-52fc-41fc-afa3-d404eb934158","[-0.039884124,0.023197986,-0.0020930043,-0.08984849,-0.016147563,0.00026019086,-0.03544639,-0.00034575313,0.03573381,-0.014472459,0.0030923171,-0.01629331,0.010799906,0.017350947,0.10640167,0.029744266,-0.026174283,0.0026378853,0.00629314,-0.023068743,0.0009292847,0.0014888081,-0.00458124,-0.025637392,-0.014030611,-0.00054362783,0.03375765,0.008784846,0.035061892,-0.003116321,0.014109251,0.022156669,0.022845874,0.020318279,0.01341501,-0.00012955442,0.007063792,0.01259447,0.013708802,-0.009927758,-0.017058302,-0.01968722,0.01987458,0.0019431893,-0.023754766,0.0042477585,0.002242896,-0.028318318,0.02591927,-0.017468829,-0.01961358,0.011500423,-0.023919493,-0.15287958,0.011415822,-0.007898347,-0.0063366955,0.011185243,-0.018705353,0.015500062,-0.009765813,-0.0054769795,-0.023586871,-0.040070638,0.0015054106,-0.02353989,-0.0046351,-0.003785236,-0.014476449,-0.009857652,-0.028368937,-0.005104926,0.019662194,-0.018973,0.009260415,-0.01068464,0.0026820728,-0.00042540225,0.009666878,0.018164046,0.002853813,-0.004691147,0.015238936,-0.014095935,-0.0090309,-0.0032680368,-0.00405881,0.01602624,-0.013602636,0.0049474235,-0.0050023445,0.015174989,0.0019176976,-0.017661735,0.013571794,-0.015763707,-0.019743867,-0.008218576,0.00038053634,0.03135908,0.001464945,-0.011071477,0.03621367,0.035095204,0.026249688,-0.033724904,-0.000763895,-0.0039156363,0.0024440947,0.0029069467,0.011202143,-0.015772706,-0.0028413057,-0.00073719886,0.0066590994,-0.15159398,-0.010657265,0.0031835528,-0.005207835,0.011054206,-0.01941734,0.0042438745,-0.0019928499,0.03148948,0.021460844,-0.005323926,0.0195013,-0.001453535,-0.020500662,-0.0019171394,-0.0134449415,0.023219168,0.004319476,0.01464934,-0.015196632,0.023133896,-0.006714567,-0.03443305,-0.00012038811,0.009433467,-0.006284282,0.021334428,0.010217194,-0.0056829844,-0.023250096,0.0075585977,-0.0390448,0.020308515,0.009606619,-0.014777258,0.018702017,-0.0055561746,-0.00054904487,-0.017069327,0.051380288,-0.033395108,-0.016536718,-0.0059206486,-0.011722773,0.007835525,0.020844469,-0.00515695,-0.027161881,0.021336213,0.0062409583,0.038073543,0.010973233,-0.01416874,0.018335652,0.010127261,-0.012551598,-0.007822822,-0.0032900616,-0.011105646,0.020109503,-0.011448935,-0.009561177,-0.009210026,-0.012277835,0.02697335,0.0014297072,0.003769465,-0.018971603,-0.00776095,0.025668066,0.014364966,-0.018891755,0.029475119,0.02715569,0.025650855,-0.019537235,-0.010013596,0.015380399,-0.028610291,-0.026111204,-0.022330647,-0.00680715,0.0074920086,-0.002629731,0.041655406,-0.0011937817,0.0014723282,0.019960364,0.021925986,0.0071768076,0.014089878,-0.0067906277,-0.0034381815,0.0023751569,0.007724419,-0.009427103,0.012969505,0.016077517,-0.013161476,-0.009660397,-0.022382842,-0.031452876,-0.00386413,-0.0070883594,-0.026531596,0.018962147,-0.00775948,0.020871311,-0.004367313,-0.0032568555,-0.05163108,-0.011502356,-0.0026400923,-0.016732706,0.010308113,-0.013340869,0.025031837,-0.015820824,0.007049389,0.041769646,-0.020141626,-0.024768163,-0.0018863425,0.013988799,0.01550511,-0.017744713,0.020610131,-0.0069530997,0.010638473,0.00897718,0.01701086,-0.0020992053,0.0006095421,0.0169238,0.018616939,-4.4394208e-05,0.008406207,-0.009521781,-0.0068322727,0.011754631,-0.022979911,-0.011151304,0.0097322455,0.0040087732,-0.020986952,-0.0049638585,0.015715187,-0.01715316,-0.006790284,0.02618224,0.009845961,0.02915498,0.0051356093,-0.029012226,0.0044478574,-0.0138097275,0.015095973,0.012326437,0.002445031,-0.013628138,-0.014649207,-0.052720867,-0.0039683837,0.0102478815,-0.006636355,0.009423269,-0.005633939,-0.01136115,0.017608186,-0.018807491,-0.014493034,0.008191508,-0.015283426,-0.012108788,-0.013236721,0.007247037,0.0040800595,0.012031858,-0.005348311,-0.001669088,-0.02746299,-0.006070154,0.003246665,-0.0028789402,-0.0033354068,0.036793645,-0.0045690383,-0.022828657,0.053742778,-0.0024569447,-0.037034832,-0.021071387,0.012834974,0.025430616,-0.025143867,-0.01548245,-0.009850634,-0.027049191,-0.023891604,0.011144938,0.0058235005,-0.0145834815,-0.011380532,-0.014028512,0.0020786137,0.013842755,-0.009434659,-0.024057234,0.011051387,-0.02322095,0.011001566,0.008647833,-0.003361662,0.0063227564,-0.029200139,-0.0043968544,0.02493572,0.002199776,-0.027474744,0.002866789,0.0023274939,-0.0018472942,0.0029693784,0.014065217,0.0006213036,0.0127340155,-0.008548841,-0.001475863,0.022843936,-0.008550001,0.036862537,0.012062794,-0.013741549,-0.006030133,-0.02052458,0.005366979,0.00542844,-0.0003624603,-0.03099061,-0.010672984,0.0069807908,-0.019124134,0.007982886,0.008514782,0.023943348,-0.00790902,-0.0034727638,0.008869289,0.00076608243,0.033007998,-0.01672246,0.010002095,0.00074838754,-0.027718382,0.0011423368,0.012551949,-0.014648492,-0.0028944504,0.013956833,-0.017743792,0.0038952152,-0.015311457,-0.041288234,-0.018438095,-0.015463254,0.012436638,0.040479146,-0.046285484,-0.011016546,-0.022363478,0.01975084,0.011282332,-0.0060630017,-0.009186133,0.031021025,0.023730278,-0.015336608,-0.035223667,-0.0030454274,-0.01839312,0.0079339715,-0.004607839,0.008032196,-0.013733962,0.003255138,-0.013683859,0.0016179957,-0.009012744,-0.03447779,-0.006027832,0.015338412,-0.022914067,0.008823421,0.0011334439,-0.024089657,-0.019108512,-0.015979309,0.016017497,-0.0012163696,0.023795472,0.021234522,-0.00253053,-0.009769926,-0.015960218,0.009873365,-0.014805062,-0.007073904,0.0014578849,-0.0024852667,-0.013270179,0.013783603,-0.014946119,-0.015502633,0.012970898,-0.06147117,0.01756831,0.003903116,-0.012742872,-0.037994467,-0.017375717,-0.02174617,-0.039339196,0.014252409,0.0038789455,-0.0068394816,0.0017376955,-0.011784291,-0.008203243,0.0011975712,0.024912251,-0.023755575,0.0066688876,0.0015533185,-0.03350673,0.017037464,0.026361784,0.03639879,0.011022126,-0.001565891,-0.024243547,-0.006411087,-0.031431798,-0.015864694,-0.007848296,-0.039008442,-0.0030693305,-0.010312036,-0.002945341,0.018002724,-0.008508332,0.0054182825,-0.00546532,-0.016427252,-0.025659492,0.014982296,-0.0025004037,0.0077158203,0.0033145451,0.008841935,0.018294277,0.009675915,-0.021336207,0.0012195722,0.012182147,-0.0031557982,-0.030440008,-0.012604863,0.025347617,-0.0014586622,0.014847442,0.019928053,0.03188866,0.015670061,-0.021383617,-0.006429787,0.009953437,0.034658562,0.013938422,0.008277759,-0.026247198,-0.014330766,-0.016876265,-0.0046461355,0.023291048,-0.0035271293,0.020936262,0.0026419272,0.019419923,-0.02042773,-0.027599603,-0.018404253,0.0033571906,0.0059901034,0.01816404,-0.017266564,-0.007055058,-0.019727677,0.022664364,0.036859505,0.0008309389,0.00519017,-0.0047164266,0.041144263,0.024014382,0.009438388,-0.016952025,-0.009628173,-0.021628765,0.013512652,-0.020513166,-0.01136556,0.020009894,0.0019518437,0.021587819,0.046057917,-0.007965793,0.012729824,0.010968846,-0.019234471,0.0027827984,-0.012412254,0.035408586,0.0031464377,0.0015936793,0.0047364854,-0.0055131842,0.0033668543,-0.014040032,-0.0105920555,0.028964637,-0.08663588,0.0008007274,0.022180632,-0.02307762,-0.00727285,-0.026180694,-0.00972713,-0.02515726,0.032287207,0.002674116,0.017766247,-0.011230131,-0.012341416,0.023592774,0.015249222,-0.011881543,-0.016301354,-0.01920916,0.0059068897,-0.015470719,0.027574299,0.0066912253,0.045227736,0.030039772,0.0032211393,-0.01924481,0.017595006,0.033971723,0.032628093,0.0037950808,-0.035452843,-0.009388989,-0.0014512672,0.02372112,0.0114580905,-0.007400012,-0.0004104686,0.0029174974,-0.012692971,-0.006219544,-0.0058488077,0.011494679,-0.037382144,-0.039009675,0.014689295,0.013218466,-0.0059583453,0.0066357558,-0.012469034,0.005213372,-0.011603291,-0.008675055,-0.008867484,-0.017009296,0.008577136,-0.015724713,-0.03207906,0.00790404,-0.005848717,-0.00059529464,-0.0026422408,0.003100858,0.016724475,0.03819353,-0.008078068,0.01940405,-0.013590639,0.025649818,-0.001952943,0.01572233,-0.00029483097,0.0163174,0.016366642,0.027408179,-0.0066393553,0.01253244,-0.0032437055,0.027423784,-0.012141413,-0.004876043,-0.018123064,-0.021555258,-0.08210772,-0.01852882,-0.008625134,0.029884977,0.032365095,-0.0032164247,0.014341458,-0.008520944,0.005439171,-0.0045840424,0.013888752,0.0140535645,-0.009547302,0.001178665,-0.008315367,-0.017406268,-0.0133294985,-0.010127909,-0.013554521,-0.052605104,-0.02597911,-0.014560047,0.020787248,-0.014774176,-0.019020304,0.014603562,-0.011152014,0.0034132444,0.010426948,-0.021839017,0.003416631,-0.11571484,-0.0005269248,0.0149455415,0.005254583,0.016711533,0.009602089,0.010433703,-0.00024835835,0.003404761,0.016667884,0.012829482,-0.033942778,-0.009291617,-0.0028776757,0.0067113247,0.11312072,0.008502183,-0.01699998,-0.025305362,-0.03764103,-0.0069097253,-0.038628567,-0.03359384,0.0051599126,-0.0035248422,0.015345048,0.0240064,-0.02858147,-0.03671078,0.01920449,-0.004091828,0.015180882,-0.026443174,-0.023766397,0.0018833095,0.0038471415,-0.0008168557,-0.01953669,0.021943133,-0.008197437,0.020830307,0.0067134914,0.007665937,-0.0063874223,0.0027994874,-0.014505497,-0.010065932,-0.019681377,-0.0026222202,-0.0025394456,-0.016285786,-0.06521413,-0.004548626,-0.015328132,0.00016853641,-0.0054856716,-0.037127268,-0.007674369,0.019688409,0.018721608,0.003994932,0.007219698,-0.00032003823,0.028968943,-0.020379297,-0.032964557,-0.015450089,0.026199548,-0.004218622,-0.005143567,-0.003446388,0.0030317546,-0.018501913,-0.015105977,-0.00035711823,0.010407082,-0.009226857,-0.015044611,0.012535549,0.012307648,0.014072301,0.0051611024,0.0036873291,-0.000546834,-0.0031245032,-0.024566833,-0.006675206,0.0068201534,0.004826641,-0.043907076,0.034154993,0.008176575,-0.0062968545,-0.011475233,-0.004729145,0.023550648,0.0046808375,0.0010378564,0.01115493,0.0067780395,-0.01357405,-0.008930355,0.011261797,-0.016170619,0.037236147,0.01734469,-0.0010694641,0.007881659,0.03680827,0.026223265]",{"tags":37,"relatedLang":47,"relatedPosts":51},[38,40,41,43,45],{"name":15,"slug":39},"multi-shot-video",{"name":17,"slug":17},{"name":16,"slug":42},"entity-memory",{"name":13,"slug":44},"video-generation",{"name":14,"slug":46},"long-range-consistency",{"id":27,"slug":48,"title":49,"language":50},"entitybench-long-range-video-consistency-en","EntityBench Tackles Long-Range Video Consistency","en",[52,58,64,70,76,82],{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":26},"6ca303f0-7bd4-4bb2-be58-70d80da5ec40","why-ai-safety-teams-are-wrong-blame-only-alignment-zh","為什麼 AI 安全團隊錯把問題全怪在對齊","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778947417022-ak55.png","2026-05-16T16:03:16.319335+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":26},"50b2e74e-7248-43a3-8775-451bf2569f33","why-fine-tuning-llms-domain-tasks-right-default-zh","為什麼針對領域任務微調 LLM 才是預設選項","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778916229431-9olk.png","2026-05-16T07:23:32.255569+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":26},"001e062e-f246-4bf0-aa04-27506febcf7b","refdecoder-reference-conditioned-video-decoder-zh","RefDecoder 讓影片解碼器吃參考圖","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912646805-czy9.png","2026-05-16T06:23:33.170076+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":26},"b9516feb-41d5-42a3-887e-7b47c5c9ffb7","atlas-one-token-visual-reasoning-zh","ATLAS 用一個 token 做視覺推理","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912032775-hp0w.png","2026-05-16T06:13:34.693651+00:00",{"id":77,"slug":78,"title":79,"cover_image":80,"image_url":80,"created_at":81,"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":83,"slug":84,"title":85,"cover_image":86,"image_url":86,"created_at":87,"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",[89,94,99,104,109,114,119,124,129,134],{"id":90,"slug":91,"title":92,"created_at":93},"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":95,"slug":96,"title":97,"created_at":98},"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":100,"slug":101,"title":102,"created_at":103},"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":105,"slug":106,"title":107,"created_at":108},"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":110,"slug":111,"title":112,"created_at":113},"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":115,"slug":116,"title":117,"created_at":118},"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":120,"slug":121,"title":122,"created_at":123},"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":125,"slug":126,"title":127,"created_at":128},"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":130,"slug":131,"title":132,"created_at":133},"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":135,"slug":136,"title":137,"created_at":138},"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"]