[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-atlas-one-token-visual-reasoning-zh":3,"article-related-atlas-one-token-visual-reasoning-zh":37,"series-research-b9516feb-41d5-42a3-887e-7b47c5c9ffb7":90},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":19,"translated_content":10,"views":20,"is_premium":21,"created_at":22,"updated_at":22,"cover_image":11,"published_at":23,"rewrite_status":24,"rewrite_error":10,"rewritten_from_id":25,"slug":26,"category":27,"related_article_id":28,"status":29,"google_indexed_at":30,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":31,"topic_cluster_id":35,"embedding":36,"is_canonical_seed":21},"b9516feb-41d5-42a3-887e-7b47c5c9ffb7","ATLAS 用一個 token 做視覺推理","\u003Cp data-speakable=\"summary\">ATLAS 用一個離散 \u003Ca href=\"\u002Ftag\u002Ftoken\">token\u003C\u002Fa>，把 agentic 和 latent 視覺推理合在一起，想減少中間步驟成本。\u003C\u002Fp>\u003Cp>\u003Ca href=\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.15198\">ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both\u003C\u002Fa> 盯上的，是多模態模型裡一個很實際的痛點：視覺推理一旦牽涉到中間狀態，系統就容易變得又重又難訓練。你可能要讓模型吐出中間影像、呼叫外部工具，或在隱藏空間裡維持一串 latent 狀態。這篇論文的主張很直接：不一定要為每種推理方式都做一套機制，或許一個 token 就夠了。\u003C\u002Fp>\u003Cp>這個想法的重點，不是把模型變得更會講話，而是把視覺推理壓縮成更乾淨的介面。對開發者來說，這種設計如果真的站得住腳，代表模型可以少吐很多冗長的中間內容，也不用為了視覺推理另外拉一條複雜管線。更重要的是，它希望維持標準的 next-token prediction 形式，不必為了新的推理模式改掉整個訓練框架。\u003C\u002Fp>\u003Ch2>這篇論文想解什麼問題\u003C\u002Fh2>\u003Cp>論文先把視覺推理拆成兩條常見路線。第一條是 agentic reasoning，也就是透過程式碼或工具呼叫來做推理。第二條是 latent reasoning，讓模型在可學習的隱藏嵌入裡處理視覺資訊。兩種做法都能做事，但代價不一樣，而且各自有明顯缺點。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912032775-hp0w.png\" alt=\"ATLAS 用一個 token 做視覺推理\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>agentic 路線的問題，主要是 context-switching latency。因為它要依賴外部執行，所以模型和工具之間來回切換會拖慢速度。latent 路線則避開了外部執行的開銷，但論文指出它常常缺乏任務泛化能力，而且在 autoregressive parallelization 的訓練方式下不太好搞。至於直接在統一模型裡生成影像，論文也把它描述成計算成本高、架構上也不簡單。\u003C\u002Fp>\u003Cp>ATLAS 要處理的，就是這種「每條路都能走，但每條路都不夠順」的狀況。它不是再加一個新的推理模組，而是把一個離散詞元當成共同單位，讓同一個 token 同時扮演 agentic 操作與 latent 視覺推理單元。這個方向很像是在問：如果最後都要進到 token 流裡，能不能乾脆把推理也收斂到 token 層級？\u003C\u002Fp>\u003Ch2>ATLAS 的方法到底怎麼運作\u003C\u002Fh2>\u003Cp>ATLAS 的核心概念叫做 functional token。你可以把它想成 tokenizer 字典裡的一個標準 token，但它不只是文字單位，還內含一個視覺操作。論文的說法是，這個 token 可以同時對應 agentic operation 與 latent visual reasoning unit，也就是說，一個 token 可以同時承擔「動作」和「隱藏式視覺運算」兩種角色。\u003C\u002Fp>\u003Cp>這裡最關鍵的地方，是它仍然走一般的 next-token prediction。換句話說，模型不需要一條特殊的解碼路徑，也不需要為了這種推理方式另外設計一個全新的架構。論文強調，這讓它能和 vanilla 的 scal\u003Ca href=\"\u002Fnews\u002Flovable-backs-atech-vibe-coding-hardware-zh\">able\u003C\u002Fa> supervised fine-tuning，以及 \u003Ca href=\"\u002Ftag\u002Freinforcement-learning\">reinforcement learning\u003C\u002Fa> 相容，而且不用做架構或方法上的修改。這點對工程實作很重要，因為很多多模態新方法最後卡住的不是效果，而是怎麼塞進既有訓練流程。\u003C\u002Fp>\u003Cp>另一個值得注意的地方，是這些 functional tokens 不需要 visual supervision。這代表訓練資料不一定要有額外的視覺標註來教模型這個 token 該做什麼。從多模態資料製作的角度看，這是很吸引人的設計，因為視覺標註通常比純文字標註更麻煩，也更貴。ATLAS 想做的，就是讓離散 token 成為視覺推理的共同接口，減少對複雜中間表示的依賴。\u003C\u002Fp>\u003Cp>如果用白話講，ATLAS 的策略不是讓模型「多想一點」，而是讓模型「用更少的東西想」。它想避免生成冗長的中間視覺內容，也不想把推理拆成太多外部步驟。這種設計的野心很明確：保留結構化推理能力，但把推理成本壓到 token 層級。\u003C\u002Fp>\u003Ch2>論文實際證明了什麼\u003C\u002Fh2>\u003Cp>從摘要能看到的結果是，作者做了 extensive experiments and analyses，並聲稱 ATLAS 在具挑戰性的 b\u003Ca href=\"\u002Fnews\u002Fentitybench-long-range-video-consistency-zh\">ench\u003C\u002Fa>mark 上拿到更好的表現，同時還保有清楚的 interpretability。不過，摘要沒有公開完整 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 細節，也沒有列出具體數字、測試集名稱或比較對象，所以目前只能確認它有正向結果，不能從這份 raw 資料推導出更細的排名或分數。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912032687-tf6t.png\" alt=\"ATLAS 用一個 token 做視覺推理\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>摘要裡另一個很重要的資訊，是作者也碰到了訓練上的真問題：functional tokens 在 reinforcement learning 中是 sparse 的。這種稀疏性會讓訓練變難，因為 token 出現頻率低，梯度訊號就容易不穩。這不是理論上的小瑕疵，而是足以影響方法能不能訓練起來的實務問題。\u003C\u002Fp>\u003Cp>為了解這個問題，論文提出 Latent-Anchored GRPO，簡稱 LA-GRPO。它的作用是用一個 statically weighted auxiliary objective 來錨定 functional tokens，讓這些特殊 token 在 RL 裡能得到更穩定的更新。白話一點說，就是當主要獎勵訊號太稀疏、太飄的時候，作者額外加了一個固定權重的輔助目標，幫這些 token 抓住訓練方向。\u003C\u002Fp>\u003Cp>這個設計透露出一件事：ATLAS 不是那種「丟一個漂亮概念就結束」的方法。它在 repr\u003Ca href=\"\u002Fnews\u002Fvibe-research-ai-tools-workflows-zh\">ese\u003C\u002Fa>ntation 上很乾淨，但在 optimization 上仍然有明顯難度。也就是說，單一 token 的想法雖然簡潔，卻還是得靠額外訓練技巧把它扶起來。對研究者或工程師來說，這通常意味著方法的優點和成本是綁在一起的。\u003C\u002Fp>\u003Cp>論文也一直強調 interpretability。因為 functional token 本身還是 token，所以推理過程不會完全沉到黑盒 latent 裡。這不代表它就能被人類完全看懂，但至少比起純 hidden pipeline，表面上更容易檢查。對需要做模型分析、debug 或安全審視的人來說，這是很實際的差異。\u003C\u002Fp>\u003Ch2>對開發者有什麼影響\u003C\u002Fh2>\u003Cp>如果你在做多模態系統，ATLAS 最有意思的地方，是它試圖把高階視覺推理塞回一般語言模型的工作流裡。這代表它想降低架構特殊化的程度，讓模型訓練還是沿用標準的 SFT 和 RL，而不是每次都為了視覺推理重寫整套 pipeline。對團隊來說，這種兼容性通常比單次 demo 的華麗更重要。\u003C\u002Fp>\u003Cp>從實作角度看，這篇論文至少帶來幾個值得注意的方向：\u003C\u002Fp>\u003Cul>\u003Cli>它可能減少中間推理內容的冗長度，讓輸出更精簡。\u003C\u002Fli>\u003Cli>它把推理維持在 token stream 裡，較容易接到既有 LLM 工具鏈。\u003C\u002Fli>\u003Cli>它試著避開外部工具執行帶來的 latency。\u003C\u002Fli>\u003Cli>它提供一種不靠 visual supervision 的訓練思路。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>但這不代表問題都解了。摘要自己就點出，functional tokens 在 RL 中很稀疏，還需要 LA-GRPO 這類輔助目標來穩定訓練。這表示方法在概念上很俐落，實際上卻可能對訓練設定相當敏感。尤其當資料分布、任務型態或獎勵設計變動時，這種稀疏 token 的穩定性很可能是關鍵風險。\u003C\u002Fp>\u003Cp>另一個限制，是摘要沒有提供完整 benchmark 細節。雖然作者說有更好的表現，但我們從這份 raw 資料裡看不到哪些任務最受益、差距多大、或在哪些情境下表現會掉下來。對開發者來說，這意味著你現在還不能只憑摘要就判斷它是否適合自己的產品場景。\u003C\u002Fp>\u003Ch2>這篇研究真正值得注意的點\u003C\u002Fh2>\u003Cp>ATLAS 的價值，不只是「用一個 token 做很多事」，而是它把視覺推理重新定義成 token 設計問題。這個角度很有意思，因為它把 agentic 路線和 latent 路線拉到同一個框架下比較，最後用單一離散單位把兩者合併。對關心模型介面設計的人來說，這是一種很乾淨的抽象。\u003C\u002Fp>\u003Cp>不過，乾淨不等於容易落地。摘要已經明講，稀疏性會讓 RL 訓練變難，所以作者才需要 LA-GRPO。這也提醒我們，很多多模態方法的真正瓶頸不是推理概念，而是怎麼讓它在訓練中穩定出現、穩定學到、穩定泛化。ATLAS 在這裡提供了一個方向，但還不是終局答案。\u003C\u002Fp>\u003Cp>如果把這篇論文濃縮成一句話，就是：它想讓視覺推理像普通語言建模一樣運作，卻又保留足夠的結構與可解釋性。這個方向對未來的多模態模型很有吸引力，特別是當大家都在想怎麼降低推理成本、維持訓練相容性、又不要把系統做得太複雜時。\u003C\u002Fp>\u003Cp>目前能確定的只有這些：ATLAS 提出單一 functional token 來統一 agentic 與 latent 視覺推理，使用 next-token prediction 保持訓練相容，並用 LA-GRPO 處理 RL 下的稀疏性問題。至於它在更完整的實際場景裡能走多遠，還得看完整論文的實驗細節。\u003C\u002Fp>","ATLAS 提出用單一離散 token 同時承擔 agentic 與 latent 視覺推理，想降低中間步驟成本，並維持標準 next-token 訓練流程。","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.15198",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778912032775-hp0w.png",[13,14,15,16,17,18],"visual reasoning","functional token","latent reasoning","agentic reasoning","reinforcement learning","GRPO","zh",2,false,"2026-05-16T06:13:34.693651+00:00","2026-05-16T06:13:34.463+00:00","done","b0a4bc50-6036-469b-b918-423fd7fab01d","atlas-one-token-visual-reasoning-zh","research","2a05602e-4f77-4e7a-a073-0f3878a9d9de","published","2026-05-16T09:00:16.869+00:00",[32,33,34],"ATLAS 用單一離散 token 統一 agentic 與 latent 視覺推理。","方法維持標準 next-token prediction，並宣稱可相容 SFT 與 RL。","作者另外提出 LA-GRPO，處理 functional tokens 在 RL 中稀疏、難訓練的問題。","0c35a120-52fc-41fc-afa3-d404eb934158","[-0.0028132657,0.025634341,0.023025308,-0.08588443,-0.022651564,-0.024947383,-0.028358374,-0.005859305,0.012860417,-0.004153979,-0.024811475,-0.014771749,0.029219538,-0.00785462,0.13549727,0.018391812,-0.029396905,0.013810177,0.0066530574,-0.039808523,0.014633166,0.0131948255,-0.0077862996,-0.008982486,-0.016765963,0.022360258,0.022080388,-0.011985048,0.04182974,0.0158605,-0.028093241,0.009264325,0.007963533,0.021774633,-0.006108776,-0.012123416,0.01431926,-0.004135032,0.023946634,0.015105068,-0.01758952,-0.019921947,0.01191281,0.0047166254,-0.0006186413,-0.003412518,-0.007621857,-0.043856807,-0.008007644,0.010411326,-0.003235516,0.06010345,-0.011341452,-0.15103866,-0.008274127,-0.004109397,-0.008303507,0.0122863185,0.017507635,0.026203446,0.0031362516,-0.012368376,-0.0025559105,0.00039237036,0.008589835,-0.013996988,0.027930176,0.003973121,-0.03571361,-0.010233548,-0.015935808,-0.014398546,0.012376413,-0.011483639,0.030623619,-0.02658585,0.0049113003,0.016664479,0.014049367,0.007524508,0.0083079785,-0.016761007,-0.006695416,4.1346797e-05,0.0124079995,0.0032554418,0.005933875,0.0031378798,-0.0056122327,0.018244406,-0.017961584,0.019251864,0.033369325,-0.007249951,0.00015100557,-0.019292884,-0.023904018,-0.0016242263,0.0041272966,0.02958217,-0.04202865,-0.02025218,0.015410018,0.01856593,0.01003397,-0.014445111,-0.0058434266,0.005595464,-0.008962404,0.028880218,0.0063316766,-0.0006910624,0.004836452,-0.011419036,-0.015656779,-0.122282304,0.0057996465,-0.0060280315,0.006323499,0.019485397,-0.023018673,-0.014249531,0.007312487,0.008848712,-0.0030864724,0.014720272,0.015622208,0.0028738128,-0.034977585,-0.01466046,-0.01619098,0.015367721,-0.009904068,0.014015607,-0.004284687,0.019214327,-0.005848193,0.0014208137,-0.030374302,-0.012138201,-0.0153865665,0.037233118,0.0099985385,0.013287714,-0.02939998,-0.0061906674,-0.044701044,0.012738449,0.014492257,-0.027305346,0.01618839,-0.012218991,-0.02128575,0.012326299,0.021341788,-0.028388586,-0.00026893397,0.018669507,0.02289961,0.024224367,-0.0034684695,-0.02454918,0.0008492974,0.008739147,-0.015838012,0.012502906,0.0037588903,-0.010144655,0.019594008,0.021253202,-0.03656502,-0.015375283,-0.022285998,0.004645057,0.023725532,-0.0064158975,-0.004908083,-0.006179383,-0.0046185236,-0.00236256,0.013322979,-0.022804815,0.011216781,-0.005308573,0.010890835,0.012757592,0.001093921,0.030528657,0.013158747,0.005945357,-0.033072583,-0.002977722,0.005860264,-0.004418312,0.0031278532,0.001919105,-0.010335878,0.00087899494,0.008258529,0.024506131,-9.627115e-05,0.001307956,0.041391425,-0.014868194,-0.009451896,-0.026028479,-0.003148701,-0.009167608,-0.011244597,0.010264936,0.012746779,0.009820307,0.01857965,-0.028338129,0.019670494,0.004294071,-0.030297095,-0.0011836669,-0.008955031,-0.025802163,0.038812287,-0.00040153903,0.0050204736,3.1359407e-06,-0.0037172793,-0.010939666,-0.011999815,-0.024189051,-0.011796918,0.00619589,-0.011548636,0.022520393,0.026985738,-0.018446026,0.023900045,-0.0035866944,-0.00035186324,0.0011363166,0.00076747034,0.019634683,0.0007920293,0.009474084,-0.00019695066,0.001818096,0.034705576,-0.011819011,0.020086927,0.0077950824,-0.0013199378,0.0115849385,-0.009364325,-0.027974885,-9.694812e-05,0.003375429,-0.0062725274,0.015388236,-0.014767848,0.006374169,0.005842932,0.00322045,-0.0005775124,-0.011078648,-0.0058491644,-0.019487517,-0.009448444,0.017351242,0.008553044,0.006103586,-0.012679939,-0.00930176,-0.0023203392,0.016705444,-0.004216442,0.009434169,-0.0026515145,0.010500218,-0.048189323,0.046621237,-0.010753966,-0.041147884,-0.009752778,-0.0074950634,-0.00082329375,-0.002095722,0.009742012,0.013867774,0.0068286294,-0.0041994276,-0.0064385147,0.0077870116,0.003198379,0.049039718,-0.0056250114,-0.005766154,0.008721936,-0.007088935,-0.009426331,0.0017584094,-0.0055572446,0.015827136,0.016817465,-0.013915241,0.00026889908,0.020494817,0.006273251,-0.005642561,0.011201243,0.021565901,0.0025210194,-0.0013289837,-0.03754833,-0.037285343,0.013276333,-0.0067569166,0.008171882,-0.011543336,0.002372081,0.011248842,-0.0019196898,0.009736177,0.0046838005,-0.0141554875,-0.04152903,-0.022726156,-0.020151917,0.012515557,0.008017971,0.0022002123,0.009736232,-0.010553282,-0.005460879,0.014106467,0.055207595,-0.05458465,0.0077021723,0.010286028,-0.0269373,-0.019420912,-0.014877942,0.017570611,-0.0019703924,0.012463584,-0.004691549,-0.014565087,-0.02648605,0.02436369,0.027548444,-0.00079666375,0.0058909096,-0.02796723,0.02788206,0.0009715188,-0.029052839,-0.032058712,-0.028611377,0.016446691,-0.0070212116,0.0032124515,0.012115158,0.00013108394,0.009171288,-5.470684e-05,-0.035443358,-0.0014929969,0.01096163,-0.01944165,0.026259523,0.014294374,-0.01010151,0.0038695582,0.005761043,0.011168539,-0.0075429557,-0.016594045,-0.024304718,-0.0049606925,0.00026913648,-0.008330229,-0.023287453,0.028028117,-0.005945065,0.022797586,-0.020555757,-0.012340784,-0.025271965,0.0058474066,0.009409862,0.005412088,0.00094743696,0.01873626,-0.008322672,0.008236933,-0.004547874,-0.017934,0.0021782496,0.013287434,0.002739349,0.004782921,-0.012515113,0.011748923,0.016813988,0.00029442232,-0.007103438,-0.012411729,-0.0074512004,0.003484978,0.00901566,0.024689179,-0.011425561,-0.018423662,-0.0032951364,-0.0039775446,0.011714966,-0.015506753,0.037615173,0.012396791,-0.02136601,-0.0060517895,0.016356813,-0.01574114,0.007154279,-0.006653254,-0.01396267,0.015729534,-0.01047175,0.012478494,-0.008422675,-0.004184456,-0.014153456,-0.036543954,-0.019077076,-0.0062782476,-0.005349138,-0.003603117,0.01301328,-0.0029448592,-0.022135949,0.004846726,-0.0027045782,-0.027811794,-0.0065452494,-0.025733849,0.0058975113,0.023208713,0.0008080206,0.0039749406,0.002454038,-0.016044702,-0.027064746,0.0080430405,0.0334356,0.01753788,-0.009885994,-0.00822553,-0.019255359,-0.025053076,-0.0022624524,0.017098324,-0.018702578,-0.013766583,-0.0026062888,0.00865205,0.010588617,0.02249128,0.026631448,-0.029883008,-0.007959243,-0.019644633,-0.024501698,0.015531765,-0.007472464,0.01811088,0.03606552,0.017036939,-0.010015988,0.00048999005,-0.022550002,0.007144218,0.013250058,0.0066509484,-0.0028519062,-0.026987154,0.006750117,-0.022175387,-0.019931644,0.017645238,-0.020487914,0.0102788275,-0.00890771,-0.0007327746,0.010156333,0.023451228,0.006672991,0.0027449117,-0.0029229256,-0.0032031774,-0.010105511,-0.0074822525,0.015137948,-0.010909521,0.009313401,-0.012276804,0.011716135,-0.0040016496,-0.0037175512,0.0045240223,-0.02925843,0.009178388,0.018730927,-0.036116224,0.00074398745,-0.0038502342,-0.009481671,0.042600714,0.007950313,-0.022292726,-0.0143663315,0.02296754,0.006857107,0.006788054,-0.0015504664,0.0036197999,0.007900715,0.00918235,-0.014821307,0.00843933,0.011454817,-0.010740379,0.036859132,0.023706274,-0.03527201,0.034890465,0.04300897,-0.023037763,-0.0018522129,-0.004705581,0.011443164,0.032735065,-0.010289118,-0.019632155,-0.024715202,0.00015138647,0.014993773,0.005686697,0.055830058,-0.087244615,0.014277782,0.0054429956,-0.029063378,0.0031674253,-0.0015263931,-0.017177438,-0.014158963,-0.009414523,0.0022411558,0.013006134,-0.01916693,0.016224789,-0.0017094557,-0.009718764,0.0023340438,-0.02369975,-0.034663647,0.0029187966,-0.030110613,0.016459428,-0.012836054,0.01623352,0.046541035,0.0156043535,-0.003995657,0.0056910627,0.022317259,0.03182072,0.018006131,-0.02704733,-0.027613068,-0.013476994,0.029174436,0.02416341,-0.004791739,0.016309317,-0.011434752,0.02102925,0.002024134,-0.0068986933,0.0060524745,-0.040946838,-0.020393182,-0.00395828,0.007920159,-0.008687544,-0.010176819,-0.010957242,-0.023820803,-0.039145704,-0.025648499,-0.009286982,0.0016558788,0.004742986,-0.008470269,-0.030668762,0.016570436,-0.011758999,0.007893169,-0.029201811,-0.0025635816,-0.010470873,0.0039053601,0.00932271,0.030774748,0.01996125,0.028125333,-0.0009234944,0.022280704,0.009500723,-0.01221853,0.010301874,0.014964263,0.010262031,0.0005264498,0.0017054568,0.029201275,-0.00023536384,0.0013240901,-0.039092526,-0.011193513,-0.10108397,-0.015325976,-0.015576055,0.0071996576,0.0081074415,-0.012329046,0.0039498536,-0.012041222,-0.02927418,-0.012455365,0.002788975,0.009413445,-0.02575736,-0.023611858,0.0002512642,0.003829798,0.0015313977,0.0068709,0.0023646655,-0.029785452,-0.03337418,0.0016463172,0.03388867,-0.018472636,-0.019342348,-0.013054259,-0.0062934067,-0.018933568,0.013778669,-0.011351395,-0.008628465,-0.12129888,-0.0016974699,0.0040477994,-0.018506028,0.010642462,-0.0065943394,0.022209844,0.0059057795,-0.0167967,-0.04315446,-0.017483484,-0.011869582,-0.017069269,-0.0014929362,0.0066404534,0.12061038,-0.011025062,-0.023808027,-0.012707268,-0.024119504,-0.0043366803,-0.018473726,-0.0071524275,0.01608182,0.011190075,-0.025810387,0.037696224,-0.009416807,-0.007632583,0.024726542,0.0026254812,0.011227151,-0.015608308,-0.0041288957,0.029581573,-0.0026937118,0.028803943,-0.019944115,-0.003797354,0.031186625,0.021631164,-0.00047875658,-0.00911895,0.0047952123,-0.010404883,-0.020319495,-0.0048877145,-0.033359326,-0.0070476686,0.00076458964,-0.018853603,-0.06986927,-0.0038293556,0.0025948207,0.022707624,-0.00030887238,0.007102869,0.008581727,0.034068745,0.017429842,0.028281664,0.00020772518,0.0022377372,-0.007639753,0.0033702622,-0.0017886937,0.0035353613,0.024884114,-0.002949208,0.016644368,-0.017616063,-0.01096026,-5.2578856e-05,-0.009379904,0.013928423,-0.04116685,-0.0070764422,0.021830615,0.01649704,-0.0068728384,0.0046745143,-0.006929474,0.0032645992,-0.02487901,0.024619829,-0.0007474247,0.0020992788,-0.02237843,0.010503933,-0.05132444,0.0033182423,0.006036659,-0.002013903,0.013399451,0.0092322705,0.013585694,-0.017386092,0.007871382,0.011757742,0.00517691,0.012623131,-0.044029415,0.0013887249,-0.03459158,0.004235624,0.03112491,0.0020857302,0.027529007,0.025197007,0.007227713]",{"tags":38,"relatedLang":49,"relatedPosts":53},[39,41,43,45,47],{"name":13,"slug":40},"visual-reasoning",{"name":16,"slug":42},"agentic-reasoning",{"name":17,"slug":44},"reinforcement-learning",{"name":14,"slug":46},"functional-token",{"name":15,"slug":48},"latent-reasoning",{"id":28,"slug":50,"title":51,"language":52},"atlas-one-token-visual-reasoning-en","ATLAS Makes Visual Reasoning Use One Token","en",[54,60,66,72,78,84],{"id":55,"slug":56,"title":57,"cover_image":58,"image_url":58,"created_at":59,"category":27},"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":61,"slug":62,"title":63,"cover_image":64,"image_url":64,"created_at":65,"category":27},"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":67,"slug":68,"title":69,"cover_image":70,"image_url":70,"created_at":71,"category":27},"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":73,"slug":74,"title":75,"cover_image":76,"image_url":76,"created_at":77,"category":27},"bfd03801-a200-4222-9370-8b441be41483","entitybench-long-range-video-consistency-zh","EntityBench 盯住長片一致性","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778911845686-4mc8.png","2026-05-16T06:10:27.85068+00:00",{"id":79,"slug":80,"title":81,"cover_image":82,"image_url":82,"created_at":83,"category":27},"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":85,"slug":86,"title":87,"cover_image":88,"image_url":88,"created_at":89,"category":27},"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",[91,96,101,106,111,116,121,126,131,136],{"id":92,"slug":93,"title":94,"created_at":95},"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":97,"slug":98,"title":99,"created_at":100},"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":102,"slug":103,"title":104,"created_at":105},"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":107,"slug":108,"title":109,"created_at":110},"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":112,"slug":113,"title":114,"created_at":115},"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":117,"slug":118,"title":119,"created_at":120},"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":122,"slug":123,"title":124,"created_at":125},"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":127,"slug":128,"title":129,"created_at":130},"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":132,"slug":133,"title":134,"created_at":135},"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":137,"slug":138,"title":139,"created_at":140},"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"]