[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"topic-zh-longmemeval-v2-agent-":3},{"cluster":4,"timeline":17},{"id":5,"slug":6,"title":7,"pinned":8,"status":9,"summary":10,"category":11,"language":12,"created_at":13,"merged_into":14,"article_count":15,"first_seen_at":13,"last_updated_at":16},"0c35a120-52fc-41fc-afa3-d404eb934158","longmemeval-v2-agent-","LongMemEval-V2：測 agent 長期記憶",false,"active","LongMemEval-V2 用 451 題測試 agent 能否記住 Web 環境經驗，而不只是使用者歷史；結果顯示以 coding agent 蒐證的記憶法準確率最高，但延遲也更高。","research","zh","2026-05-13T06:20:30.5093+00:00",null,218,"2026-07-17T12:40:00.07888+00:00",[18,25,32,39,46,53,60,67,74,81,88,95,102,109,116,123,130,137,144,151,158,165,172,179,186,193,200,207,214,221,228,235,242,249,256,263,270,277,284,291,298,305,312,319,326,333,340,347,354,361],{"id":19,"slug":20,"title":21,"summary":22,"category":11,"image_url":23,"cover_image":23,"published_at":24,"is_canonical_seed":8},"ff3d62ca-fd2d-4779-87e3-3a8817cdaa33","meanflownft-forward-process-rl-average-velocity-zh","MeanFlowNFT 讓少步生成也能做 RL","MeanFlowNFT 把前向流程 RL 接到 MeanFlow 的平均速度生成器上，保住少步採樣的速度，也讓對齊訓練變得可行。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1784271777053-13dr.png","2026-07-17T07:02:26.952+00:00",{"id":26,"slug":27,"title":28,"summary":29,"category":11,"image_url":30,"cover_image":30,"published_at":31,"is_canonical_seed":8},"ed59677b-bc56-4c01-b1e8-163b6c6744dc","analysis-driven-transformer-linearization-zh","線性化 Transformer 但不掉品質","這篇論文證明，凍結骨幹後做分析驅動的線性化改造，能更接近原本 Transformer 的品質與長上下文行為。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783580571758-gqu7.png","2026-07-09T07:02:26.188+00:00",{"id":33,"slug":34,"title":35,"summary":36,"category":11,"image_url":37,"cover_image":37,"published_at":38,"is_canonical_seed":8},"3e8fbc00-9a1f-4e79-bfc2-ca933bf09eb9","co-lmlm-continuous-query-limited-memory-models-zh","Co-LMLM 讓 LLM 持續查知識","Co-LMLM 把固定 KB 查詢改成連續向量查詢，讓模型能在不把事實寫進權重的前提下，提高事實精準度。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783578779210-wi5d.png","2026-07-09T06:32:30.391+00:00",{"id":40,"slug":41,"title":42,"summary":43,"category":11,"image_url":44,"cover_image":44,"published_at":45,"is_canonical_seed":8},"404bac33-b9b4-41bb-bb9a-1d98a63aa536","evaluation-protocols-fine-tuned-llms-2026-zh","2026 微調 LLM 評估流程","建立一套可落地的微調 LLM 評估流程，涵蓋任務指標、LLM 評審、安全檢查與人工複核。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783101776530-b8eu.png","2026-07-03T18:02:24.567+00:00",{"id":47,"slug":48,"title":49,"summary":50,"category":11,"image_url":51,"cover_image":51,"published_at":52,"is_canonical_seed":8},"8f3122c8-9eb1-4aa6-b780-3b62003b3418","deepspec-data-regeneration-pipeline-qwen3-eagle3-zh","DeepSpec 應被視為資料重生管線，而不是訓練技巧","DeepSpec 最好的理解方式，是把它當成對對話資料做重生的管線，而不是一個單純的訓練技巧。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783080165006-321z.png","2026-07-03T12:02:18.361+00:00",{"id":54,"slug":55,"title":56,"summary":57,"category":11,"image_url":58,"cover_image":58,"published_at":59,"is_canonical_seed":8},"6cfddc0d-ce6e-4a14-baf7-3531bf32bc5d","program-as-weights-fuzzy-functions-zh","PAW把提示詞編成可重用工具","PAW把自然語言任務規格編成可離線執行的小型神經工具，讓模糊任務能重複使用、降低推理成本。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783062178440-pnt0.png","2026-07-03T07:02:32.577+00:00",{"id":61,"slug":62,"title":63,"summary":64,"category":11,"image_url":65,"cover_image":65,"published_at":66,"is_canonical_seed":8},"5bd0dc27-5a7f-4563-8086-acccc98eb2fc","lacuna-llm-unlearning-localization-testbed-zh","LACUNA：檢驗 LLM 真的有沒有忘記","LACUNA 用已知參數位置，直接測 LLM unlearning 是否真的擦掉記憶，而不只是讓模型表面上不說。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783060373883-d92j.png","2026-07-03T06:32:31.274+00:00",{"id":68,"slug":69,"title":70,"summary":71,"category":11,"image_url":72,"cover_image":72,"published_at":73,"is_canonical_seed":8},"ff17d0f0-f249-41e3-b62e-658282631451","persistent-state-ai-agents-attack-surface-zh","持久狀態 AI 代理的新攻擊面","這篇論文證明，能跨 PR 持續工作的 AI coding agent 可以把攻擊拆散到多次提交，讓傳統 diff 監控更難抓到。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783058580349-ldhu.png","2026-07-03T06:02:30.275+00:00",{"id":75,"slug":76,"title":77,"summary":78,"category":11,"image_url":79,"cover_image":79,"published_at":80,"is_canonical_seed":8},"4c1c0228-6f8e-4be6-b948-61bc48e67746","language-critiques-imitation-learning-zh","語言批註讓模仿學習更準","研究顯示，把自然語言批註直接拿來訓練策略，比只用分數回饋更能從次佳示範中學到有效行為。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782975775937-7kd6.png","2026-07-02T07:02:28.753+00:00",{"id":82,"slug":83,"title":84,"summary":85,"category":11,"image_url":86,"cover_image":86,"published_at":87,"is_canonical_seed":8},"5b59165e-18fd-4c10-afa4-1307e39a11f0","one-transformer-layer-can-carry-rl-gains-zh","單層 Transformer 也能扛住 RL 增益","這篇研究指出，強化學習後訓練的主要增益，可能集中在 Transformer 少數層，甚至只訓練一層就能拿回大部分效果。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782973979895-px83.png","2026-07-02T06:32:29.168+00:00",{"id":89,"slug":90,"title":91,"summary":92,"category":11,"image_url":93,"cover_image":93,"published_at":94,"is_canonical_seed":8},"269ae2f5-ce51-4e00-8771-eab2f264e074","bineval-binary-questions-llm-evals-zh","BINEVAL 用二元問題評估 LLM 輸出","BINEVAL 把原本模糊的 LLM 評分拆成一連串是／否題目，主打更可檢查、可除錯的評估流程，並在多個 benchmark 上對上 G-Eval 與 UniEval。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782927171316-0dkv.png","2026-07-01T17:32:23.648+00:00",{"id":96,"slug":97,"title":98,"summary":99,"category":11,"image_url":100,"cover_image":100,"published_at":101,"is_canonical_seed":8},"0ee8cc51-c309-4477-8914-82f7824161e3","rlmf-teaches-llms-express-uncertainty-better-zh","RLMF 讓 LLM 更會表達不確定","RLMF 用元認知回饋訓練 LLM，讓模型的自信表達更貼近真實不確定性，且保留原本準確率。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782887572465-ag3v.png","2026-07-01T06:32:28.695+00:00",{"id":103,"slug":104,"title":105,"summary":106,"category":11,"image_url":107,"cover_image":107,"published_at":108,"is_canonical_seed":8},"00a1011e-5e65-4d16-9ed4-208b129055d8","qval-dense-supervision-testbed-long-horizon-agents-zh","QVal 先測密集監督再訓練","QVal 提供一種免訓練的方式，先比較長鏈路 LLM agent 的密集監督訊號，再決定要不要投入完整訓練。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782886677076-dhmx.png","2026-07-01T06:17:33.643+00:00",{"id":110,"slug":111,"title":112,"summary":113,"category":11,"image_url":114,"cover_image":114,"published_at":115,"is_canonical_seed":8},"678566b8-297f-4efc-8b78-9e80c4cc1140","self-explanation-training-tracks-model-behavior-zh","固定解釋資料也能追上模型行為","這篇研究指出，固定不更新的解釋資料，仍能讓模型學會描述自己當下的行為，而且會跟著行為變化一起移動。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782885778930-88u5.png","2026-07-01T06:02:30.465+00:00",{"id":117,"slug":118,"title":119,"summary":120,"category":11,"image_url":121,"cover_image":121,"published_at":122,"is_canonical_seed":8},"8c68f33b-8ab7-40f8-96d6-cb854eb9b654","worldevolver-self-evolving-world-models-llm-planning-zh","WorldEvolver 讓 LLM 代理自我修正前瞻","WorldEvolver 透過測試時記憶修訂，讓 LLM 代理在不改權重下更新前瞻與規劃能力。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782801181629-6zpm.png","2026-06-30T06:32:28.892+00:00",{"id":124,"slug":125,"title":126,"summary":127,"category":11,"image_url":128,"cover_image":128,"published_at":129,"is_canonical_seed":8},"c8de2902-230f-4a9f-a6c2-75bb234ca422","levo-2-full-length-song-generation-zh","LeVo 2 用分層建模做完整歌曲生成","LeVo 2 透過分層表示與漸進式後訓練，改善完整歌曲生成的穩定性、可控性與音樂性。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782800283385-laim.png","2026-06-30T06:17:32.006+00:00",{"id":131,"slug":132,"title":133,"summary":134,"category":11,"image_url":135,"cover_image":135,"published_at":136,"is_canonical_seed":8},"e6db7892-cfae-4a48-ae72-0b56f71e737a","vlk-synthetic-humanoid-loco-manipulation-zh","VLK 用合成場景訓練人形機器人","VLK 證明可用重建室內場景合成視覺、語言與運動監督，訓練人形機器人完成導航與單物件搬運。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782799374468-q7z1.png","2026-06-30T06:02:29.609+00:00",{"id":138,"slug":139,"title":140,"summary":141,"category":11,"image_url":142,"cover_image":142,"published_at":143,"is_canonical_seed":8},"d6f25c66-98f5-4971-8d1d-487fb5fe1881","claude-sonnet-46-sre-benchmark-rootly-zh","Claude Sonnet 4.6 對上 SRE 工作更接近 Opus","Rootly 的 SRE benchmark 顯示，Claude Sonnet 4.6 在事故調查上已接近 Opus 4.6，且每百萬輸出 Token 成本低約 40%。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782750780131-xelc.png","2026-06-29T16:32:28.446+00:00",{"id":145,"slug":146,"title":147,"summary":148,"category":11,"image_url":149,"cover_image":149,"published_at":150,"is_canonical_seed":8},"29321237-6e9a-4271-b9fb-e43e798d5dff","glm-52-beats-claude-semgrep-idor-test-zh","GLM 5.2 在 IDOR 測試贏過 Claude","Semgrep 的 IDOR benchmark 顯示，GLM 5.2 在純提示詞條件下 F1 贏過 Claude Code，且每個漏洞成本約 0.17 美元。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782749882713-7i5n.png","2026-06-29T16:17:31.903+00:00",{"id":152,"slug":153,"title":154,"summary":155,"category":11,"image_url":156,"cover_image":156,"published_at":157,"is_canonical_seed":8},"5172bfc7-34c8-4477-a177-ffa615497ecf","opd-distillation-skills-without-bruteforce-rl-zh","OPD 讓你把技能蒸餾進模型","我拆 On-Policy Distillation 的做法，整理成可直接套用的後訓練模板，少碰硬拼 RL。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782730101413-5wjx.png","2026-06-29T10:47:57.447+00:00",{"id":159,"slug":160,"title":161,"summary":162,"category":11,"image_url":163,"cover_image":163,"published_at":164,"is_canonical_seed":8},"6f5be102-5764-44f1-ab3f-722fc5c32c23","google-deepmind-turns-science-into-tools-zh","Google DeepMind把AI變研究工具","我拆 Google DeepMind 這頁，看看它怎麼把 AI 從口號包成研究工具，讓開發者能抄走定位、流程與模板。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782721105628-g4op.png","2026-06-29T08:17:57.702+00:00",{"id":166,"slug":167,"title":168,"summary":169,"category":11,"image_url":170,"cover_image":170,"published_at":171,"is_canonical_seed":8},"c649adb7-c8ae-4ade-a092-2c0d53beeb71","measuring-llm-behavior-portability-zh","LLM 行為不一定可移植","這篇研究指出，LLM 在一個情境學到的行為，常常無法穩定轉移到報酬等價但表面不同的環境。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782717472977-na8g.png","2026-06-29T07:17:29.583+00:00",{"id":173,"slug":174,"title":175,"summary":176,"category":11,"image_url":177,"cover_image":177,"published_at":178,"is_canonical_seed":8},"637c3016-e364-4bfe-904e-5e60a18ed678","prompt-injection-ai-security-problem-zh","Prompt injection 已是 AI 資安問題","Prompt injection 會用隱藏文字操控 LLM。近期測試顯示，像 DeepSeek-R1 這類模型，仍可能在注入攻擊下失手。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782716580916-m1nm.png","2026-06-29T07:02:36.16+00:00",{"id":180,"slug":181,"title":182,"summary":183,"category":11,"image_url":184,"cover_image":184,"published_at":185,"is_canonical_seed":8},"118680f5-6212-4535-986a-50c4a0e71699","solver-choice-nash-equilibrium-selection-zh","求解器會改變納許均衡","這篇論文指出，在多重納許均衡的零和博弈裡，不同求解器不只會收斂，還會系統性挑不同的均衡。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782714784181-t42d.png","2026-06-29T06:32:31.048+00:00",{"id":187,"slug":188,"title":189,"summary":190,"category":11,"image_url":191,"cover_image":191,"published_at":192,"is_canonical_seed":8},"89159fcf-2fbb-4b72-9e05-7928e609a925","dexcompose-reuses-dexterous-policies-across-tasks-zh","DexCompose 讓手部技能可重用","DexCompose 用手指級的動作分工，把已訓練好的靈巧手策略組成多任務操作，並在 16 個任務上達到 77.4% 平均成功率。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782712975186-mj1e.png","2026-06-29T06:02:28.127+00:00",{"id":194,"slug":195,"title":196,"summary":197,"category":11,"image_url":198,"cover_image":198,"published_at":199,"is_canonical_seed":8},"b67223ce-e3cb-4161-9df6-b384e364eb87","hawor-hand-motion-mano-params-zh","HaWoR 把手部重建收斂成 MANO","我拆 HaWoR 之後，只剩一個重點：它不是在猜手的網格，而是在預測 MANO 參數，整個 pipeline 會乾淨很多。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782705793656-d9q2.png","2026-06-29T04:02:46.408+00:00",{"id":201,"slug":202,"title":203,"summary":204,"category":11,"image_url":205,"cover_image":205,"published_at":206,"is_canonical_seed":8},"84f239c4-94f9-4a73-a284-21fb271139cd","nvidia-30000-grant-usc-health-ai-zh","NVIDIA 3 萬美元補助瞄準 USC 健康 AI","USC 公布 NVIDIA 3 萬美元學術補助，主打健康與 AI 研究，申請截止日是 2026 年 6 月 30 日，內容是 H100 GPU 時數，不是現金。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782692274946-mih0.png","2026-06-29T00:17:30.598+00:00",{"id":208,"slug":209,"title":210,"summary":211,"category":11,"image_url":212,"cover_image":212,"published_at":213,"is_canonical_seed":8},"5431a65e-76da-4a2a-96c5-73a6a7635903","cuda-toolkit-13-3-fixes-nested-divergence-bug-zh","CUDA 13.3 修掉巢狀分歧編譯錯誤","CUDA Toolkit 13.3 修掉一個從 12.8 就存在的編譯器錯誤。這個 bug 會在巢狀分歧的 GPU kernel 裡弄壞暫存器值，結果可能是算錯，不是當掉。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782676982948-afr9.png","2026-06-28T20:02:39.335+00:00",{"id":215,"slug":216,"title":217,"summary":218,"category":11,"image_url":219,"cover_image":219,"published_at":220,"is_canonical_seed":8},"37acb4f1-36aa-4cbd-8c2f-0733c39a074f","eagle3-real-speedup-kimi-k25-mi325x-zh","EAGLE3 才是 Kimi-K2.5 在 MI325X 上真正的加速器","我認為 Kimi-K2.5-W4A8 在 AMD MI325X 上變快，主因是 EAGLE3 的 speculative decoding，不是 kernel 微調；真正改變的是解碼幾何，而不是單一算子效率。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782640968852-4e46.png","2026-06-28T10:02:26.123+00:00",{"id":222,"slug":223,"title":224,"summary":225,"category":11,"image_url":226,"cover_image":226,"published_at":227,"is_canonical_seed":8},"7c4c30b3-b2a8-48a7-b2ea-96c40c16ae19","llm-fine-tuning-turns-generic-models-into-domain-tools-zh","LLM 微調把通用模型變專用工具","我把企業 LLM 微調拆成一套可直接抄的流程：先判斷該不該微調，再做資料清理、模型選擇、評估與上線。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782569910494-nhtn.png","2026-06-27T14:17:56.595+00:00",{"id":229,"slug":230,"title":231,"summary":232,"category":11,"image_url":233,"cover_image":233,"published_at":234,"is_canonical_seed":8},"a90ab5b6-f647-4cef-85af-35ff7bb21a93","autoregressive-boltzmann-generators-ditch-flows-zh","ArBG 改用自回歸做分子採樣","ArBG 把 Boltzmann Generator 從 flow 改成自回歸建模，主打更快、更可擴展的平衡態分子採樣。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782455577323-vrvt.png","2026-06-26T06:32:30.033+00:00",{"id":236,"slug":237,"title":238,"summary":239,"category":11,"image_url":240,"cover_image":240,"published_at":241,"is_canonical_seed":8},"93b19c63-dbfd-4277-92b5-b5a60946fd65","river-llm-reinforcement-learning-without-answers-zh","RiVER 讓 LLM 不靠標準答案也能學","RiVER 證明 LLM 可以只靠執行回饋與分數校準，在沒有標準答案的任務上學出更好的策略。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782454671897-i8l3.png","2026-06-26T06:17:26.969+00:00",{"id":243,"slug":244,"title":245,"summary":246,"category":11,"image_url":247,"cover_image":247,"published_at":248,"is_canonical_seed":8},"cd38b72e-b309-493d-b36f-684745ff5f7e","danceopd-on-policy-generative-field-distillation-zh","DanceOPD：把修圖技能蒸餾進同一模型","DanceOPD 用 on-policy 蒸餾，把文生圖與編輯能力放進同一個 flow-matching 模型，減少彼此互相干擾。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782453784592-x1gk.png","2026-06-26T06:02:33.114+00:00",{"id":250,"slug":251,"title":252,"summary":253,"category":11,"image_url":254,"cover_image":254,"published_at":255,"is_canonical_seed":8},"af1a155b-d8e6-4575-a014-959aef283098","microsoft-ai-team-collaboration-cfp-2026-zh","Microsoft 砸錢研究團隊協作 AI","Microsoft Research 開出 2026 春季 CFP，研究 AI 怎麼幫團隊協作。每案約 5 萬到 7.5 萬美元，重點放在會議、分工、共識與群體生產力。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782415981776-jikr.png","2026-06-25T19:32:33.146+00:00",{"id":257,"slug":258,"title":259,"summary":260,"category":11,"image_url":261,"cover_image":261,"published_at":262,"is_canonical_seed":8},"2cc1973d-a7a5-4031-8ed3-e05ca5d335fd","ai-papers-code-music-rare-disease-zh","3 篇 AI 論文：程式、音樂、罕病診斷","知乎整理 3 篇 arXiv AI 論文，涵蓋程式生成、即時音樂與罕病診斷。重點不在聊天，而是不同架構如何處理結構、延遲與專業推理。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782372792462-buxp.png","2026-06-25T07:32:27.26+00:00",{"id":264,"slug":265,"title":266,"summary":267,"category":11,"image_url":268,"cover_image":268,"published_at":269,"is_canonical_seed":8},"f9ec6d6f-80a9-4a8e-b3ea-1eb5231aa796","new-nlp-papers-agent-memory-tool-use-zh","新 NLP 論文盯上代理記憶與工具使用","6 月 24 日的 arXiv 論文整理，聚焦 agent 記憶、工具使用評估與對話式搜尋，對做 AI 代理和搜尋助理的人很實用。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782371888802-40t8.png","2026-06-25T07:17:39.038+00:00",{"id":271,"slug":272,"title":273,"summary":274,"category":11,"image_url":275,"cover_image":275,"published_at":276,"is_canonical_seed":8},"a875d002-f6f0-4139-abc1-f1602bc42fee","self-distillation-shrinks-output-diversity-zh","自蒸餾會縮小模型多樣性","這篇論文指出，自蒸餾能拉高 pass@1，卻會壓縮輸出多樣性，讓模型在分布外情境更脆弱。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782369171288-egwp.png","2026-06-25T06:32:26.547+00:00",{"id":278,"slug":279,"title":280,"summary":281,"category":11,"image_url":282,"cover_image":282,"published_at":283,"is_canonical_seed":8},"80a6e921-dfde-4861-ba61-382e195ec94c","revengebench-reverse-engineering-game-policies-zh","RevengeBench：反推遊戲政策的測試框架","RevengeBench把隱藏遊戲政策的反向工程做成可測試任務，證明主動探測能讓 LLM 更接近還原可執行策略。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782368284240-86sh.png","2026-06-25T06:17:29.001+00:00",{"id":285,"slug":286,"title":287,"summary":288,"category":11,"image_url":289,"cover_image":289,"published_at":290,"is_canonical_seed":8},"978e67d0-1acb-479e-af06-9ead35e4eb74","learning-action-priors-cross-embodiment-manipulation-zh","先學動作先驗，再對齊多模態","這篇論文證明，先用動作軌跡學出 motion prior，再做視覺語言對齊，能讓跨具身操作訓練更快、成功率更高。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782367376604-ffk9.png","2026-06-25T06:02:29.656+00:00",{"id":292,"slug":293,"title":294,"summary":295,"category":11,"image_url":296,"cover_image":296,"published_at":297,"is_canonical_seed":8},"4a0bbfe8-be40-4add-95c8-7ed1d38a641f","opsd-user-feedback-training-loop-zh","OPSD 讓你把點擊變訓練","我把 OPSD 拆成一個可直接抄的閉環：怎麼把隱性用戶回饋變成校正資料，再持續訓練模型。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782335103935-0efp.png","2026-06-24T21:04:40.402+00:00",{"id":299,"slug":300,"title":301,"summary":302,"category":11,"image_url":303,"cover_image":303,"published_at":304,"is_canonical_seed":8},"a2242009-98d7-409c-9f22-d825a81fef2e","ultraquant-4bit-kv-caching-agents-zh","UltraQuant：4-bit KV 快取加速長代理","UltraQuant 證明 4-bit KV 快取能讓長篇多輪代理在更少記憶體下維持更多上下文，並在後段輪次明顯加速服務。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782331375909-uhyy.png","2026-06-24T20:02:32.542+00:00",{"id":306,"slug":307,"title":308,"summary":309,"category":11,"image_url":310,"cover_image":310,"published_at":311,"is_canonical_seed":8},"261f4fc9-e9c8-413c-b222-a31008ec2bcf","flux3d-3d-gaussian-generation-diffusion-zh","FLUX3D 讓 3DGS 保住細節","FLUX3D 透過對齊稀疏 3D latent 與密集 2D token，改善影像轉 3D Gaussian 時的細節流失問題。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782284582760-3ja7.png","2026-06-24T07:02:37.284+00:00",{"id":313,"slug":314,"title":315,"summary":316,"category":11,"image_url":317,"cover_image":317,"published_at":318,"is_canonical_seed":8},"5f0498a5-626f-4217-8c67-3b3404c7c172","insight-vla-self-guided-skill-acquisition-zh","InSight 讓 VLA 自學新技能","InSight 讓 vision-language-action 政策在沒有目標任務人類示範的情況下，自行拆解原始動作、找出缺口，並收集新技能資料。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782282781514-vo4g.png","2026-06-24T06:32:30.777+00:00",{"id":320,"slug":321,"title":322,"summary":323,"category":11,"image_url":324,"cover_image":324,"published_at":325,"is_canonical_seed":8},"29ea0e09-dbd2-406d-9d74-fd851c59a4f7","anthropic-right-alarm-recursive-self-improvement-zh","Anthropic 警告遞迴自我改進是對的，但真正的問題是 AI 控制已經失速","Anthropic 對遞迴自我改進的警告是正確的，但更大的問題是 AI 的治理速度已經跟不上能力擴張。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782263867507-ive3.png","2026-06-24T01:17:20.48+00:00",{"id":327,"slug":328,"title":329,"summary":330,"category":11,"image_url":331,"cover_image":331,"published_at":332,"is_canonical_seed":8},"e6906894-cfe6-48a2-84a2-cd34e6a95186","openai-bug-hunt-chrome-safari-firefox-zh","OpenAI 一週挖出三大瀏覽器漏洞","OpenAI 研究團隊一週內找出 Chrome、Safari、Firefox 的可利用漏洞，顯示瀏覽器核心引擎仍有不少攻擊面，Mozilla 也因此提前修補 Firefox。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782258476786-r07m.png","2026-06-23T23:47:30.613+00:00",{"id":334,"slug":335,"title":336,"summary":337,"category":11,"image_url":338,"cover_image":338,"published_at":339,"is_canonical_seed":8},"19c48417-946e-4c23-865f-87ffcc754d1a","llm-fine-tuning-production-2026-zh","2026 生產環境 LLM 微調指南","AgamiSoft 的 2026 指南整理了生產環境 LLM 微調選型，從開源模型、資料整理、評估到部署，重點放在成本、延遲與可維護性。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782252178755-rwnv.png","2026-06-23T22:02:33.154+00:00",{"id":341,"slug":342,"title":343,"summary":344,"category":11,"image_url":345,"cover_image":345,"published_at":346,"is_canonical_seed":8},"8531d5f9-60f1-4a4b-94a3-323b82990f06","lifescibench-tests-biotech-models-zh","LifeSciBench 讓模型先過科研關","我拆 LifeSciBench 怎麼把生命科學模型評估拉回真實科研工作，順手給你一份可直接抄的評測模板。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782198202904-lzgm.png","2026-06-23T07:02:47.173+00:00",{"id":348,"slug":349,"title":350,"summary":351,"category":11,"image_url":352,"cover_image":352,"published_at":353,"is_canonical_seed":8},"f6fbee54-8ee4-4ad1-a6bb-a3f2ac890430","coordex-humanoid-loco-manipulation-priors-zh","CoorDex 讓人形機器人邊走邊操作","CoorDex 把人形機器人的身體與手部控制壓成 latent priors，讓它能在移動中完成精細操作。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782196378261-6x4x.png","2026-06-23T06:32:32.255+00:00",{"id":355,"slug":356,"title":357,"summary":358,"category":11,"image_url":359,"cover_image":359,"published_at":360,"is_canonical_seed":8},"7171fed6-f304-4f46-9efe-f691ea304b65","randomized-yarn-long-context-reasoning-zh","Randomized YaRN 讓長上下文更穩","Randomized YaRN 透過隨機化位置編碼與長度課程，讓只看過短上下文訓練的 LLM，更能推廣到 16K 到 128K 的長推理窗口。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782195475543-rsm6.png","2026-06-23T06:17:32.36+00:00",{"id":362,"slug":363,"title":364,"summary":365,"category":11,"image_url":366,"cover_image":366,"published_at":367,"is_canonical_seed":8},"56da2379-5b47-4f3d-827f-e50d8be5d015","autodex-automates-dexterous-grasp-data-collection-zh","AutoDex 自動蒐集靈巧抓取資料","AutoDex 把靈巧抓取的實體試驗、成功失敗標記與重置流程全自動化，讓資料蒐集不再卡在人工作業。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782194579214-qano.png","2026-06-23T06:02:31.186+00:00"]