[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-askstream-proactive-asking-streaming-video":3,"topics-all":38,"news-related-643e4f3e-2ec0-47eb-b732-4d8da1b18e70":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"643e4f3e-2ec0-47eb-b732-4d8da1b18e70","视频模型该学会追问:离线52.18%,流式只剩7.22%","中科院自动化所等机构发布 AskStream 基准，给流式视频理解补上第三个动作：追问。基准含 438 个视频、1401 个问题和 2714 个人工标注时间窗口。论文评测里离线模型 Ask-F1 最高也只有 52.18%，原生流式主动模型最高仅 7.22%。消融实验进一步显示，模型缺的不是回答能力，而是判断何时该问。","流式视频的智能体已经能\"边看边答\"了，可一旦画面里该看的东西暂时没出现、用户的问题又含糊其辞，模型就卡在两难：继续等，等到没完；直接答，答案大概率是幻觉。中科院自动化所、北京智源等机构发布的 AskStream 基准（[论文原文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38678)）把这个两难摆到了台面上，并给出了一个新答案：主动交互不该只有\"等\"和\"答\"两种动作，还得有第三种——问。\n\n## 从\"等\"和\"答\"到\"等、问、答\"\n\n论文的出发点很朴素。现有主动视频理解默认了一个前提：回答所需的信息最终会在视频里出现。于是模型只需在\"继续等\"和\"给出答案\"之间二选一。可现实场景经常打破这个前提——镜头始终没拍到被遮挡的物体，用户只说\"提醒我那个人出现\"却没说清是谁。这种时候等待不会带来新信息，硬答只能编。\n\nAskStream 把动作空间扩成 wait、ask、answer 三类，并把\"不确定性\"拆成两个来源。视频不确定性包括关键视觉证据缺失、视角信息缺失、个体概念缺失和个性化动作概念缺失；用户不确定性包括问题歧义、问题前提错误、用户偏好缺失和偏好冲突。四类任务、八个子任务。\n\n## 438 个视频和 2714 个动作窗口\n\n基准规模不算大，标注却下得很重：438 个视频、1401 个问题，覆盖 9 个领域，视频时长从 16.72 秒到 118 分钟不等，平均每个视频 3.2 轮交互。作者标的不是单个时间戳，而是\"有效动作窗口\"——wait、ask、answer 三类窗口的累计时长分别是 33.92、30.47 和 12.00 小时，共 2714 个人工标注的帧级时间戳。标注可靠性也做了检验：抽 160 个样本由两名标注员独立重标，ASK 窗口的平均时序交并比 0.962，ANSWER 窗口 0.971。\n\n评测分选择题和开放式两套。开放式由 Gemini-3.8 Flash 当裁判，作者随机抽 300 轮找人工复核，裁判与人工共识的一致率 93.50%，而且对 Gemini 自己生成的回复（93.75%）和别家模型的回复（93.27%）看不出明显自偏好。\n\n## 离线 52.18%，主动流式模型 7.22%\n\n分数是最扎眼的部分。论文的选择题评测里，离线闭源模型中 Gemini-3.8 Flash 的 Ask-F1 总分最高，为 52.18%，其中\"问\"的子分 44.63%、\"答\"的子分 62.80%；GPT-5.6-Sol 是 20.50%，Claude Sonnet 5 是 22.48%。开源离线模型整体更低，MiMo-V2.5 14.66%、Kimi K2.6 16.52%、Qwen-3.8 Flash 17.60%、DeepSeek-V4.1-Flash 19.69%。\n\n真正尴尬的是那批带主动能力的流式模型：论文评测中 MOSS-VL-Realtime、MMDuet2、VideoLLM-online 大量子任务直接记 0，MiniCPM-o 4.5 是 0.12，TimeChat-Online 4.18，JoyAI-VL-Interaction 4.91，其中分数最高的 AURA 也只有 7.22%。换到开放式设置，Gemini-3.8 Flash 掉到 29.19%，AURA 只剩 4.01%。论文的判断很直白：这些模型不是不太会主动，是不会\"问\"。\n\n## 真正的瓶颈是\"何时问\"\n\n论文用两组消融，把\"问\"的两个维度拆开看。\n\n第一组是 oracle 实验。在关键视觉证据缺失任务上，只要把\"什么时候该问\"的正确答案告诉模型，Ask-F1 就从 37.17% 涨到 58.30%，涨了 21.13 个点；同时给出\"何时问\"和\"问什么\"两项信息，整体增益论文报为 43.83%↑。也就是说，缺的主要是时间判断，不是措辞。\n\n第二组是提问预算。把允许提问的次数从 1 一路放宽到 16，分数一路涨，但在 52.35% 附近就饱和了。多给提问机会，并不能替代\"知道该问\"。\n\n还有个容易被忽略的细节：作者先做了模糊实验，验证追问本身有没有价值。画面被模糊后模型准确率 42.3%，模型主动索要并拿到正确证据后升到 62.3%，已经接近完全清晰画面下的 63.1%；但补进来的若是不相关证据，收益就基本没有。真正有用的不是多看几帧，而是拿到与当前问题匹配的信息。\n\n记忆结构也被单独算了账。用分层记忆框架 StreamHarness（保留最近 10 帧、每 10 帧压一条中期记忆、每四条中期记忆合并成一个长期记忆块，最多留八个）时 Ask-F1 是 52.18%，换成只留最近 10 帧的滑动窗口掉到 41.70%，均匀采样 128 帧只有 26.72%，64 帧 16.99%。\n\n## 这个基准在提示什么\n\nAskStream 有意把难度做高：所有模型都在\"看不清\"或\"说不清\"的前提里被测，评测每秒给一次动作机会，容差 ±1 秒。论文自己也列了四条局限——每轮只允许一次追问、只覆盖流式视频理解而没涉及具身智能体、每段交互都由明确的用户提问触发、真实场景的不确定性会跨越更长的时间尺度。\n\n但这套设计指向一个行业里容易被跳过的环节。\"知道什么时候该问\"这件事，现在几乎没有模型做得好：离线强模型也只拿到一半出头的分数，而真正为实时交互设计的流式模型基本交白卷。\n\n流式视频的下半场，看起来不在\"看得更清\"，而在\"知道自己没看清\"。AskStream 把这个能力量化成了一道 52.18 分的门槛——而主流流式模型离及格线还远。下一次这类榜单刷新时，值得看的可能不是谁的分数又涨了，而是哪个模型第一次敢在正确的时刻开口问一句。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38678","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"da116f20-377f-449f-9382-821d56f6fa64","streaming",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c3890278-03ab-4e89-96a9-1ff846bc4016","en","Video models should learn to ask: 52.18% vs 7.22%","AskStream adds a third action to streaming video: ask. The paper's evaluation puts offline models at 52.18% Ask-F1, streaming models at 7.22%.","Streaming video agents can already watch and answer at the same time. But the moment the thing worth seeing has not surfaced on screen yet, or the user's question is vague, the model is squeezed between two bad options: keep waiting, or answer and hallucinate. A new benchmark called AskStream, published by researchers from the Institute of Automation at the Chinese Academy of Sciences and Beijing Academy of Artificial Intelligence among others (paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38678), puts that squeeze on the table and proposes a third way out. Active interaction should not only offer \"wait\" and \"answer\" — it needs a third action: ask.\n\n## From wait-and-answer to wait, ask, answer\n\nThe premise behind existing proactive video understanding is simple: the information needed to answer will eventually show up in the video. Models are therefore left to choose between waiting longer and answering now. Real scenarios break that premise constantly. The camera never pans to the object hidden behind the cup, or the user says \"remind me when that person appears\" without saying which person. Waiting adds no new information there; answering means making something up.\n\nAskStream widens the action space to three: wait, ask, and answer, and it splits \"uncertainty\" along two axes. Video-side uncertainty covers missing key visual evidence, missing viewpoint information, missing individual concepts, and missing personalized action concepts. User-side uncertainty covers ambiguous questions, questions with false premises, missing user preferences, and conflicting preferences. That yields four task categories and eight subtasks.\n\n## 438 videos and 2,714 action windows\n\nThe benchmark is not huge, but the annotation is heavy. It contains 438 videos and 1,401 questions across nine domains, with clip durations ranging from 16.72 seconds to 118 minutes and an average of 3.2 interaction turns per video. Rather than marking single timestamps, the authors mark valid action windows: wait, ask, and answer windows accumulate to 33.92, 30.47, and 12.00 hours of effective duration respectively, across 2,714 manually annotated frame-level timestamps in total. Annotation reliability was checked too — 160 sampled instances were re-annotated independently by two annotators, reaching a mean temporal intersection-over-union of 0.962 for ASK windows and 0.971 for ANSWER windows.\n\nEvaluation runs in both a multiple-choice and an open-ended setting. In the open-ended setting Gemini-3.8 Flash serves as the judge; the authors sampled 300 turns for human review and reported 93.50% overall agreement between the Gemini judge and human consensus, with no clear self-preference (93.75% on Gemini-3.8 Flash's own outputs versus 93.27% on outputs from other models).\n\n## 52.18% offline, 7.22% streaming\n\nThe scores are the sharpest part. In the paper's multiple-choice evaluation, the highest offline closed-source Ask-F1 belongs to Gemini-3.8 Flash at 52.18%, with the asking sub-score at 44.63% and the answering sub-score at 62.80%. GPT-5.6-Sol reaches 20.50% and Claude Sonnet 5 reaches 22.48%. Offline open-source models sit lower still: MiMo-V2.5 at 14.66%, Kimi K2.6 at 16.52%, Qwen-3.8 Flash at 17.60%, and DeepSeek-V4.1-Flash at 19.69%.\n\nThe awkward part is what happens to models built specifically for streaming interaction. In the paper's evaluation MOSS-VL-Realtime, MMDuet2, and VideoLLM-online score zero on most subtasks; MiniCPM-o 4.5 reaches 0.12, TimeChat-Online 4.18, JoyAI-VL-Interaction 4.91, and AURA — the highest-scoring of the group — only 7.22%. Switch to the open-ended setting and the picture drops further: Gemini-3.8 Flash falls to 29.19% and AURA to 4.01%. The paper's read is blunt: these models are not bad at being proactive in general, they are bad at asking.\n\n## The real bottleneck is when to ask\n\nTwo sets of ablations separate the two dimensions of asking.\n\nOne ablation is an oracle experiment. On the subtask about missing key visual evidence, telling the model the correct answer to \"when should I ask\" alone lifts Ask-F1 from 37.17% to 58.30%, a gain of 21.13 points. Supplying both \"when to ask\" and \"what to ask\" raises overall Ask-F1 by 43.83% per the paper. The missing piece is timing, not wording.\n\nThe second ablation varies the ask budget. Letting the model ask more often — sweeping the budget from 1 up to 16 — lifts the score but saturates around 52.35%. More opportunities to ask do not substitute for knowing that you should ask.\n\nThere is also a detail worth keeping: before any of this, the authors ran a blur experiment to check whether asking is worth anything at all. With a blurred frame the model answers correctly 42.3% of the time. After the model requests and receives the right evidence, accuracy rises to 62.3%, close to the 63.1% it reaches with a fully clear input. Feed in mismatched evidence instead and the benefit largely disappears. What helps is not seeing more frames — it is obtaining information that matches the current question.\n\nMemory structure gets its own accounting. With StreamHarness, a hierarchical memory framework that keeps the most recent 10 frames at 1 fps, summarizes every 10 frames into one mid-term memory, consolidates every four mid-term memories into one long-term block, and retains up to eight long-term blocks, Ask-F1 is 52.18%. Dropping to a sliding window that keeps only the last 10 frames gives 41.70%, uniform sampling of 128 frames gives 26.72%, and 64 frames gives 16.99%.\n\n## What the benchmark is pointing at\n\nAskStream deliberately makes the task hard: every model is evaluated under conditions where it either cannot see clearly or cannot tell what is being asked, and the model gets one action opportunity per second with a tolerance of plus or minus one second. The paper also lists four limitations of its own — at most one ASK action per turn, coverage limited to streaming video understanding without embodied agents, every interaction initiated by an explicit user query, and real uncertainty that plays out over longer and messier time horizons.\n\nBut the design points at a stage the industry keeps skipping. Almost no model handles \"knowing when to ask\" well: even a strong offline model clears only a bit more than half of the bar, and the streaming models built for real-time interaction mostly score zero.\n\nThe next phase of streaming video looks less about seeing more clearly and more about knowing that you cannot see clearly. AskStream turns that ability into a 52.18% threshold, and mainstream streaming models are still far below it. So the next time a leaderboard like this refreshes, the number worth watching may not be whose score climbed — it may be which model dares to open its mouth and ask at the right moment.","askstream-proactive-asking-streaming-video","2026-11-02T01:35:00Z","2026-11-02T01:15:42.461609Z","2026-11-02T01:15:42.461622Z",true,"agent",212,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"950e298b-7e72-443b-939e-ef8ba07ee677","多模态社工攻破7个智能体:验证越多被骗越狠","multimodal-sea-social-engineering-agent-security-benchmark","2026-11-08T09:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"bff1ec03-4e40-4d63-be35-82da33948969","图像生成的长上下文考试:32张参考图拼一张图","longbanana-multi-reference-image-synthesis-benchmark","2026-11-05T17:05:54+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"fd0e593c-bb2c-44f9-8913-16f8dcef6999","苹果8B智能体:真实iOS上追平Gemini-3.1-Pro","apple-online-rl-ui-agent-live-ios","2026-10-31T11:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"9109e6b0-bf85-472d-abf6-ba8e31dc17c5","Co-Design 基准:七成准则过,26.4% 完整交付","co-design-bench-agent-delivery-gap","2026-10-30T05:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"db8f2b94-8565-49fe-9fd8-08310f441161","HiSTEVE 1.0:像人一样玩我的世界,只靠屏幕和键鼠","histeve-hierarchical-vla-minecraft-agent","2026-10-26T15:13:44+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5be712c9-ff84-4e5a-aefa-673f880c8ed1","视频深度研究报告基准:AI排版强内容弱,最高63.3分","omnidr-bench-video-deep-research","2026-10-25T19:15:23+00:00"]