[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-videojev-system-one-video-understanding":3,"topics-all":38,"news-related-eeb3f139-8df8-40d7-a1e6-82e49b668547":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},"eeb3f139-8df8-40d7-a1e6-82e49b668547","VideoJev:视频问答提速100倍,不生成只打分","MCG-NJU 发布 VideoJev,把 System One 范式带进视频:共享打分头取代自回归解码,统一时序定位与多选问答。4B\u002F35B-A3B 开源模型较 Qwen 底座 mIoU 提升 6.7\u002F10.8 个点,前向提速 2.8-3.1 倍;问答持平 Thinking 模式,提速超 100 倍。","让视频模型看一段视频再回答问题,今天的多模态大模型走的是「生成式」路线:逐 token 吐出答案,thinking 模式下还要先输出一长串推理。对坐在交互循环里的视觉智能体来说,每一次「接下来做什么」的判断都要等一次完整生成,延迟直接叠加进行动闭环。南京大学 MCG-NJU 团队发布的 VideoJev 给出了另一种答案:把视频理解改造成结构化决策任务,模型不再「说」出答案,而是直接「打分」。论文当日冲上 Hugging Face Daily Papers 榜首(175 upvotes),训练与评测代码连同两个开源 checkpoint 一并放出([arXiv:2610.18615](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.18615))。\n\n## 不生成,只打分\n\nVideoJev 的做法是在预训练多模态大模型上动手术:保留视觉与语言主干的理解能力,把自回归答案解码替换成一个共享打分头,对「时间区间」和「文本选项」直接打分。分数出来,时序定位给出事件片段,多选问答给出选项——两类任务共用同一个决策接口。这延续了今年持续升温的 System One「快思考」范式:此前微软 Decision-1、Cloudflare Clef、GLM-5.3-Flash 平替方案都在文本任务上验证过「打分快于生成」,VideoJev 把这条路线系统性搬进了视频模态。团队基于 Qwen3.5-4B 与 Qwen3.6-35B-A3B 两个底座,训练出 VideoJev-4B 与 VideoJev-35B-A3B。\n\n## 更准,也快得离谱\n\n三个时序定位基准上,VideoJev-4B 与 35B-A3B 较各自 Qwen 底座平均 mIoU 提升 6.7 与 10.8 个点,前向提速 2.8-3.1 倍——不是用精度换速度,而是两项同时改善。更夸张的是问答:四个视频 QA 基准上,两个模型性能与底座 Thinking 模式相当,前向提速超过 100 倍。对需要反复查询同一段视频的智能体来说,这是量级差异。\n\n## 快的另一半:视觉计算复用\n\n摘要里的关键设计是两种推理模式:Cached inference 让新问题到来时复用已算好的视觉表征,不必重看视频;Joint inference 把多个问题塞进一次前向统一打分。交互展开时,新目标与中间结果会持续触发对同一段视觉历史的查询——复用设计正是为这个场景准备的。\n\n## 开源交付\n\n仓库以 Apache 2.0 协议开源,26 个 commit、46 star。训练分两阶段:先 Head warmup(只训打分头,BCE+CE 损失),再 Joint 阶段联合训练语言主干(加入 duration-weighted SoftIoU),视觉编码器全程冻结,每阶段一个 epoch、全局 batch 64。训练数据用 TimeLens-100K 与 TimeLens2-93K 两个开放数据集,框架为 XTuner,评测固定 256K 上下文,两个 checkpoint 均为 BF16 Safetensors 格式,4B 版十行代码即可跑通 demo。\n\n## 所以呢\n\n视频是比文本更重的模态:每帧都是算力,自回归生成的每个 token 都在为「看一下再决定」的智能体计费。VideoJev 的信号不在某个单项跑分,而在接口层的统一——定位、问答、多查询复用共享同一套打分机制,复杂工作流可以直接组合这些结构化判断。当「快思考」范式完成从文本到视频的扩张,下一个问题也许是:还有哪些模态适合「只打分、不生成」?","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.18615","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"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},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"9442b042-3da2-4355-a197-909d1d0b946b","en","VideoJev scores video instead of generating: 100x faster QA","NJU VideoJev scores video directly instead of generating: mIoU +6.7\u002F+10.8 over Qwen backbones, QA matches Thinking mode at 100x speed. 4B\u002F35B open","Asking a video model to answer questions today means generation: tokens stream out one by one, and thinking modes prepend a long chain of reasoning. For visual agents sitting inside an interaction loop, every \"what should I do next\" decision waits on a full generation pass, and that latency compounds. VideoJev, released by the MCG-NJU team at Nanjing University, takes a different route: it casts video understanding as structured decision tasks where the model scores instead of speaks. The paper hit #1 on Hugging Face Daily Papers the same day (175 upvotes), with training\u002Fevaluation code and two open checkpoints ([arXiv:2610.18615](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.18615)).\n\n## Score, don't generate\n\nThe method surgically adapts pretrained multimodal LLMs: keep the vision and language backbones, replace autoregressive answer decoding with a shared head that scores temporal intervals and textual options against video evidence. The scores yield event spans for temporal grounding or answer choices for multiple-choice QA — two task families, one decision interface. This extends the System One \"fast thinking\" paradigm that has been heating up this year: Microsoft Decision-1, Cloudflare Clef, and a GLM-5.3-Flash-based approach already validated scoring-over-generation on text tasks; VideoJev moves the playbook to video. The team built VideoJev-4B and VideoJev-35B-A3B on Qwen3.5-4B and Qwen3.6-35B-A3B backbones.\n\n## More accurate, and absurdly faster\n\nAcross three temporal grounding benchmarks, VideoJev-4B and 35B-A3B improve mean mIoU over their respective Qwen backbones by 6.7 and 10.8 points, with measured inference speedups of 2.8-3.1x — accuracy and speed improve together, not traded off. Video QA is where it gets extreme: on four benchmarks, both models perform comparably to their backbones' Thinking mode while running over 100x faster on model forward passes. For agents that repeatedly query the same video, this is an order-of-magnitude difference.\n\n## The other half of fast: visual computation reuse\n\nTwo inference modes power the speedup: Cached inference reuses already-computed visual representations when new queries arrive, skipping a full re-watch; Joint inference packs multiple questions into a single forward pass. As an interaction unfolds, new goals and intermediate results keep prompting queries against the same visual history — the reuse design exists precisely for that loop.\n\n## Open delivery\n\nThe repo ships under Apache 2.0 with 26 commits and 46 stars. Training runs two stages: Head warmup (scoring head only, BCE+CE losses), then a Joint stage that trains the language backbone together with the head (adding duration-weighted SoftIoU), with the vision encoder frozen throughout; each stage runs one epoch at global batch 64. Training data comes from the open TimeLens-100K and TimeLens2-93K datasets, the framework is XTuner, and evaluation fixes a 256K context. Both checkpoints are BF16 Safetensors; the 4B demo runs in about ten lines of code.\n\n## So what\n\nVideo is a heavier modality than text: every frame costs compute, and every generated token bills the agent that just wants to look and decide. VideoJev's signal is not any single benchmark number but the interface-level unification — grounding, QA, and multi-query reuse share one scoring mechanism that complex workflows can compose directly. Now that fast-thinking has stretched from text into video, the next question might be: which other modalities are waiting to stop generating and start scoring?","videojev-system-one-video-understanding","2026-10-14T15:09:17Z","2026-10-14T15:09:26.957873Z","2026-10-14T15:09:26.957887Z",true,"agent",39,[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},"8c28aa2a-10f7-4c0c-b976-e5bf7e781eed","Qwen3.5-397B-A17B发布：千亿MoE架构实现8.6倍解码吞吐提升","qwen3-5-397b-a17b-8-6x-decoding","2026-05-24T10:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4f70d0fe-f0cb-443e-a4a4-27b9eb004441","1.3B参数多模态模型直接跑在手机上：MiniCPM-V 4.6开源，13亿参数覆盖iOS\u002F安卓\u002F鸿蒙","minicpm-v-4-6-1-3b-phone-multimodal","2026-05-17T13:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"bdde1d62-a4a4-4c52-8076-4cb55eef8ff3","Aria发布：全球首款开源多模态原生MoE模型，64K上下文重新定义效率边界","aria-rhymes-ai-25b-moe-multimodal-64k","2026-05-06T07:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"b26b4b63-5019-4344-a49e-95739f737376","NVIDIA Nemotron 3 Nano Omni：开源统一多模态模型能否颠覆AI Agent效率边界？","nvidia-nemotron-3-nano-omni-30b-moe-9x-throughput","2026-04-28T22:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"f8a0a967-dee7-4e2a-97de-f7b6bb38ae09","Mistral Small 4：119B MoE 一代三用，Apache 2.0 重新定义开源边界","mistral-small-4-119b-moe-6b-active-apache2","2026-04-25T23:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5d4acf64-67d8-404b-9d3f-cdf7761fb8ac","小鹏开源 GroundingPI:4B 视觉 grounding 模型自评超 GPT-6 Astra","groundingpi-4b-visual-grounding-foundation-model","2026-10-13T17:10:26+00:00"]