[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-intreal-interleaved-generation-benchmark":3,"topics-all":38,"news-related-2f9ad534-b737-490e-b97e-bf01f2b6d26b":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},"2f9ad534-b737-490e-b97e-bf01f2b6d26b","交错生成新基准:闭源旗舰文本赢,图像一致性输给开源","IntReal 用 10 万条真实图文交错样本加七个维度评测统一多模态模型。闭源 NanoBanana2 文本逻辑性 97.49,远超开源 Emu3.5 的 75.17,但图像多样性 83.56、上下文一致性 76.70 反低于开源最优值。SFT 涨文本 29.04% 却让图像美学掉 5.65%,RL 才拉回均值。","刷模型榜单有个问题没人给干净答案：模型到底会不会\"图文交错地讲故事\"。10 月底挂上 arXiv 的论文（[arXiv:2610.38425](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38425)）把它拆成七个可测维度，结论反直觉——闭源旗舰在纯文本维度把开源甩开两倍以上，却在真正决定交错能力的两项指标上反被开源压过去。\n\n## 交错生成为什么一直难评\n\n\"交错生成\"指模型一次性吐出图文交替排列的连续内容，比如一份带配图的菜谱。模型得同时管住叙事推进、图像数量，以及每张图跟前文是否对得上。\n\n论文开篇说得很直白：统一多模态模型（UMM）已能联合生成文本和图像，但交错生成仍\"largely underexplored\"。卡点有两个——缺少被明确定义的任务和训练数据，以及缺少能分维度诊断的评测。\n\n数据侧空白更刺眼。作者从七个允许公开抓取和再分发的站点收集了百万量级原始语料，最终只留约 10%：长序列里任何一段有问题，整条样本就扔。\n\n## IntReal 的两件套：10 万条数据 + 600 条评测集\n\n保留下来的 IntReal-100K 有 100,000 条样本、538,452 张图、146.6M 个文本 token，平均每条 5.38 张图、1466.06 个 token。数据按文本与图像的协作方式分四类：Planning（决策规划）、Knowledge（知识讲解）、Tutorial（教程步骤）、Story（叙事）。\n\n七个来源里下厨房 34,771 条、维基百科 21,138 条、Wikivoyage 21,055 条，另有 Wikitravel 5,055、Curious George 7,298、Rabbids Invasion 3,949、The Land Before Time 6,734。Story 类平均每条 8.6 张图，交错序列最长；Planning 类只有 2.42 到 2.68 张。\n\n评测集从 100K 里分层抽样出 600 条，每类 150 条，分粗到细两档：整体档（holistic）让模型一次生成完整交错回复；细粒度档（fine-grained）固定前文图文上下文，只生成下一张图，用来定位哪步跑偏。代价是每个模型要生成 6,000 多张图文交错图。\n\n七个维度各配了\"专家模型\"：文本丰富度和文本逻辑性交给 LLM 专家 Gemma4，图像美学交给 IQA 模型 RALI，图像多样性用 DINOv3 算相邻图相似度，数量合规直接数张数，上下文一致性由 VLM 专家判断，主体一致性用 DINOv3 对照参考图。\n\n## 评测本身可不可信\n\n作者做了三组验证。指标之间不重复：任意两个维度的样本级 Spearman 相关绝对值都低于 0.42，最突出的一对是上下文一致性跟主体一致性。跟人对不对得上：五个主观维度上做了 750 次专家成对比较，自动评分与专家判断平均一致率 83.3%。换个裁判还稳不稳：把 Gemma4 换成同量级的开源 Qwen3.6-35B-A3B，模型排名的相关系数都超过 0.9（文本丰富度 0.905、文本逻辑性 0.929、上下文一致性 0.976）。\n\n## 跑分结果：闭源赢在文本，输在交错\n\n四个 UMM 家族被拉来对比：Bagel、Emu3.5、SenseNova-U1，以及论文标注为 Gemini3.1-Flash-Image 的闭源 API NanoBanana2。\n\n结果不是\"谁全面领先\"，而是能力画像错位。NanoBanana2 文本丰富度拿到 90.80、文本逻辑性 97.49，而开源里文本逻辑性排在头前的 Emu3.5 是 75.17。但轮到真正跟交错绑定的两项，它反而落后：图像多样性 83.56，低于开源最优的 93.31；上下文一致性 76.70，低于开源最优的 87.29。论文给出的相对差值是，文本丰富度比开源最优高 222.33%，图像多样性低 10.45%，上下文一致性低 12.13%。\n\n读表有个细节值得点破：论文用的\"开源最优\"是每项指标各自的最优包络，而非某一个统一模型，而这个包络里已包含在本数据集上专门微调过的 checkpoint。闭源 API 实际是跟一群开过小灶的开源模型比。\n\n## 后训练暴露的取舍\n\n作者还顺手验证了自家指标能不能当训练信号。用 IntReal-100K 对 SenseNova-U1-8B-MoT 做 SFT，文本丰富度涨 29.04%、文本逻辑性涨 6.68%、主体一致性涨 6.53%，但图像美学掉 5.65%、上下文一致性掉 4.89%，七项平均只涨 0.60%。这跟 UMM 里\"理解与生成存在内在优化冲突\"对上了。\n\n再上强化学习（Flow-GRPO，用 IQA、DINO 和 rubric 三类奖励，10K 条精标样本），图像美学从 -5.65% 收回 -1.34%，图像多样性转正到 +5.60%，均值涨到 +2.69%，在开源模型里排头前。\n\n作者从中得出的判断更值得记：单模态生成能力的提升可外溢到交错生成，未必需要每阶段都做联合优化。训练与评测严格不重叠，99.4K 训练样本，评测那 600 条剔除。\n\n## 值得泼的几盆冷水\n\n论文自己承认的局限：高质量真实交错数据要靠激进过滤换来，保留率偏低；为保持实验设置干净，没探索混用单模态数据后训练。\n\n但几处值得掂量。人类一致性平均 83.3%，意味着约六分之一的成对比较，自动评分和专家判断并不一致——这基准终究是模型评模型。闭源那侧更微妙：论文明确说，因为系统没公开，无法判断 NanoBanana2 的交错能力是原生支持还是套了 agentic pipeline，所以\"闭源输给开源\"要打折。素材许可也逐源处理：Story 类来自 Curious George、Rabbids Invasion、The Land Before Time 三部儿童动画，每张图单独授权。\n\n还有个工程现实：整套评测跑在一张 80GB 显存的 A100 或 H100 上。门槛不算高，但这解释了评测集为何只有 600 条——作者说再扩大会大幅推高成本。\n\n## 所以呢\n\n这份评测的价值不是又一张榜单，而是把\"交错能力\"从模糊印象拆成七个互不重叠、可复现、可当训练信号的刻度。文本强不等于交错强，单模态的领先在图文交替场景里会被原样抹平——这是最反常识的一条。\n\n10 月 30 日，作者已把数据集、榜单和微调后的模型同步放到 Hugging Face，代码仓库一并公开。想复现的可直接开跑，想蹭数据的也别客气——它本就是为训练准备的。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38425","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",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},"98fcec2e-a43a-44d2-a61f-a312351214ad","en","IntReal: Closed-Source Wins Text, Loses Interleaving","IntReal benchmarks unified multimodal models on 100K interleaved samples. NanoBanana2 wins text logic 97.49 versus 75.17, but trails on context fidelity.","Scrolling model leaderboards leaves one question unanswered: can the model actually interleave text and images into a coherent piece. A paper posted to arXiv at the end of October ([arXiv:2610.38425](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.38425)) breaks that into seven measurable dimensions, and the result is counterintuitive. The closed-source flagship wins big on pure text, yet on the two dimensions that define interleaving, it trails the open-source models.\n\n## Why interleaved generation has been hard to evaluate\n\nInterleaved generation means a model emits a continuous sequence where text and images alternate, like a recipe with a photo next to every step. This is not one image plus one caption: the model has to hold narrative progression, image count, and whether each image matches the preceding context all at once.\n\nThe paper states the situation plainly. Unified Multimodal Models (UMMs) can already jointly generate text and images, but interleaved generation remains \"largely underexplored.\" Two blockers are named: the lack of well-defined interleaved generation tasks with corresponding training data, and the absence of fine-grained evaluation protocols that can reliably diagnose model performance across capability dimensions.\n\nThe gap on the data side is starker. The authors collected a million-scale raw corpus of interleaved text-image content from seven sources that permit public collection and redistribution. Only about 10 percent survived: if any segment of a long interleaved sample is defective, the whole sample goes.\n\n## What IntReal ships: 100K samples plus a 600-sample suite\n\nThe surviving IntReal-100K contains 100,000 samples, 538,452 images, and 146.6M text tokens, averaging 5.38 images and 1466.06 text tokens per sample. The data splits into four categories by how text and images collaborate: Planning (decisions), Knowledge (explanation), Tutorial (ordered steps), and Story (narrative).\n\nThe seven sources contribute 34,771 samples from Xiachufang, 21,138 from Wikipedia, 21,055 from WikiVoyage, 5,055 from Wikitravel, 7,298 from Curious George, 3,949 from Rabbids Invasion, and 6,734 from The Land Before Time. Length distribution tells its own story: Story averages about 8.6 images per sample, the longest interleaved sequences, while Planning averages only 2.42 to 2.68, with text carrying decisions and images visualizing the spatial and temporal changes.\n\nThe evaluation set draws 600 samples by stratified sampling, 150 per category, in two granularities. Holistic evaluation has the model generate the complete interleaved response at once; fine-grained evaluation pins the preceding text-image context and asks only for the next image, so you can localize which step went wrong. The cost is steep: each model has to generate more than 6,000 images interleaved with text.\n\nEach of the seven dimensions gets a dedicated expert model. Text Richness and Text Logicality go to an LLM judge running Gemma4, Image Aesthetic to the RALI IQA model, Image Diversity to DINOv3 similarity between adjacent generated images, Quantity Compliance to plain image counting, Context Fidelity to a VLM judge, and Subject Consistency to DINOv3 comparison against reference images.\n\n## Can the evaluation itself be trusted\n\nThe authors ran three checks. Metrics do not overlap: the absolute pairwise sample-level Spearman correlation between any two dimensions stays below 0.42, with the most strongly correlated pair being Context Fidelity against Subject Consistency. It agrees with people: across five subjective dimensions they collected 750 expert pairwise comparisons, with automatic scores matching expert judgments 83.3 percent of the time on average. It survives a different judge: swapping Gemma4 for the similarly sized open-source Qwen3.6-35B-A3B keeps model rankings correlated above 0.9 (Text Richness 0.905, Text Logicality 0.929, Context Fidelity 0.976).\n\n## The results: closed source wins on text, loses on interleaving\n\nFour UMM families were brought in: Bagel, Emu3.5, SenseNova-U1, and the closed-source API NanoBanana2, which the paper labels Gemini3.1-Flash-Image.\n\nThe outcome is not one model dominating, it is capability profiles that do not line up. NanoBanana2 posts 90.80 on Text Richness and 97.49 on Text Logicality, while the leading open-source Text Logicality score, Emu3.5, reaches 75.17. But on the two dimensions tied to interleaving, it falls behind: Image Diversity 83.56 against an open-source high of 93.31, and Context Fidelity 76.70 against an open-source high of 87.29. The relative deltas in the table put Text Richness 222.33 percent above the open-source high, Image Diversity 10.45 percent below, and Context Fidelity 12.13 percent below.\n\nOne detail in reading that table is worth flagging. The open-source comparison is a per-metric envelope across checkpoints, not one single model, and that envelope already includes checkpoints specifically fine-tuned on this very dataset. The closed-source API is therefore measured against open-source models that had already studied for this test.\n\n## The trade-off post-training exposes\n\nThe authors also checked whether their own metrics can work as training signal. Fine-tuning SenseNova-U1-8B-MoT on IntReal-100K lifts Text Richness by 29.04 percent, Text Logicality by 6.68 percent, and Subject Consistency by 6.53 percent, but costs 5.65 percent on Image Aesthetic and 4.89 percent on Context Fidelity, leaving the seven-dimension mean up just 0.60 percent. That lines up with the widely reported optimization conflict between understanding and generation inside UMMs.\n\nReinforcement learning on top (Flow-GRPO with IQA, DINO, and rubric rewards over 10K carefully annotated samples) pulls Image Aesthetic back from -5.65 percent to -1.34 percent, turns Image Diversity positive at +5.60 percent, and lifts the mean to +2.69 percent, the top score among open-source models.\n\nThe broader conclusion from the authors matters more than the deltas: improving unimodal generation capability generalizes to interleaved generation, so interleaved generation does not require joint text-image training at every stage, and modality-specific optimization is a viable path. Training and evaluation are strictly disjoint: 99.4K training samples, with the 600 evaluation samples removed entirely.\n\n## A few cold-water caveats\n\nThe stated limitations in the paper: high-quality real interleaved data requires aggressive filtering, which keeps the retention rate low, and to keep the experimental setting clean the authors did not explore mixing in unimodal data for post-training, leaving synthetic data augmentation and mixed-data post-training for future work.\n\nSeveral things are worth weighing independently. Human agreement averages 83.3 percent, meaning roughly one in six pairwise comparisons has the automatic score disagreeing with the expert judgment, so this benchmark ultimately has models grading models. The closed-source side is more delicate: the paper explicitly notes that because the system is not publicly disclosed, the authors cannot determine whether the interleaved capability of NanoBanana2 is natively supported or implemented through an agentic pipeline, so reading \"closed source loses to open source\" needs a discount. Licensing is handled source by source: all Story material comes from Curious George, Rabbids Invasion, and The Land Before Time, with each image licensed individually, and user-contributed Xiachufang content requires attribution and is limited to non-commercial use.\n\nOne engineering reality: the whole evaluation runs on a single 80GB GPU such as an A100 80GB or H100 80GB. That is not a high bar, but it explains why the evaluation set stops at 600 samples, since the authors note that scaling further would substantially increase evaluation cost across multiple model families.\n\n## So what\n\nFor anyone building models, the value here is not another leaderboard. It is that \"interleaving ability\" gets decomposed from a vague impression into seven mutually distinct, reproducible dimensions that can even serve as training signal. Being strong at text is not the same as being strong at interleaving, and a lead in the unimodal modality gets erased in text-image alternating scenarios. That is the most counterintuitive finding in these results.\n\nOn October 30, the authors put the dataset, the leaderboard, and the fine-tuned models on Hugging Face simultaneously, with the code repository open alongside. Anyone reproducing this can start running; anyone wanting the training data is welcome to it, since it was prepared for training in the first place.","intreal-interleaved-generation-benchmark","2026-11-01T05:27:08Z","2026-11-01T05:27:30.037114Z","2026-11-01T05:27:30.037121Z",true,"agent",31,[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},"26801a16-3f57-4bd8-adad-3b49a1d28f08","OmniCompass基准:开源模型音视频理解仍落后","omnicompass-omni-modal-video-grounding-benchmark","2026-10-16T15:10:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"5d3c50e8-5087-43e9-a8f1-c64973f712c1","Qwen 拆掉 ASR 管道:音视频原生对话靠合成数据练成","qwen-omnivchat-native-audio-visual-dialogue","2026-09-21T15:14:15+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"ff0bc92a-295a-4707-be8d-76115fe9eeee","PerceptionBench 出炉:16 个前沿多模态模型,视觉感知无一及格","moonshot-perceptionbench-atomic-perception","2026-08-26T13:15:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"6f2c5b92-a390-4b36-9eff-534daf04f4aa","WorkFactory:先修任务再微调,879涨到1565","workfactory-workpilot-agent-training","2026-10-31T23:10:20+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"8db153fd-969d-4d3f-bd34-11304a3806d2","Ollama 官方库 7 个标签是坏的,文件本身却没坏","ollama-official-tags-non-functional-quantized-llm-audit","2026-10-31T03:15:00+00:00"]