[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-sumi-tohoku-udlm-7b-1-5t-native":3,"topics-all":36,"news-related-2c7dc4a2-6a5d-4cd3-837e-4486628c39a0":55},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"2c7dc4a2-6a5d-4cd3-837e-4486628c39a0","Sumi 把 UDLM 从零拉到 7B\u002F1.5T：扩散语言模型终于有了「原生参考点」","扩散语言模型（Diffusion LM）过去两年一直是「下一代 LLM 范式」的热门候选，但 Uniform Diffusion Language Model（UDLM）这条更灵活的技术路线（任意 token 在任意步可被覆盖）至今缺乏从零训练的大规模开源参照系。Google 的 DiffusionGemma 26B-A4B 是从预训练 AR 模型改造而来，Inception Labs 的 Mercury 2 走商业化推理路径，腾讯 WeDLM、EPIC、d3LLM 等则聚焦推理加速——「原生 UDLM」这一定位始终空缺。日本东北大学（Jun Suzuki、Keisuke Sakaguchi 团队）6 月 17 日发布的 Sumi 论文填补了这一空缺。\n\nSumi 是 7B 参数、1.5T tokens 训练的原生 UDLM，基于 GIDD 框架的 SNR 重参数化版本（连续 log-SNR λ∈[−9,9]），36 层 LLaMA 风格双向 Transformer + OLMo 3 tokenizer（词表 100,278），288 张 H100 跑 43,308 GPU-hours。预训练 1.3T tokens 后接两阶段 mid-training 共 250B tokens，数据全部来自公开语料并完整记录配比。\n\n关键性能：在 MMLU（51.1）、GSM8K（32.8）、HumanEval（22.6）、MBPP（26.6）、TruthfulQA（46.6）、RACE（41.4）上达到或超过同 token 预算的 Llama 2-7B、Falcon-7B、OLMo-7B；但 PIQA（66.4）、HellaSwag（60.0）、WinoGrande（60.0）明显落后，作者归因为教育数据过重——这是已知的 quality filtering 副作用，并非架构问题。\n\nSumi 真正深层的贡献在 §4 探索性实验，对 UDLM 三条乐观叙事做了冷静的边界测试：(1) 可用 canvas 长度严格受训练长度限制（1184 \u002F 4864），过短生成显著劣化；(2) confidence 采样器自动诱导出 task-shaped 提交顺序，让顺序无关的模型获得结构化能力，并在编程任务上支持 k=4 tokens\u002Fstep 并行解码（多步算术任务仍敏感）；(3) 显式 revision 预算不带来 self-correction：58-100% revision 步会覆盖已提交 token，但净变化 \u003C 1%，多为 A→B→A 往返——简单堆算力换不来自我修正。\n\n对从业者的判断：Sumi 第一次让社区能严肃研究「从零训练的 UDLM 在数据富集区究竟如何」，提供干净的 scaling reference point。但它也明确划出 UDLM 的能力边界——灵活性不是免费午餐，commitment order、canvas 长度、sampler 选择共同决定实际质量。下一阶段关键问题不是「扩散能否替代自回归」，而是「在哪些任务上，UDLM 的灵活性值得放弃 AR 的成熟工程栈」。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.19005","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"dfb12944-4c26-49a4-ac68-73b642a3312b","en","Sumi trains UDLM from scratch: 7B params, 1.5T tokens","arXiv 2606.19005 introduces Sumi, a 7B-parameter diffusion language model trained on 1.5T tokens. The standout: Sumi is the largest open diffusion LLM with a \"native reference point\" — i.e., it can be used as a starting point for downstream fine-tuning, evaluation, and deployment, just like an AR base model.\n\nThe \"scaling diffusion LLMs\" challenge: diffusion LLMs have historically been trained at small scales (1-3B parameters) and on limited data (100-500B tokens). The community lacked a \"base model\" comparable to Llama-3-7B or Qwen2.5-7B — a model that downstream developers can build on.\n\nThe Sumi fix: train a 7B diffusion LLM on 1.5T tokens (the same scale as Llama-3-7B), using a \"masked diffusion\" architecture with a novel \"noise schedule\" that improves training stability. The model is open-sourced with full training code and data pipeline.\n\nThe benchmark: Sumi-7B hits parity with Qwen2.5-7B on MMLU (within 1 point) and beats it on several code-generation tasks. The model is fully open-sourced, including the training data, training code, and evaluation scripts.\n\nThe bigger takeaway: \"open diffusion LLM base model\" is a critical missing piece. Sumi fills that gap, and the open-source release will accelerate the diffusion LLM research community. For the industry, this means diffusion LLMs are now a viable alternative to AR models for production deployments, with comparable quality and the additional benefits of diffusion (parallel generation, controllable generation, iterative refinement).","sumi-tohoku-udlm-7b-1-5t-native","2026-06-21T02:01:00Z","2026-06-21T02:12:30.605211Z","2026-08-19T02:08:40.142862Z",true,"agent",143,[37,46],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"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":47,"tag_slug":47,"title_zh":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56},[57,62,67,72,77,82],{"id":58,"title":59,"news_slug":60,"published_at":61},"4455b8ee-eab9-463b-9934-f1df4b1b4fb3","扩散语言模型的适配断点被接上:dQwen3.5 只花一半 token","dqwen3-5-hybrid-attention-diffusion-language-models","2026-09-18T19:20:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"8def771a-d936-4859-930d-02c3011dc55c","LimiX-2 开源：一个模型吃下分类回归插补，表格三榜登顶","limix-2-tabular-foundation-model","2026-09-17T21:09:27+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"176b4807-da61-479f-a514-9381cd13319e","SP3O:3 个锚点修复 PPO critic 的平坦化","sp3o-sparse-critic-supervision","2026-09-17T17:10:01+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"2e27016d-b90e-45c7-825a-41fd1e435c80","JHU 新研究:组合持续学习机制,百任务记忆留存从 1.2% 提到 34.9%","compose-cl-long-horizon-memorization","2026-09-16T15:10:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"4c4a2a9e-f69b-4985-bd42-97ab2ef4e2ac","Spark-X2.5-4B 开源:4B 跑 1M 上下文,22 项基准打 9B 级 Qwen3.5","spark-x2-5-4b-apache-open-source","2026-09-16T01:30:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"2731ed1c-17c3-4d85-9174-983cf50743e3","地铁售票机上的 AI 大考:2.6GB 端侧模型 91.32 分超 GPT-5.6,规则基线也拿 84.6","metrollm-bench-transit-kiosk-llm","2026-09-12T23:08:18+00:00"]