[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hc-dlm-hierarchical-continuous-diffusion-lm":3,"topics-all":35,"news-related-2f43c66f-9994-4dc6-901c-3220143322b2":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"2f43c66f-9994-4dc6-901c-3220143322b2","HC-DLM：让连续潜变量替扩散语言模型打草稿","扩散语言模型并行解码时每个 token 只看自己的边缘分布，依赖被切断。UIUC 团队的 HC-DLM 把离散 token 与连续潜轨迹耦合进同一个去噪过程，目标由变分界导出，潜变量成为唯一持久生成状态。6M 参数下 Hard Sudoku 72.41%，LM1B 生成困惑度 75.5，优于同规模扩散基线。","扩散语言模型想抢自回归的饭碗，卡在一个结构性瓶颈上：并行解码时每个 token 只从自己的边缘分布独立采样，同批 token 之间的统计依赖被整体切断。数独这类全局约束任务直接暴露问题。连续扩散语言模型绕开了它——所有 token 在一个共享连续潜状态里联合规划——但去噪器只看得见潜状态，直到最后一步才解码，中间没有「必须是合法词元组合」的约束。UIUC 团队 10 月 1 日提交到 arXiv 的 HC-DLM 把两条路线缝在一起：离散 token 生成与连续潜轨迹耦合进同一个去噪过程，训练目标由 token 似然的变分界导出。\n\n## 潜变量当唯一草稿纸，token 每步纠偏\n\nHC-DLM 让潜变量成为唯一的持久生成状态：每步先从潜状态 x_t 读出 token 草稿 k_t，草稿再反过来指导下一次潜状态更新。与把连续上下文挂在自包含离散链上的混合方法不同，这里离散链本身不存在——token 只是潜状态的读出物，却承担「锚」的角色。训练端一个变分界拆三项：token 读出的重构项、token 引导的流匹配项、防编码器坍缩的熵正则项，端到端训练。\n\n## 跑分：同规模扩散阵营里的最优解\n\n项目页对比四类设计：离散扩散（MDM、LLaDA 类）一步内 token 相互独立；连续扩散（Plaid 类）只在 t=0 解码；混合方法（CCDD）靠条件耦合。HC-DLM 两头都占。数字（项目页口径）：\n\n- **Hard Sudoku**（6M 参数）：72.41%，比 CCDD 高 1.68 个百分点，比 MDM（top-prob. margin）高 22.53 个百分点；7 倍大的自回归模型只有 32.57%。\n- **Countdown CD5**（6M 参数）：37.52%，比 CCDD 高 12.17 个百分点，且领先随规划深度拉大——CD4 只领先 3.2。\n- **LM1B 生成困惑度**（118M 参数）：75.5，低于 Plaid 的 77.3、LangFlow 的 92.2、MDM 的 103.9。\n\n冷水也在同一页：Easy Sudoku 上 CCDD 94.65% 仍略高于 HC-DLM 的 94.21%；LM1B 上自回归 Transformer（108M）的 66.7 仍低于所有扩散方案。「领先」只在扩散阵营内部成立，自回归的城墙还没破口。\n\n## 消融：两层缺一不可\n\n去掉 token 引导退化为纯潜扩散，Hard Sudoku 从 72.41% 掉到 24.74%；去掉潜变量退化为纯离散链，掉到 49.88%。前向噪声同样关键：吸收态噪声最好 48.25%，均匀态加并行更新才到 72.41%——均匀腐蚀的序列仍是完整假设，mask 不携带支架信息。两条过程曲线值得看：中间预测准确率在 t\u002FT=0.6 前爬过 90%（无引导版本封顶 57%）；去噪步数从 100 砍到 10，HC-DLM 稳在 52%-76%，MDM 从 56% 跌到 20%——少步数场景正是扩散对自回归的速度优势所在，这条曲线直接关系到路线的实用价值。\n\n## 所以呢\n\n这一年扩散语言模型从「能不能生成」卷到「并行解码的依赖怎么保」：Flash-dLLM 做免训加速、dQwen3.5 接适配断点，HC-DLM 则重构生成状态本身。对做推理服务的人，价值不在 6M 玩具规模，而在验证了联合规划与逐步锚定可以共存、少步数下不塌。6M\u002F118M 的结论能否推到大模型等后续工作回答，但「潜变量打草稿、token 每步纠偏」这个范式值得进论文清单。\n\n参考：arXiv:2610.02193（hc-dlm.github.io）","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02193","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"84a98429-72c1-42e4-97d7-e23e8c4214c6","en","HC-DLM: Continuous Latent Drafts for Diffusion LMs","HC-DLM couples tokens with a continuous latent in one denoising process, anchoring parallel decoding: 72.41% Hard Sudoku at 6M params, LM1B PPL 75.5.","Diffusion language models want to challenge autoregressive generation, but they are stuck on a structural bottleneck: when decoding in parallel, each token is sampled independently from its own marginal distribution, severing the statistical dependencies among tokens decoded together. On a globally constrained task like Sudoku, two cells in the same row have no idea the other exists, so accuracy stalls. Continuous diffusion LMs sidestep this by planning all tokens jointly in a shared continuous latent — but their denoiser only sees the latent itself, and nothing ties it to a valid token configuration until the very last step. HC-DLM (Hierarchical Continuous Diffusion Language Models), submitted to arXiv on October 1 by a UIUC team, stitches the two routes together: discrete token generation and a continuous latent trajectory are coupled into a single denoising process, with a training objective derived from a variational bound on token likelihood.\n\n## The latent as the only persistent draft, tokens correct it every step\n\nThe key design makes the latent the only persistent generative state: at every step the model reads a token draft k_t out of the latent x_t, and that draft guides the next latent update. Unlike recent hybrid methods that attach continuous context to a self-contained discrete chain, there is no separate discrete chain here — tokens are readouts of the latent, yet they serve as its anchor. On the training side, one variational bound splits into three terms: a reconstruction term for token readout, a token-guided flow-matching term for the latent denoiser, and an entropy regularizer that keeps the encoder from collapsing, all trained end to end.\n\n## Numbers: the best in the diffusion camp at matched scale\n\nThe project page lines up four design families: discrete diffusion (MDM, LLaDA-style) leaves tokens independent within a step; continuous diffusion (Diffusion-LM, Plaid) decodes only at t = 0; hybrids (CCDD) couple via conditioning. HC-DLM keeps both strengths — tokens are planned jointly in the latent and corrected by tokens at each step. Per the project page:\n\n- **Hard Sudoku** (6M params): HC-DLM reaches 72.41%, +1.68 points over CCDD and +22.53 over MDM (top-prob. margin); a 7x larger autoregressive model manages only 32.57% here.\n- **Countdown CD5** (6M params): 37.52%, +12.17 points over CCDD; the lead widens with planning depth — +3.2 on CD4, +12.17 on CD5.\n- **LM1B generative perplexity** (118M params): 75.5, below Plaid's 77.3, LangFlow's 92.2, and MDM's 103.9.\n\nThe cold water is on the same page: on Easy Sudoku, CCDD's 94.65% still edges out HC-DLM's 94.21%; on LM1B, an autoregressive Transformer (108M) at 66.7 remains clearly below every diffusion option. HC-DLM's \"lead\" holds only within the diffusion camp — the autoregressive wall is yet to crack.\n\n## Ablations: neither level works alone\n\nRemove token guidance and the model degenerates into a plain latent diffusion model: Hard Sudoku drops from 72.41% to 24.74%. Remove the latent and it becomes a purely discrete chain: 49.88%. The forward-noise choice matters too: absorbing-state (mask-style) noise tops out at 48.25%, while uniform-state corruption with parallel updates reaches 72.41% — a uniformly corrupted sequence is still a complete hypothesis, whereas masks carry no scaffold information. Two process curves deserve attention: the accuracy of intermediate predictions climbs past 90% before t\u002FT = 0.6 (the no-guidance variant plateaus near 57%); and when denoising steps shrink from 100 to 10, HC-DLM holds the 52%-76% range while MDM falls from 56% to 20% — the few-step regime is exactly where diffusion's speed advantage over autoregressive decoding lives, so this curve is what ties the method to practical value.\n\n## So what\n\nOver the past year the diffusion-LM conversation has moved from \"can it generate\" to \"how to preserve dependencies under parallel decoding\" — Flash-dLLM attacked it with training-free acceleration, dQwen3.5 with adapter breakpoints, and HC-DLM restructures the generative state itself. For inference teams, the value is not the 6M toy scale but the validated idea that joint planning and step-wise anchoring can coexist without collapsing at few steps. Whether conclusions at 6M\u002F118M transfer to frontier scale is left to future work — but the paradigm of \"the latent drafts, tokens correct every step\" earns a slot on your reading list.\n\nReference: [arXiv:2610.02193](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02193) · [Project page](https:\u002F\u002Fhc-dlm.github.io\u002F)","hc-dlm-hierarchical-continuous-diffusion-lm","2026-10-02T17:11:05Z","2026-10-02T17:11:10.658989Z","2026-10-02T17:11:10.658996Z",true,"agent",350,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"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":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"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":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"a78b61ba-6899-4b4f-86b0-8d6c1753308b","亚马逊 ALoDLM:扩散语言模型学会按难度分配算力","alodlm-adaptive-looped-diffusion","2026-10-07T13:15:53+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"c448a366-372a-47dc-936f-8ea89103d238","Pivot-SD:200道题让扩散LLM数学涨6分","pivot-sd-diffusion-lm-self-distillation","2026-10-05T17:09:21+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"4899d809-5d52-454e-8c23-115f597e82d4","离散扩散模型终于被「拉直」:22 位作者把 Tokenization、Masking、Score 三条路线焊进同一框架","discrete-diffusion-unified","2026-07-18T02:10:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"bce625bf-8835-47d1-a12e-bf0cc111b905","dOPSD：让扩散 LLM 用「自身去噪轨迹」当老师，Dream 与 LLaDA 数学、代码双涨","dopsd-diffusion-llm-self-distillation","2026-07-07T06:05:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"d6afa8e3-b342-41da-839b-840c02c42cc8","ICML 2026 杰出论文砸场子:扩散语言模型的「任意顺序」是个陷阱","icml-2026-flexibility-trap","2026-07-06T10:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"79c21822-4e80-43f7-afea-baa14af4ba0a","Subliminal Clocks: 扩散语言模型里那块\"潜时钟\"被找到了","subliminal-clocks-dlm-latent","2026-07-06T06:00:00+00:00"]