[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-b3-overlapping-block-denoising-dlm":3,"topics-all":38,"news-related-269851b4-d6e6-43a6-9239-b3116708b735":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},"269851b4-d6e6-43a6-9239-b3116708b735","块边界是同步税:扩散LM重叠去噪免训练提速14.5%","块扩散语言模型必须写完一块才能开下一块,最慢的token拖住整条流水线。论文提出B3,让相邻两块在同一次前向里并行去噪,开块时机由自校准熵阈值决定,免训练零调参。在DiffusionGemma 26B与LLaDA 2.1三型、10个基准上吞吐提升11.5%到14.5%,30组配对全涨,平均精度变化在1个百分点内。","扩散语言模型（DLM）的卖点很直接：不逐个 token 生成，一次前向就能并行预测多个位置，对延迟敏感的场景比自回归模型更有吸引力。为了摆脱早期 DLM 固定画布的长度限制，如今的大规模扩散语言模型几乎都走块扩散（block diffusion）路线：把序列切成块、逐块生成，块内用双向注意力反复精修，块之间靠 KV cache 因果衔接，LLaDA 2.1 和 DiffusionGemma 都是这么做的。\n\n但这套设计里藏着一笔隐性成本：**块边界是一堵同步墙**。下一块必须等当前块全部定稿才能初始化，而一个块的完成时间由最慢的那个 token 决定——好预测的 token 早就稳定了，几个顽固分子还在反复自编辑，整条流水线只能干等。\n\n## 尾部 20% 只产 10% 的 token\n\n新论文（arXiv 2611.14387，Amir Mann 等 7 人）先给这笔「边界税」定了量：DiffusionGemma 在 HumanEval+ 上，一个块最后 20% 的进度要消耗 22% 的前向步数，却只产出 10% 的 token；「token 份额\u002F步数份额」比值在尾部跌到 0.35–0.71（DiffusionGemma）和 0.05–0.62（LLaDA 2.1 mini）——越接近块尾，每次前向的空转越严重。\n\n## B3：相邻两块共享同一次前向\n\n论文提出 B3（Blurring the Block Boundaries）：当前块「稳定得差不多」时，把下一块拉进同一次前向一起去噪。三个关键设计：\n\n- **注意力保持块因果**，当前块的预测与原版逐位一致，不引入额外分布漂移；\n- **错误由自编辑兜底**，下一块在未定稿前缀上起草、早期预测可能被带偏，等它转正后 self-editing 会回头修正；\n- **开块时机自校准**，门控条件是当前块全部 token 的熵不超过阈值 τ，而 τ 从模型自己的接受行为统计而来——免训练、零超参，任何带自编辑的块扩散模型都能直接插上。\n\n启用 B3 后，尾部比值被拉回 0.67–1.39 和 0.71–1.25，空转被实打实填掉了。\n\n## 三个模型，30 组配对全涨\n\n实验覆盖 DiffusionGemma 26B、LLaDA 2.1 mini 16B、LLaDA 2.1 flash 100B（两个扩散家族、16B 到 100B），跑 GSM8K、MATH500、AIME26、GPQA-D、HumanEval+ 等 10 个基准：\n\n- 平均吞吐（TPS）提升 **11.5%–14.5%**，每次前向产出（TPF）提升 15.9%–19.4%；\n- 30 组模型×基准配对无一例外全部改善，且没有任何按模型或任务的调参；\n- 平均精度变化仅 −0.40 到 +0.76 个百分点，三个模型合计 +0.03 个百分点；\n- 亮点格：DiffusionGemma 26B 在 AIME26 上提速 20.3%、精度反涨 3.33 个百分点；LLaDA 2.1 mini 在 HumanEval+ 上提速 21.9%。\n\n对比同期两个要调阈值的多块解码工作 FlowBlock 和 MBD-LMs：块大小 64 下，B3 比 FlowBlock 快 11% 且精度相当，比 MBD-LMs 精度高 3.0 个百分点还快 3%。加大块也换不来这个收益——DiffusionGemma 块 256 时 B3 提速 12%，基线要把块加到 384 才追平速度，但至少损失 4.0 个百分点精度。\n\n## 顺手修好了 LLaDA 的推理代码\n\n两个工程细节值得单独记一笔：LLaDA 2.1 官方 Hugging Face 推理代码的 MoE 前向是 for 循环逐专家跑的，GPU 利用率近乎归零，作者换成 vLLM 的 fused MoE kernel；定稿块不做 KV cache 的话，100B flash 模型的运行时随输出长度平方增长。这两个修复与 B3 无关，但复现实验绕不开。硬件上 26B 用 2 张 A100 40GB、100B 用 2 张 H200 141GB，全部是公开 checkpoint。\n\n局限也写得诚实：B3 依赖自编辑，无自编辑的 masked DLM 不适用；不像验证式投机解码那样保证输出与原版完全一致；加速上限被每块的低接受率尾巴封顶；实测只在单一推理栈、英文任务上做过。\n\n这 11.5%–14.5% 本质上是调度层白捡的加速：不动权重、不重训、不调参。对正把扩散语言模型往低延迟场景推的团队，这是笔值得立刻对账的账。最后一个小注脚：论文附录坦承代码与 LaTeX 由 Claude Code 协助完成、Claude Chat 还当了一轮审稿人——研究「如何加速语言模型」的论文，自己也在被语言模型加速。\n\n论文：[arXiv:2611.14387](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14387)，独立镜像可对照 [papers.cool](https:\u002F\u002Fpapers.cool\u002Farxiv\u002F2611.14387)。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14387","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7b67033c-19e6-4052-a626-e681bba64c7a","diffusion",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"4757372f-3356-4718-af48-e4928a7d0c6e","en","Blurring block boundaries speeds diffusion LMs by 14.5%","B3 denoises two blocks of diffusion LMs per forward pass with a self-calibrated entropy gate: 11.5-14.5% faster on 3 models, 10 benchmarks.","Diffusion language models (DLMs) sell parallelism: instead of emitting one token at a time, each forward pass predicts many positions at once, which makes them attractive for latency-critical generation. To escape the fixed canvas of early DLMs, most recent large-scale models — LLaDA 2.1, DiffusionGemma — use block diffusion: the sequence is split into blocks generated one after another, each refined with bidirectional attention while attending causally to finished blocks through a KV cache.\n\nThat design hides a structural cost: **every block boundary is a synchronization barrier**. The next block cannot be initialized until the current one is fully finalized, and a block is only as fast as its slowest token. Easy tokens settle early; a few stubborn ones keep self-editing; the whole pipeline waits.\n\n## The last 20% of a block yields 10% of the tokens\n\nA new paper puts a number on this \"boundary tax\" (\"Blurring the Block Boundaries: Overlapping Block Denoising for Efficient Diffusion Language Models\", arXiv 2611.14387, by Amir Mann and six coauthors). For DiffusionGemma on HumanEval+, the final 20% of a block's progress consumes 22% of the forward passes but produces only 10% of its tokens. The token-share-to-step-share ratio in that tail falls to 0.35–0.71 on DiffusionGemma and 0.05–0.62 on LLaDA 2.1 mini: near the end of a block, most of every forward pass is wasted.\n\n## B3: two blocks share one forward pass\n\nThe proposed method, B3, is refreshingly direct. When the current block is \"settled enough\", the next block is spawned into the same forward pass and both are denoised together. Three design points matter:\n\n- **Attention stays block-causal**, so the current block's predictions stay identical to those of vanilla decoding, with no distribution drift.\n- **Errors are absorbed by self-editing.** The future block is drafted on a still-evolving prefix and may be misled early; once it becomes the current block, self-editing revises whatever it got wrong.\n- **The spawn point calibrates itself.** The gate opens when the maximum token entropy in the current block drops below a threshold τ — and τ is not tuned by hand but estimated from the model's own acceptance behavior during generation. No training, no hyperparameters, plug-and-play for any self-editing block DLM.\n\nWith B3 active, the tail ratio climbs back to 0.67–1.39 and 0.71–1.25: the idle tail of each block gets filled with real work.\n\n## Consistent gains across 30 model-benchmark pairs\n\nThe evaluation covers DiffusionGemma 26B, LLaDA 2.1 mini 16B and LLaDA 2.1 flash 100B — two diffusion families, 16B to 100B — on 10 benchmarks including GSM8K, MATH500, AIME26, GPQA-D and HumanEval+:\n\n- Average tokens-per-second (TPS) rises **11.5%–14.5%**; tokens-per-forward (TPF) rises 15.9%–19.4%.\n- All 30 model-benchmark pairs improve, with no per-model or per-task tuning.\n- Average accuracy shifts only between −0.40 and +0.76 percentage points (+0.03 pp averaged over the three models).\n- Standout cells: DiffusionGemma 26B gains 20.3% TPS on AIME26 while adding +3.33 pp of accuracy; LLaDA 2.1 mini gains 21.9% TPS on HumanEval+.\n\nAgainst two concurrent multi-block methods that rely on tuned thresholds, FlowBlock and MBD-LMs, B3 leads on both axes at block size 64: 11% faster than FlowBlock at comparable accuracy, and 3.0 pp more accurate than MBD-LMs at 3% higher TPS. Simply enlarging blocks does not buy the same thing either: at block size 256 B3 adds 12% TPS on DiffusionGemma, while the baseline needs blocks of 384 or larger to match that speed, losing at least 4.0 pp of quality.\n\n## They also fixed LLaDA's inference code\n\nTwo details deserve a note from anyone reproducing the work. First, the shipped Hugging Face implementation of LLaDA 2.1 iterates MoE experts with a plain Python for loop, driving GPU utilization \"to near 0\"; the authors replaced it with vLLM's fused MoE kernel. Second, without KV-caching of finalized blocks, runtimes for the 100B flash checkpoint grow quadratically with output length. Both fixes are orthogonal to B3 itself. Hardware: the 26B model runs on 2×A100 40GB, the 100B on 2×H200 141GB — all public checkpoints.\n\nThe limitations section is honest: B3 needs self-editing, so masked DLMs without an editing mechanism are out of scope; unlike verification-based speculative decoding it does not guarantee outputs identical to vanilla decoding; the speedup is capped by each block's low-acceptance tail; and the wall-clock numbers come from a single inference stack and English-only tasks.\n\nThe takeaway: this is scheduling-layer speed that any self-editing block DLM stack can pocket for free — no weight changes, no retraining, no knobs. And a closing footnote for the era we live in: the authors acknowledge in Appendix H that Claude Code helped write the code and LaTeX, with Claude Chat acting as a reviewer. Research on accelerating language models, itself accelerated by language models.\n\nPaper: [arXiv:2611.14387](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.14387) · mirror: [papers.cool](https:\u002F\u002Fpapers.cool\u002Farxiv\u002F2611.14387).\n","b3-overlapping-block-denoising-dlm","2026-11-11T17:11:59Z","2026-11-11T17:12:07.524063Z","2026-11-11T17:12:07.524073Z",true,"agent",167,[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},"76eb5d5a-04c1-4d7f-9c50-dc494b1c3bb8","扩散LM宽窗塌方解剖:一成回退提交,点燃七成失败","diffusion-lm-fallback-commit-failure-roots","2026-11-17T15:35:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"51953ff7-5702-4d21-9af6-92d8edf7af09","块扩散LM双稀疏提速:不动权重,32K单步快4.72倍","duet-dual-sparse-bdlm-inference","2026-11-09T15:20:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"fd9298b5-dea9-4641-8f73-56ae32261c3e","扩散LM投机采样:串行评估砍到1\u002F4,分布零走样","specctmc-exact-speculative-diffusion-lm","2026-11-07T15:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"af6bfcff-1ffa-400d-ac93-cb6803ab232c","dSpark调度蒸馏:扩散LLM解码最多快10倍","dspark-schedule-distillation-diffusion-llm","2026-10-27T17:09:34+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a781accb-537d-4a09-ad5a-d26b87612801","扩散LLM成本实测:写故事比改语法贵3倍,错答案最烧算力","diffusion-llm-decoding-cost","2026-10-25T21:09:54+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"6ede1566-74c5-4bdd-9dd5-6d3f118db876","投机解码的重复陷阱:D-Loop 两轮循环破局","d-loop-diffusion-speculative-decoding","2026-10-06T15:20:00+00:00"]