[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-entmtp-entropy-speculative-decoding":3,"topics-all":36,"news-related-a453eb28-7fb0-4e07-adc1-0d0575850758":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},"a453eb28-7fb0-4e07-adc1-0d0575850758","EntMTP 用熵信号给多 token 推测装上调速器：让 LLM 自适应匹配上下文可预测性","Multi-Token Prediction（MTP）头已经被 DeepSeek-V3、Llama-3 当成标配——它把训练数据密度提上去，还能直接挂成 self-speculative decoding 的草稿器。但现有实现有一个被默认的假设：树形注意力的拓扑在整段生成中是静态的，推测深度不会跟着上下文变。\n\n这与自然语言的熵分布天然不匹配。一段连贯叙事（低熵）值得把推测推到 4-5 步，草稿几乎都能被验证器接住；进入逻辑分支或代码边界（高熵），同样的深度会让验证计算白白浪费在大概率被拒的草稿上。\n\nCarrie Chen 等人的 EntMTP（arXiv:2606.27550） 给出优雅解法：把局部生成熵作为在线调度信号，在一组任务相关的 Pareto 最优树之间动态切换。它完全 training-free，把\"哪棵树适合当前上下文\"做成运行时决策——用 task-specific Pareto 树作为候选池，根据滑动窗口内的熵估计选择当下的拓扑深度。\n\n效果算不上惊艳但足够说明问题：在 Humaneval、ShareGPT、GSM8k、Litbench 四个基准上对 Hydra 稳定拿到 1.15× 加速，对 Medusa 峰值 1.36×。提速不算激进，但它没引入额外训练成本，也没改模型权重，可直接挂到任何已训练 MTP 头的生产模型上。\n\n这条路线真正值得关注的是它改变了\"speculative decoding 工程优化\"的主战场：之前大家比的是\"我能写出更复杂的树\"（JetSpec 的并行树草稿、DSpark 的半自回归调度），现在变成\"我能更聪明地选哪棵树\"。当加速比逼近硬件上限，软件层的下一个红利是调度智能，而不是更深的草稿。EntMTP 样本虽小，却提示 MTP 推理栈下一步的演化方向——context-aware 的运行时策略层，正在变成推理优化的新前哨。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.27550","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"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},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"7981509a-20f3-48f4-877a-53aef5e4b157","en","EntMTP: entropy speed control for multi-token prediction","Multi-Token Prediction (MTP) heads have become standard in DeepSeek-V3 and Llama-3 — they increase training-data density and can be plugged in as a self-speculative-decoding drafter. But today's implementations carry an implicit assumption: the tree-attention topology stays static throughout generation, and the speculation depth never adjusts to context.\n\nThat assumption sits awkwardly with the natural entropy distribution of language. A coherent passage (low entropy) is worth pushing 4-5 steps deep — almost every draft token will be accepted by the verifier; the moment you hit a logical branch or a code boundary (high entropy), the same depth wastes verification compute on drafts that are most likely to be rejected.\n\nEntMTP (arXiv:2606.27550) by Carrie Chen et al. offers an elegant fix: take local generation entropy as an online scheduling signal, and switch dynamically across a task-specific set of Pareto-optimal trees. It is fully training-free — the \"which tree fits the current context\" decision is made at runtime, with task-specific Pareto trees as the candidate pool and a sliding-window entropy estimate selecting the right topology depth for the moment.\n\nThe results are not dramatic but illustrative: on HumanEval, ShareGPT, GSM8k, and Litbench, EntMTP delivers a stable 1.15× speedup over Hydra, and peaks at 1.36× over Medusa. The speedup is modest, but the approach adds no training cost and changes no model weights — it can be dropped into any production model that already has a trained MTP head.\n\nThe truly interesting takeaway is that it shifts the main battlefield of \"speculative-decoding engineering optimization\" — from \"who can write a more complex tree\" (JetSpec's parallel tree drafts, DSpark's semi-autoregressive scheduling) to \"who can pick the right tree more smartly.\" As acceleration ratios approach hardware ceilings, the next software-side dividend is scheduling intelligence, not deeper drafts. The samples in EntMTP are small, but they point the way: the MTP inference stack's next evolution is a context-aware runtime policy layer — the new frontier of inference optimization.","entmtp-entropy-speculative-decoding","2026-06-29T12:21:51Z","2026-06-29T12:21:57.345252Z","2026-08-19T02:08:40.142862Z",true,"agent",228,[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},"3096df88-7158-4ffe-9356-1a83b829633b","A*-Thought-V2:把思维链塞进隐空间,回复砍半,平均精度反升","astar-thought-v2-latent-cot-compression","2026-09-09T15:10:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"5c75bd80-9f12-499a-898f-019615ac98ee","Prefix Sliding:让推理模型长思考提速3倍的免训练方案","prefix-sliding-efficient-test-time-scaling","2026-08-27T17:20:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"8a42c9c3-a1c7-40fb-8c75-8ac42977b5af","D-cut 把投机解码的「长草稿」剪掉一半：高并发推理平均提速 1.65×、MoE 跑出 3×","d-cut-speculative-draft-cut","2026-07-18T10:10:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"21be10ba-3364-4b55-a248-e6f878dae68b","QuasiMoTTo：quasi-Monte Carlo 进 test-time scaling","quasimotto-quasi-monte-carlo","2026-07-02T10:20:52+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"9a3cd449-e29a-4730-814b-f1be5c2685c6","复旦FFD让Flash Attention退役？11.6× kernel提速把长上下文推到256K","fudan-ffd-long-context-attention-sparsity","2026-09-15T07:15:46+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"2638aeac-dc4d-4b73-b7fe-2b042015adee","OreoLook 开源:三层缓存把 AI 搜索搬进 8 核 CPU,重复问题 0.1 毫秒出答案","oreolook-three-layer-cpu-cache","2026-09-10T23:08:36+00:00"]