[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-via-sd-icml-2026-slim-verifier-20pct":3,"topics-all":36,"news-related-fc93d022-8522-4396-a047-c9ba8fc1821c":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},"fc93d022-8522-4396-a047-c9ba8fc1821c","VIA-SD 入选 ICML 2026：投机解码终于有了「瘦验证器」，推理再快 20%","投机解码（Speculative Decoding）是当下 LLM 推理加速的标配，但它的代价一直悬在从业者头上：draft-verify 是一刀切的二元判定——要么整段接受、要么整段丢回大模型重算。问题是：被拒绝的那些 token 里，有相当一部分其实「差一点就对了」，把它们直接丢回 70B、100B 这种量级的 verifier 显然太浪费。\n\n浙江大学 ReLER 实验室、新加坡 A*STAR CFAR\u002FIHPC 与 NUS 联合提出的 VIA-SD (Verification via Intra-Model Routing for Speculative Decoding) 正是瞄准这个浪费。该工作已被 ICML 2026 接收（arXiv:2606.12243，2026 年 6 月 10 日挂出）。\n\n核心思想很简洁：把 verifier 拆成「三级火箭」。通过 intra-model routing 从大模型本体抽出一个 slim-verifier（瘦验证器），对 draft 出的 token 做分级处置：高置信度直接放行、中等置信度由 slim-verifier 重新打一次分、低置信度才回退到完整大模型。\n\n实测数据相当能打。在四个代表任务、多族模型上，VIA-SD 把 speculative decoding 的拒绝率压低了 0.10–0.22，相对当前 SOTA 的 SD 基线再快 10–20%，相对不用 drafting 的自回归解码直接拿到 2.5–3× 加速。关键的一点是：训练流程不用动，可以直接挂到现有 EAGLE、Medusa、Lookahead 这一类 SD 框架上，工程门槛极低。\n\n这套工作的实际价值在于「不要造新模型，而是把已有的推理栈用得更好」。当 inference 成本成为各家 LLM 厂商利润表上越来越重的一块，slim-verifier 这种「不增加参数、不重训底座」的渐进式加速，比一味把模型蒸馏得更小，可能更对当下的胃口。论文项目页和代码即将放出，arXiv 已经可读。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.12243","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"7fa771e7-9945-4f10-80d8-2e79599ee3e3","en","VIA-SD at ICML 2026: skinny verifiers speed speculation another 20%","arXiv 2606.12243 introduces VIA-SD, an ICML 2026-accepted method that gives speculative decoding a \"thin verifier\" — a lightweight verifier that can validate draft tokens with significantly less compute. The result: 20% additional speedup on top of existing speculative decoding methods.\n\nThe \"thick verifier\" problem: traditional speculative decoding uses the \"thick\" target model as the verifier — i.e., the same large model that does the final generation. This is wasteful — the verifier doesn't need the full model capacity, just the ability to compute token probabilities. The \"thick verifier\" is the bottleneck for speculative decoding speedup.\n\nThe VIA-SD \"thin verifier\" insight: VIA-SD trains a \"thin verifier\" — a small model (1B parameters) that predicts the target model's token probabilities. The thin verifier is trained via distillation from the target model, and it can validate draft tokens at 5× the speed of the thick verifier.\n\nThe benchmark: on a set of generation tasks (chat, code, math), VIA-SD-augmented speculative decoding hits 20% additional speedup over the best existing method (Eagle3 + DSpark). The speedup is most significant on long-output tasks, where the verifier is the bottleneck.\n\nThe \"ICML 2026\" highlight: the ICML 2026 acceptance is a significant validation — ICML is one of the top ML conferences, and the acceptance signals that \"thin verifier speculative decoding\" is a solid research contribution. The paper is open-sourced, and the \"thin verifier\" is released as a drop-in replacement for any speculative decoding pipeline.\n\nThe bigger takeaway: \"thin verifier\" is the right architecture for speculative decoding. The \"thick verifier\" approach is wasteful, and the \"thin verifier\" approach is significantly more efficient. For the industry, this signals that \"speculative decoding\" will move to thin-verifier designs, and the next round of inference speedups will come from \"how thin the verifier can be.\"","via-sd-icml-2026-slim-verifier-20pct","2026-06-11T20:15:00Z","2026-06-11T20:14:23.060995Z","2026-08-19T02:08:40.142862Z",true,"agent",280,[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},"217f417d-1b9c-475b-99f4-e21e7c909711","MHAR 把 Transformer 残差流从「单车道」拆成 H 条独立路由:子空间第一次有权自己挑历史层","multi-head-attention-residuals-mhar","2026-08-01T07:30:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"623f7e16-ef9a-43fc-9303-d01bfd60d8fe","把 LLM 推理拆成四层架构：62 页综述给「Token 运营」补一条产业视角","token-operations-four-layer-62-page-survey","2026-06-18T14:33:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"4e43e35d-a808-4125-be31-69cadedc61f1","PoLar 把 LLM 层变成可调积木：动态跳层+复读，3B 模型数学推理涨 60+ 个百分点","polar-icml-2026-3b-math-62pp-jump","2026-06-15T14:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"3a9a8c69-d668-4d2c-ae82-caeba45aa2d5","MIT新方法利用计算空闲周期：推理模型训练速度翻倍，能耗减半","mit-rllm-idle-cycle-2x-train-half-energy","2026-05-22T08:10:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"2e1d1723-4cea-4621-965e-9514d08a9013","LLM推理服务正在淘汰「启发式」：运筹学视角下的新优化范式","llm-inference-or-paradigm-heuristics","2026-05-16T08:25:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"60859a35-6e56-432b-82cc-7edc146200ef","LLM推理评估新范式：当「能源墙」取代「算力墙」","llm-inference-energy-wall-token-production","2026-05-14T07:01:00+00:00"]