[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-when2think-difficulty-aware-length-control":3,"topics-all":35,"news-related-813ad679-51dd-43d7-afcc-0baf48d2ef5f":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},"813ad679-51dd-43d7-afcc-0baf48d2ef5f","When2Think:推理模型该想多久,先看题有多难","大型推理模型常把简单题想复杂、难题又想不够。成均馆大学与微软团队的 When2Think 后训练框架为每道题预计算难度与参考 token 预算,用 IDAC 奖励塑形控制推理深浅:AIME24 上 Pass@3 提升 10 个百分点,token 反而少 27.9%。","推理模型有个长期被默认容忍的怪癖:不管题目难易,思考预算几乎一样。9 月 17 日挂上 arXiv 的 When2Think 论文(成均馆大学与微软合作,通讯作者来自微软)给了一个量化注脚:在 MATH-500 最简单的 Level 1 题目上,1.5B 的 R1-Distill 平均要写 1,199 个 token 才肯作答,而这些题几乎不需要推理;到了难题一侧,思考又常常不够用。\n\n## 一刀切的「效率税」\n\n现有省 token 方案多走两条路:统一长度惩罚,或刚性的 Think\u002FNoThink 路由。论文把它们的共同代价称为「效率税」——简单题省了,难题精度跟着掉。在 AIME24 上,LC-R1 与 AdaptThink 都压低了推理成本,精度却分别下降 10.0 与 1.3 个百分点;AdaptThink 虽然只用了 5,806 token,Pass@3 跌到 44.7%,比基座模型还低。\n\n## 按题分配:IDAC 奖励塑形\n\nWhen2Think 把高效推理形式化成「按实例分配算力」:训练前先用参考策略对每道题预计算两个统计量——以参考准确率代理题目难度,以平均轨迹长度充当该题的 token 预算,再由 IDAC 项对超出参考长度的轨迹施加更强的奖励衰减。配合可验证奖励与批内标准化优势(BWS),整套 PPO 式优化不需要学习型 reward model、critic,也不需要在训练中在线查询参考模型。基座是 R1-distill-1.5B——DeepSeek-R1 蒸馏、Qwen2.5-Math-1.5B 的底子。\n\n## 结果:省得多,还涨分\n\nAIME24 上 Pass@3 从 46.0% 升到 56.0%,提升 10.0 个百分点,平均 token 少 3,959 个、降幅 27.9%;AIME25 拿到 40.0% Pass@3,按论文对比优于压缩类与纯路由类基线。行为层面的证据更细:Level 1 的平均 token 从 1,199 降到 619,精度保持 95.8%;Level 5 在保住精度的同时省下 2,276 个 token;面对 GSM-Plus 的对抗扰动样本,它反而主动多花 462 个 token,换来 6.3 个百分点的精度提升。Think 比例随难度从约 0.2 单调升到 0.7 以上——算力分配确实跟着难度走,而不是学会了新的一刀切。\n\n## 所以呢\n\n局限写得很直白:依赖可验证奖励,直接适用域是数学推理这类可判定对错的场景;难度估计靠离线参考统计,题目难度本身模糊时会失真;团队也承认更大模型自身推理效率更强,这种自适应控制的价值在 1.5B 尺度最显著。对工程侧的启发或许更实际——按题分配推理预算不需要在线额外查询,一次离线统计就够,推理成本敏感的 API 服务与端侧部署都可以直接抄这条思路。\n\n参考:arXiv:2609.19671 https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.19671 · 论文页 huggingface.co\u002Fpapers\u002F2609.19671","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.19671","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"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},"7b443ebf-94b5-470e-86f9-19aad492746a","en","When2Think: reasoning depth should follow problem difficulty","SKKU and Microsoft's When2Think shapes RL rewards with per-problem difficulty and token budgets: AIME24 Pass@3 +10 points on 27.9% fewer tokens.","Reasoning models carry a long-tolerated quirk: their thinking budget barely changes with problem difficulty. When2Think, posted to arXiv on Sep 17 by a Sungkyunkwan University and Microsoft team (the corresponding author is at Microsoft), offers a quantified footnote: on the easiest Level 1 problems of MATH-500, the 1.5B R1-Distill still writes an average of 1,199 tokens before answering, though those problems barely need any reasoning — while on hard problems the thinking often falls short.\n\n## The efficiency tax of one-size-fits-all\n\nExisting token-saving approaches mostly take one of two roads: uniform length penalties, or rigid Think\u002FNoThink routing. The paper calls their shared cost an \"efficiency tax\" — easy instances get cheaper, and hard-instance accuracy drops along with it. On AIME24, LC-R1 and AdaptThink both cut inference cost, yet accuracy falls by 10.0 and 1.3 points respectively; AdaptThink spends only 5,806 tokens but its Pass@3 drops to 44.7%, below the backbone model.\n\n## Allocation per instance: IDAC reward shaping\n\nWhen2Think formalizes efficient reasoning as instance-adaptive computation allocation. Before training, a reference policy pre-computes two statistics for every problem: reference accuracy serving as a difficulty proxy, and average trajectory length serving as a per-problem token budget. An IDAC term then imposes stronger reward decay on trajectories that exceed the reference length. Combined with verifier-based rewards and batch-wise standardization (BWS), the PPO-style optimization needs no learned reward model, no critic, and no online reference-model queries during training. The backbone is R1-distill-1.5B — distilled from DeepSeek-R1 on a Qwen2.5-Math-1.5B base.\n\n## Results: fewer tokens, higher scores\n\nOn AIME24, Pass@3 rises from 46.0% to 56.0% — up 10.0 points — while average tokens drop by 3,959, a 27.9% reduction. On AIME25 it reaches 40.0% Pass@3, which the paper reports as beating compression-only and routing-only baselines. The behavioral evidence is finer-grained: on Level 1, average tokens fall from 1,199 to 619 with accuracy held at 95.8%; on Level 5, it saves 2,276 tokens while maintaining accuracy; on adversarially perturbed GSM-Plus instances it deliberately spends 462 more tokens and gains 6.3 points of accuracy. The Think ratio rises monotonically from about 0.2 at Level 1 to over 0.7 at Level 5 — computation genuinely follows difficulty rather than learning a new uniform cut.\n\n## So what\n\nThe limitations are stated plainly: the method depends on verifiable rewards, so its direct domain is math-style tasks with checkable answers; difficulty estimation relies on offline reference statistics and gets shaky when difficulty itself is ambiguous; and the team acknowledges that larger models are inherently more reasoning-efficient, so adaptive control matters most at the 1.5B scale. The engineering takeaway may be the most practical one — per-problem reasoning budgets need no online queries, just one offline statistics pass, which API services and on-device deployments sensitive to inference cost can adopt directly.\n\nReference: arXiv:2609.19671 https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.19671 · paper page huggingface.co\u002Fpapers\u002F2609.19671","when2think-difficulty-aware-length-control","2026-09-19T19:08:00Z","2026-09-19T19:08:02.738699Z","2026-09-19T19:08:02.738709Z",true,"agent",46,[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},"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":62,"title":63,"news_slug":64,"published_at":65},"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",{"id":67,"title":68,"news_slug":69,"published_at":70},"178aa5e5-2a4f-4a87-a97c-0da16295d96f","EMNLP 2026 OCGQuant:用通道配对治 NVFP4 陪葬误差,Qwen3-1.7B 接近 FP16","ocgquant-nvfp4-outlier-companion-grouping","2026-09-10T09:15:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"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":77,"title":78,"news_slug":79,"published_at":80},"c83af54b-79ed-445c-9482-07d98c26c36b","BeaconKV:长推理会回头看,只压最近窗口的 KV 缓存注定丢东西","beaconkv-beacon-query-kv-cache-compression","2026-09-09T11:25:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"0fe9ceb8-6411-4924-8e02-8cee3665fc6f","Cohere 开源 megakernel 推理引擎：单 CUDA 文件，H100 解码吃到 62% 带宽光速","cohere-megakernel-north-mini-code-h100","2026-09-08T21:13:46+00:00"]