[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-moe-router-reasoning-behavior-steering":3,"topics-all":38,"news-related-d456840a-1a1d-412d-90ac-73ca5a5bbe10":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},"d456840a-1a1d-412d-90ac-73ca5a5bbe10","MoE 路由器读出推理状态:砍规划省三成 token","UCLA 团队发现,MoE 推理模型的路由器 logit 能直接读出六类推理行为,还能在推理时不改任何权重地改写路由,四个模型四个基准共十六组设置里,生成 token 中位数省下 30.4%、最高 63.3%,准确率最多只掉 0.5 个百分点。更反直觉的是该砍规划而非反思——压反思专家反而让轨迹变长 5%。","推理模型\"想太多\"是个老毛病：答案早算出来了，还在反复验算。常规解法是加一句系统提示词——\"简洁点，别重复检查\"。但四个 MoE 推理模型、四个基准的十六组设置里，这种提示词的省 token 效果**没有一组**比得上直接改路由器；路由器干预中位数省 30.4%，最好一组 63.3%（[arXiv:2611.05919](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.05919)）。\n\n另一半反着来：直觉上该砍的是\"反思\"，实测砍它反而让轨迹**变长** 5%；能缩短轨迹的是砍**规划**。\n\n## 路由器是天然的行为仪表盘\n\n给推理过程逐段打上\"这一步在干什么\"的标签，再看内部路由能不能读出这个标签。行为分六类：规划、执行、复述、事实、反思、结论。标注用 GPT-5.5 当裁判，把 GSM8K、NuminaMath、AIME 2023 的 330 条推理轨迹切成句级或子句级片段，共 78,955 个带标签片段。在 Qwen3-30B-A3B 上做线性探针，读单层 128 个 pre-top-k 路由 logit：L12 层 macro-F1 是 0.608，\"永远猜最常见的行为\"只有 0.096。对照的 2048 维残差流，最好单层 L14 为 0.658；路由侧把 L8–L16 九层拼起来是 1,152 维、0.659，宽度只有残差流的 56%。控制实验显示，同样九层的残差流压到 1,152 维只有 0.657——差距不是\"层多就好\"。\n\n这个探针**不是靠表面措辞**：片段级对比里路由特征 0.709，高于输入嵌入 0.624、TF-IDF 0.566 和轨迹位置 0.124；丢掉开头五个 token 再池化仍有 0.677。\n\n## 行为和专家的绑定能跨领域复用\n\n作者用 1,980 条 Qwen3-30B-A3B 轨迹检验（覆盖数学、代码、科学、知识、常识）。\n\n第 12 层上，同一行为的专家画像相关系数 0.729，不同行为是 -0.107；各数据集 top-10 专家集合的平均 Jaccard 相似度 0.511，打乱标签做零假设只有 0.202。规划、反思、结论三种行为，六个数据集共享各自 top-10 里的 6 个专家，承载 15%–23% 的路由流量。跨域迁移更硬：第 36 层上用源数据集选出的专家集合给目标数据集排序，平均 AUC 0.768。反思 0.863、结论 0.811、复述 0.745，执行最弱只有 0.631。\n\n## 改路由，确实省 token\n\n推理时把 logit 转成 log-softmax，压制的专家压到最低值减一个 margin，激活的提到最高值加 margin，让模型重算门控——不动权重不重训。模型是 Gemma-4-26B-A4B-it、Qwen3-30B-A3B、Qwen3.6-35B-A3B、GPT-OSS-120B，基准 MATH500、AIME 2026、GPQA-Diamond、AICrypto，数学题答案由 DeepSeek-V4 Flash 当裁判判分。\n\n十六组设置全部省下 token，中位数 30.4%，最好一组是 Qwen3.6 在 AICrypto 上的 63.3%；准确率上十四组持平或更高，剩下两组最多掉 0.5 个百分点。省得最狠的是较简单的 MATH500 和 AICrypto（平均 43.2%），AIME 和 GPQA-Diamond 只有 24.9%。\n\n两个对照让结果站得住。一是随机专家：同等规模的随机专家集合接受同样干预，十六组里有十四组反而让 token 变多——省下的算力来自选对专家。二是提示词：写一句\"简洁、别重复检查\"的系统提示词，省 token 效果十六组里全部不如路由干预。\n\n## 反直觉：该砍规划，不是砍反思\n\n一百道 MATH500 题、Qwen3-30B-A3B 上的逐行为对照：压 200 个规划专家，token 掉 22%；压反思专家，轨迹**变长** 5%；压复述变长 1%，答案之后的复查轮次还翻倍以上；压执行变长 52%，下手重一点准确率掉到 40%。\n\n机制很明确：每一轮复查都由一句新的计划开头——\"再查一遍\"\"换个思路\"。所以压掉规划不是让模型停止规划，而是把**下一个计划**从\"开启新一轮\"改写成\"就此收尾\"，整轮连同里面的行为一起消失。压反思没用：模型嘴上不多说疑心了，但发起每一轮的规划还在，轮数照旧。\n\n时机也对得上：首个答案出现得很早，中位数占 Qwen3.6 轨迹的 8%、Qwen3 的 11%、Gemma-4 的 21%，而 82%–88% 的规划句在其后。切割实验证实——只在首个答案之后干预，之前的推理几乎原样保留（首个答案出现在第 422 个 token 而非 441 个），单靠这一段就复现了整条轨迹的效果，准确率从 84% 提到 97%。\n\n再往下，论文数了 314 个\"即将开启新阶段\"的切点：未干预时\"开新角度\"和\"收尾\"是 23 比 23，压掉规划后收尾 68 比 11，模型在这约 170 个 token 内结束思考的比例从 5% 升到 17%。只压紧邻的一小段则收尾倾向仍在（43 比 21）但结束率几乎不变（6% 对 5%）——干预得一直挂着。\n\n## 三个还没解决的问题\n\n标注只有一个裁判。全部标签来自单个 LLM 裁判 GPT-5.5，论文没报人类一致性，而规划与反思的边界本就模糊。这也是效率不是能力：论文写明激活行为关联的专家并不能可靠地提升准确率，路由器是\"花多少算力\"的旋钮，不是\"能不能答对\"的旋钮。\n\n规模没验证。论文覆盖的模型止于 GPT-OSS 的 120B，而开源 MoE 已到几百亿到上万亿参数。代码以 MIT 许可证开源在 GitHub（[swang-ai\u002Freasoning-moe](https:\u002F\u002Fgithub.com\u002Fswang-ai\u002Freasoning-moe)），但仓库 11 月 4 日才建立，star 还是 0。\n\n## 所以呢\n\n真正值钱的可能不是那 30.4%，而是\"路由 logit 能读出推理状态\"本身——给 MoE 模型挂上几百维探针，就能实时知道它此刻在规划、在执行还是反复怀疑。对做 Agent 的人来说这才是可观测性。\n\n另一半提醒很扎心：给后训练出来的推理模型写\"请简洁一点\"，效果接近安慰剂——十六组对照说明系统提示词对后训练养成的验证习惯几乎没有约束力。想省算力，得找到具体那个控制点。\n\n但别急着上生产：模型必须是稀疏 MoE，你得能拿到 router logits，还得先跑标注和专家筛选。\n\n对研究者来说剩下的问题也有意思：路由 logit 里读出来的\"规划\"，跟残差流里的\"规划\"，是同一个东西吗？如果是，MoE 的路由器就不只是负载均衡器，而是实时的推理状态日志。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.05919","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"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",{"id":22,"name":23,"slug":23,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"aa1c2aad-1255-4033-bf68-341eabb14baf","en","MoE Router Reads Reasoning: Cut Planning, Not Reflection","MoE router logits read reasoning state: suppressing planning cuts tokens a median 30.4% (up to 63.3%), accuracy -0.5pp. Cut planning, not reflection.","Reasoning models \"overthink\" — the answer is already computed, yet they keep re-verifying and re-deriving it. The standard remedy is a system prompt: \"keep it concise, skip unnecessary re-checks.\" A harder set of numbers says otherwise. Across sixteen model–benchmark settings built from four MoE reasoning models and four benchmarks, that prompt failed to beat direct router editing in a single setting. Router intervention cut generated tokens by a median of 30.4%, reaching 63.3% in the best case ([arXiv:2611.05919](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.05919)).\n\nThe other half runs the other way. Intuitively the thing to cut is \"reflection\" — those \"Wait, let me double-check\" sentences. Suppressing reflection-associated experts makes traces **longer** by 5%. The behavior that shortens a trace is suppressing **planning**.\n\n## The router is a natural instrument panel for reasoning state\n\nThe idea is direct. First segment the reasoning, label each span with \"what is this step doing\", then ask whether the model's internal routing exposes that label.\n\nBehaviors fall into six classes: planning, execution, restatement, fact, reflection, and conclusion. Annotation used GPT-5.5 as the judge, cutting 330 reasoning traces from GSM8K, NuminaMath, and AIME 2023 into sentence- or clause-level spans — 78,955 labeled segments in total.\n\nOn Qwen3-30B-A3B, a linear probe reading the 128 pre-top-k router logits of a single layer reaches 0.608 macro-F1 at layer L12, while \"always predict the most frequent behavior\" manages only 0.096. The best single layer of the 2,048-dimensional residual stream, L14, reaches 0.658; concatenating router layers L8–L16 gives 1,152 dimensions and 0.659, at 56% of the residual stream's width. The authors also ran the control: residual states from those same nine layers, compressed to the same 1,152 dimensions, reach 0.657. The gap is not an artifact of using more layers.\n\nWhat matters more than the numbers is that the probe does **not** key on surface wording. In a segment-level comparison, router features reach 0.709, above input embeddings (0.624), lexical TF-IDF (0.566), and trace position (0.124). Dropping a segment's opening five tokens and pooling what remains still gives 0.677, while those five tokens alone give 0.626.\n\n## Behavior-to-expert bindings survive a change of domain\n\nIf the correspondence were an artifact of one dataset, the method would be practically useless. The authors test it with 1,980 Qwen3-30B-A3B traces, 330 from each of six datasets spanning mathematics, code, science, knowledge, and commonsense.\n\nAt layer 12, the expert-profile correlation within the same behavior is 0.729, while across different behaviors it is -0.107. The mean cross-dataset Jaccard similarity of the top-10 expert sets is 0.511, against 0.202 for a label-permutation null. For planning, reflection, and conclusion, six of the top ten experts are shared by all six datasets, carrying 15%–23% of routing mass.\n\nThe cross-domain transfer test is the harder one: at layer 36, expert sets selected on a source dataset rank segments of a target dataset with mean AUC 0.768. Reflection reaches 0.863 and conclusion 0.811, restatement 0.745, while execution is weakest at 0.631.\n\n## Editing the router really does save tokens\n\nThe intervention is not complicated. At inference time, map logits to log-softmax, push the experts to be suppressed down to the minimum score minus a margin, push the ones to be activated up to the maximum plus a margin, then let the model recompute its softmax, top-k, and gate normalization. No weight updates, no retraining.\n\nFour models were tested — Gemma-4-26B-A4B-it, Qwen3-30B-A3B, Qwen3.6-35B-A3B, and GPT-OSS-120B — against four benchmarks: MATH500, AIME 2026, GPQA-Diamond, and AICrypto, with math answers graded by DeepSeek-V4 Flash as a judge.\n\nAll sixteen settings save tokens, with a median of 30.4% and a best case of 63.3% for Qwen3.6 on AICrypto; on accuracy, fourteen settings are unchanged or better and the remaining two drop by at most 0.5 percentage points. The savings are deepest on the easier MATH500 and AICrypto (43.2% on average) and shallowest on AIME and GPQA-Diamond (24.9%) — the more the model overthinks, the more there is to reclaim.\n\nTwo control experiments make the result hold. The first is random experts: an equally sized random expert set receiving the same intervention **increased** token generation in fourteen of sixteen settings, so the savings come from choosing the right experts rather than from perturbing the router in general. The second is the prompt baseline: a system instruction asking for concise reasoning saved fewer tokens than router intervention in all sixteen settings, and matched or beat its accuracy in only fourteen.\n\n## The counterintuitive part: cut planning, not reflection\n\nThe per-behavior comparison on 100 MATH500 problems with Qwen3-30B-A3B is close to backwards. Suppressing 200 planning experts cuts generated tokens by 22%. Suppressing reflection experts lengthens traces by 5%. Suppressing restatement lengthens them by 1%, and more than doubles the number of checking rounds after the answer. Suppressing execution lengthens them by 52%, and applied more strongly, drops accuracy to 40%.\n\nThe mechanism is clear: every round of checking is opened by a fresh plan — \"let me check again,\" \"alternatively,\" \"another check.\" So suppressing planning does not stop the model from planning, it rewrites the **next plan** from \"open another round\" to \"wrap up,\" and the whole round disappears along with the behaviors inside it. This explains why suppressing reflection fails: the model voices fewer doubts, but the planning that launches each round is untouched, so the rounds keep running.\n\nThe timing matches too. Models reach an answer early — at a median of 8% of the Qwen3.6 trace, 11% for Qwen3, and 21% for Gemma-4 — and 82%–88% of planning sentences fall after that point. The cut experiment confirms it: steering only after the answer leaves the reasoning before it essentially intact (the answer arrives after 422 tokens instead of 441), and that single side reproduces the effect of steering the whole trace, raising accuracy from 84% to 97%.\n\nGoing one level deeper, the paper counts 314 cut points where a new stage is about to begin, regenerating the next ~170 tokens with the same seed: unmodified, \"open a new angle\" and \"wrap up\" split 23 against 23; with planning experts suppressed, wrap-ups outnumber new angles 68 to 11, and the model ends its thinking within those ~170 tokens 17% of the time instead of 5%. Suppressing only the immediately preceding stage still tilts the first plan toward wrapping up (43 against 21) but barely changes how often thinking ends (6% against 5%) — the intervention has to stay on.\n\n## Three unresolved problems\n\nAnnotation rests on a single judge. Every behavior label comes from one LLM judge, GPT-5.5, with no human agreement reported, and the boundary between planning and reflection is inherently ambiguous — a sentence that both voices a doubt and proposes to check could be filed either way, which would shift the expert associations built on top.\n\nThis is efficiency, not capability. The paper states it plainly: activating the experts associated with a behavior did not reliably improve accuracy. The router is here a lever on how much a model computes, not on whether it reaches the right answer.\n\nScale is unverified. The model scale studied stops at GPT-OSS's 120B, while open-weight MoE models now reach several hundred billion to over a trillion parameters with more experts per layer. The code is released under an MIT license on GitHub ([swang-ai\u002Freasoning-moe](https:\u002F\u002Fgithub.com\u002Fswang-ai\u002Freasoning-moe)), but the repository was created on November 4 and still has zero stars.\n\n## So what\n\nThe most valuable part may not be the 30.4%, but the fact that router logits can read out reasoning state at all: attach a few-hundred-dimensional probe to an MoE model and you can know in real time whether it is currently planning, executing, or second-guessing itself. For people building agents that is observability — a probe telling you \"the model is stuck in this mode of thinking right now\" is far more useful than a counter reporting only how many tokens a run cost.\n\nThe other lesson is blunter. Telling a post-trained reasoning model \"please be concise\" works about as well as a placebo. All sixteen comparisons show that system prompts barely constrain the verification habits instilled during post-training. To save compute, you have to find the specific control point.\n\nThat said, do not put this in production yet. The preconditions are real: the model must be sparse MoE, you need access to router logits, and you still have to run labeling and expert selection first. Dense models have no such path. And the paper itself flags that every behavior label rests on a single LLM judge — before pointing this at production traffic, you would want humans to check it.\n\nFor researchers the open question is equally interesting: is the \"planning\" read out of router logits the same thing as the \"planning\" read out of the residual stream? If it is, then an MoE router is not just a load balancer but a live log of reasoning state — and that may be worth more than the thirty percent of tokens it saves.\n","moe-router-reasoning-behavior-steering","2026-11-05T09:21:56Z","2026-11-05T09:22:08.567101Z","2026-11-05T09:22:08.567112Z",true,"agent",61,[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},"0e56c8bc-48db-43a4-b06c-ab9178358885","KAIST：MoE 解码的阈值校准可以整个删掉","kaist-moe-channel-selection-no-calibration","2026-10-28T09:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"50846ffe-7171-471c-889c-247a3b86d553","华为InferOpt:逐层旋钮搜着调,KV缓存省64%","inferopt-llm-inference-config-search","2026-10-14T21:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"859569a7-b657-498e-ac1c-548f16e3418f","华为论文把 MoE 专家缓存做成可训练目标:命中率提升至多33个点,流量省一半","moe-expert-cache-trainable-router","2026-09-08T19:05:30+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"a579e5c5-e386-49ec-a16a-34722f30662e","arXiv ACE 论文：免训练 MoE 推理减半,token 跳过率 50%","ace-moe-expert-skipping","2026-09-08T03:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"4f745af8-2962-450f-a4b1-612b796d7dd7","PCoMoE:把 MoE 专家拆成路径组合,推理提速 1.31 倍还涨 10% 精度","pcomoe-moe-path-composition-inference","2026-09-02T21:10:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"77a567e5-2209-4981-882b-e2a3077a0012","先猜再听:全双工语音模型把投机解码搬到音频到来之前","speculative-listening-full-duplex-speech-decoding","2026-11-06T09:05:00+00:00"]