包裹，可能有些部分需要分小节吗？原文没有明显的标题，但可以用
或
吗？原文的结构是：问题引入→方法（拆分标注、线性探测）→信号泛化→干预应用→结果和总结。可以把方法部分用
吗？比如“方法：行为标注与线性探测”？不过原文没有标题，可能需要保持原文的结构，因为规则说不得增删改内容，所以不能加标题？哦对，规则第四条是保留原文所有信息，不得增删改内容，所以不能添加原文没有的标题，对吧？那只能用
包裹每个自然段落，还有列表、引用、代码这些，但原文里没有列表、引用、代码，所以主要是
，重要句子用。 首先，原文的第一句：“Reasoning models allocate test-time compute across distinct behaviors such as planning, deriving, and verifying.” 然后下一句：“However, models frequently "overthink," continuing to verify and re-derive long after finding a correct answer.” 这部分是问题，可以放在第一个
里。然后接下来：“This wastes expensive compute and can even degrade performance. Existing interventions, such as token budgets or steering, primarily regulate reasoning duration rather than targeting specific behaviors.” 可以接着，或者合在一个
？原文的自然段落大概是：第一段讲推理模型的计算分配和过度思考的问题，以及现有干预的不足。第二段讲MoE路由器作为训练无关的行为控制接口。第三段讲拆分推理痕迹成句子片段并标注六种行为，线性探测的结果。第四段讲信号在不同数据集上的泛化。第五段讲基于此的干预，减少token的结果，准确率的变化。第六段总结MoE路由的意义。 现在开始组织HTML： 首先，第一句：“Reasoning models allocate test-time compute across distinct behaviors such as planning, deriving, and verifying.” 然后下一句：“However, models frequently "overthink," continuing to verify and re-derive long after finding a correct answer.” 可以用一个
包裹：“
Reasoning models allocate test-time compute across distinct behaviors such as planning, deriving, and verifying. However, models frequently "overthink," continuing to verify and re-derive long after finding a correct answer.
” 接下来：“This wastes expensive compute and can even degrade performance. Existing interventions, such as token budgets or steering, primarily regulate reasoning duration rather than targeting specific behaviors.” 另一个：“
This wastes expensive compute and can even degrade performance. Existing interventions, such as token budgets or steering, primarily regulate reasoning duration rather than targeting specific behaviors.
” 然后：“In this work, we demonstrate that Mixture-of-Experts (MoE) routers provide a natural, training-free interface for behavioral control.” 这个独立成：“
In this work, we demonstrate that Mixture-of-Experts (MoE) routers provide a natural, training-free interface for behavioral control.
” 接下来：“We first split reasoning traces into sentence-level segments and label each with one of six reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion), so that every token carries the behavior of its segment.” 这个：“
We first split reasoning traces into sentence-level segments and label each with one of six reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion), so that every token carries the behavior of its segment.
” 然后：“Linear probes on router logits predict this label nearly as well as probes on the residual stream, the model's full hidden state and the standard target for probing, while using about half as many features.” 这个，可以把“nearly as well as”和“while using about half as many features”用吗？比如重要句子，但原文的重要句子可能是关键结论，比如“Linear probes on router logits predict this label nearly as well as probes on the residual stream, the model's full hidden state and the standard target for probing, while using about half as many features.” 这里可以用强调吗？规则说重要句子可用，所以可以：“
Linear probes on router logits predict this label nearly as well as probes on the residual stream, the model's full hidden state and the standard target for probing, while using about half as many features.
” 接下来：“These signals generalize: experts associated with a reasoning behavior on one dataset predict the same behavior on the others, across six datasets spanning five domains (math, code, science, knowledge, and commonsense).” 这个：“
These signals generalize: experts associated with a reasoning behavior on one dataset predict the same behavior on the others, across six datasets spanning five domains (math, code, science, knowledge, and commonsense).
” 然后：“Building on this, we selectively edit expert router scores during inference to curb overthinking.” 这个，可以加：“
Building on this, we selectively edit expert router scores during inference to curb overthinking.
” 接下来：“Across four models and four benchmarks, this intervention reduces the number of generated tokens by a median of 30.4% (and up to 63.3%).” 这个，数字可以加：“
Across four models and four benchmarks, this intervention reduces the number of generated tokens by a median of 30.4% (and up to 63.3%).
” 然后：“Crucially, it encourages the model to stop once the problem is solved, rather than skipping essential reasoning steps.” 这个是重要的点，用加：“
Crucially, it encourages the model to stop once the problem is solved, rather than skipping essential reasoning steps.
” 接下来：“Accuracy never drops by more than 0.5 percentage points, and frequently improves.” 这个：“
Accuracy never drops by more than 0.5 percentage points, and frequently improves.
” 最后：“Ultimately, MoE routing offers both an interpretable lens into reasoning states and a lightweight mechanism for optimizing test-time compute.” 这个总结句，用加：“
Ultimately, MoE routing offers both an interpretable lens into reasoning states and a lightweight mechanism for optimizing test-time compute.
” 现在检查是否符合规则：没有文档级标签，每个自然段落用包裹，重要句子用了，保留了所有内容，没有增删改，没有列表、引用、代码，所以不需要那些标签。现在组合起来：
Reasoning models allocate test-time compute across distinct behaviors such as planning, deriving, and verifying. However, models frequently "overthink," continuing to verify and re-derive long after finding a correct answer.
This wastes expensive compute and can even degrade performance. Existing interventions, such as token budgets or steering, primarily regulate reasoning duration rather than targeting specific behaviors.
In this work, we demonstrate that Mixture-of-Experts (MoE) routers provide a natural, training-free interface for behavioral control.
We first split reasoning traces into sentence-level segments and label each with one of six reasoning behaviors (planning, execution, restatement, fact, reflection, and conclusion), so that every token carries the behavior of its segment.
Linear probes on router logits predict this label nearly as well as probes on the residual stream, the model's full hidden state and the standard target for probing, while using about half as many features.
These signals generalize: experts associated with a reasoning behavior on one dataset predict the same behavior on the others, across six datasets spanning five domains (math, code, science, knowledge, and commonsense).
Building on this, we selectively edit expert router scores during inference to curb overthinking.