MoE Routing Reveals Reasoning Behavior, Cutting Test-Time Compute by 30% Without Training

Reasoning models allocate test-time compute resources when answering questions, using them for behaviors such as planning, derivation, and verification. However, models often "overthink"—continuing to verify and repeatedly re-derive after they have already found the correct answer. This not only wastes expensive compute but can also degrade performance. Existing intervention methods (such as token budgets or steering) primarily regulate reasoning length rather than targeting specific behaviors. A recent study proposes that the routers in Mixture-of-Experts (MoE) models provide a natural, training-free interface for behavior control.

MoE Routers: A Natural Probe for Reasoning Behavior

The research team split reasoning traces into sentence-level segments and annotated each segment with one of six reasoning behaviors: planning, execution, restatement, fact, reflection, and conclusion. In this way, each token carries the behavior label of the segment it belongs to. They then trained linear probes on router logits to predict these labels and found that their predictive performance was nearly comparable to probes trained on the residual stream (the standard probing target using the model's full hidden states), while using only about half the number of features.

These signals generalize: experts associated with a certain reasoning behavior on one dataset can also predict the same behavior on five other datasets. These six datasets cover five domains: math, code, science, knowledge, and common sense. This means that routing patterns not only reflect the model's internal computational state but also encode transferable behavioral semantics.

Mixture-of-Experts (MoE) and Router Mechanism

A Mixture-of-Experts (MoE) model consists of multiple parallel "expert" subnetworks, each specializing in different types of input patterns. The router is a lightweight gating network that computes weight scores (logits) for each expert for every input token, and based on these decides which experts to activate and how to weight their outputs. This sparse activation mechanism allows MoE models to maintain a huge parameter count while keeping computational costs far lower than dense models of equivalent scale. Router logits thus become an observable signal of the model's internal decision process: they not only reflect which experts a token should be assigned to but also implicitly encode the model's judgment of the current input features. This study leverages this property by treating router logits as a "fingerprint" of reasoning behavior, decoding high-level behavior labels such as planning, execution, and reflection from them via linear probes.

Linear Probes and Residual Stream Probing

A linear probe is a common interpretability analysis tool: a simple linear classifier is trained on the model's internal representations (such as hidden states or router logits) to predict some externally annotated property (such as reasoning behavior labels). If a linear probe can predict labels with high accuracy, it indicates that the relevant information is linearly encoded in that internal representation. The residual stream is the "main highway" along which hidden states accumulate across layers in a Transformer; it is often regarded as the carrier of the model's full computational state and is the standard probing target. This study finds that probes trained on router logits perform nearly as well as those trained on the residual stream, while using only about half the feature dimensions, indicating that router logits encode behavioral semantics in a more compact form that is easier to decode linearly.

Selectively Editing Routing Scores to Suppress Overthinking

Building on these findings, the researchers selectively edit expert routing scores during inference to suppress overthinking. Rather than crudely truncating reasoning length, they guide the model to stop after solving the problem while preserving necessary reasoning steps. Across four models and four benchmarks, this intervention reduced the median number of generated tokens by 30.4%, with a maximum reduction of 63.3%.

Crucially, accuracy never dropped by more than 0.5 percentage points and often improved. This shows that reducing compute does not come at the expense of performance. Compared with existing methods, this technique directly targets reasoning behavior rather than merely controlling length, making it more precise.

Test-Time Compute and Overthinking

Test-time compute refers to the additional computational resources a model consumes during inference, such as generating longer chains of thought, performing multiple samplings, or self-verification. Increasing test-time compute typically improves accuracy on complex tasks, but more is not always better: once the model has found the correct answer, continuing to generate verification steps and repeated derivations wastes compute and can even degrade performance by introducing noise—a phenomenon known as "overthinking." Existing interventions such as token budgets (limiting maximum generation length) or steering prompts (encouraging concise reasoning) mainly regulate reasoning duration rather than providing fine-grained control over specific reasoning behaviors, and may therefore inadvertently harm necessary planning or verification steps. This study achieves behavior-level precise intervention by editing router scores to suppress redundant behaviors such as reflection and restatement.

Multi-Source Cross-Validation and Industry Implications

According to the original study and community analysis, MoE routing has dual value as both an explanatory tool for reasoning states and an optimization mechanism for test-time compute. Unlike other work that focuses only on reasoning length, this method provides behavior-level interpretability. Some analyses note that its training-free nature makes it easy to deploy on existing MoE models. However, others caution that the method has so far been validated only on specific models and benchmarks, and its generality still needs further testing.

For AI application developers, this means inference costs can be significantly reduced without retraining the model, while maintaining or even improving accuracy. For AI coding and AI application scenarios, this technique helps balance computational overhead and response quality in complex tasks. The open-source tooling community may also benefit, as the method does not rely on proprietary components.

Behavior Label System and Cross-Domain Generalization

The study splits reasoning traces by sentence and annotates six behaviors: planning (formulating a solution strategy), execution (performing specific calculations or derivations), restatement (repeating existing information), fact (citing external knowledge), reflection (checking or verifying intermediate results), and conclusion (giving the final answer). Although this granularity is sentence-level, it is already sufficient to distinguish functional roles in the reasoning process. The significance of cross-dataset generalization lies in the fact that if an expert is associated with "reflection" behavior on math data, it can also predict "reflection" on code or science data, indicating that routing patterns encode domain-agnostic behavioral semantics rather than task-specific surface features. This provides a basis for transferring the method to new domains and explains why it can be deployed without retraining.

Limitations and Future Directions

The study still has limitations: the granularity of behavior labels is relatively coarse (sentence-level) and may fail to capture finer-grained reasoning dynamics; the models and tasks covered by the experiments are limited, and validation in broader scenarios has not yet been conducted. Future work could explore more fine-grained behavior classification and applicability across different architectures.

Overall, this study demonstrates the potential of MoE routing as an interface for reasoning behavior and opens a new path for test-time compute optimization.

One-line Summary

Using MoE router logits to probe and edit reasoning behavior reduces token consumption by over 30% without training and without sacrificing accuracy.

Paragraph Summary

This study proposes a method for controlling reasoning behavior based on Mixture-of-Experts (MoE) model routers. The researchers split reasoning traces into sentence-level segments and annotate six behaviors (planning, execution, restatement, fact, reflection, conclusion), then train linear probes on router logits. They find that these probes achieve predictive performance comparable to residual-stream probes with about half the features and generalize across six datasets. Based on this, they selectively edit routing scores to suppress overthinking, reducing median generated tokens by 30.4% (up to 63.3%) across four models and four benchmarks, with accuracy dropping by no more than 0.5 percentage points. The method is training-free and directly targets behavior rather than length, offering a new path for test-time compute optimization, though coarse behavior-label granularity and limited experimental coverage remain current limitations.