acceptodds
Under review as a conference paper at ICLR 2027

Routes of Reasoning: Reading and Reshaping Test-Time Compute in MoE Models

Abstract

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. 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%). Crucially, it encourages the model to stop once the problem is solved, rather than skipping essential reasoning steps. Accuracy never drops by more than 0.5 percentage points, and frequently improves. Ultimately, MoE routing offers both an interpretable lens into reasoning states and a lightweight mechanism for optimizing test-time compute.

Primary area:
interpretability and explainable AI
Keywords:
Mixture-of-Experts, Interpretability, Model Steering, Test-Time Compute, Efficient Reasoning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_oKFnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper observes that tokens in different reasoning behaviors have distinct routing patterns and this observation generalizes across datasets and domains. Building on this, authors propose a method to curb overthinking during MoE LLM inference by selectively editing expert router scores. Extensive experiments across 16 models and benchmarks demonstrated their method effectively reduces generated tokens by a median of 30.4% while largely preserving accuracy.

Critical strengths
  1. The topic of this paper is interesting and meaningful. It provides a novel perspective to change MoE model’s behaviors by changing router’s behaviors, which is training-free.

  2. This paper presents a thorough and logically structured investigation into the model’s internal mechanism. Its analysis is supported by solid statistical evidence, providing strong support of existence of behavior-associated experts.

Critical weaknesses
  1. The observation of behaviors-associated experts suggests that certain reasoning behavior is associated with certain router behavior. However, it remains unclear whether editing router behavior can change reasoning behaviors causally. There is a logical gap between observation and the proposed method.

  2. Concerns about application. The additional cost of perturbing the router behavior remains unclear so the practical efficiency improvement is also unclear. Moreover, steering method requires separate tuning for each benchmark, leaving its cross-task generalization ability and efficiency under such generalization unclear.

Questions for authors
  1. Equation 4 in the paper considers relative value of the router scores to identify behaviors-associated experts, what is the typical range of the absolute value of router score of these experts? Do they generally have high router scores? If not, steering experts with different router score have the same impact on model?

  2. The paper identifies behavior-associated experts based on the original model's behavior. However, the steering method changes the model's behavior. Could these changes also alter the set of behavior-associated experts? In other words, are the identified behavior-associated experts still valid after steering?

  3. Could authors provide the comparison with other expert steering baselines? The current baselines in this paper are too simple.

  4. Could authors provide analysis about practical efficiency (e.g. wall clock time) of proposed method, which also considers the cost of identifying and steering experts?

  5. The paper splits reasoning behaviors using a judge LLM (GPT5.5), which limits the deployment of this method.

Reviewer_j8k5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper examines whether Mixture-of-Experts (MoE) routing can identify and control distinct reasoning behaviors in LLMs. The authors classify reasoning traces into six behaviors (planning, execution, restatement, fact, reflection, and conclusion) and show that linear probes predict these behaviors about as well as residual-stream probes. They also find that behavior-associated MoE experts show consistent patterns across domains. Building on this, they propose a training-free intervention that adjusts expert routing scores to reduce unnecessary reasoning. Across four MoE models and four benchmarks, this cuts generated tokens by a median 30.4% while preserving accuracy. Further analysis suggests that suppressing planning-associated experts mainly reduces redundant verification after the model has already reached an answer.

Critical strengths
  1. Novelty: The paper presents an interesting connection between MoE experts and specific reasoning behaviors, providing evidence that controlling these behaviors can improve LLM reasoning performance.

  2. Strong empirical evaluation: The experiments span multiple LLMs and reasoning benchmarks, demonstrating consistent performance gains (token reduction with preserved accuracy).

Critical weaknesses

Main weakness is that the paper is limited to MoE architectures.

Additionally, the paper would benefit from some additional baselines. For reasoning behavior detection, simple LLM-as-a-judge approach (using the same backbone LLM) could be compared against linear probes. For steering, another useful baseline would be directly modifying the selected expert's activations rather than only its routing score.

Questions for authors
  1. The proposed approach is limited to MoE architectures. Could the authors discuss whether similar behavior steering could be applied to dense LLMs, for example, by modifying intermediate hidden states rather than expert routing weights?

  2. I think the paper would benefit from discussion of when the additional reasoning is beneficial and when it is not - as current experiments suggest that additional reasoning often burns tokens without substantial accuracy improvements, which is contradictory with common findings that more reasoning can be helpful for harder problems.

Reviewer_TdGgOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether the routing patterns of Mixture-of-Experts (MoE) reasoning models encode distinct reasoning behaviors and can be manipulated to reduce redundant test-time computation. The authors categorize reasoning traces into six behaviors. These include planning, execution, restatement, fact recall, reflection, and conclusion. The paper demonstrates that linear probes of router logits can predict these behaviors comparably to probes of the larger residual-stream representations. They further identify behavior-associated experts and show that their associations generalize across multiple reasoning domains. Building on these findings, the authors propose a training-free inference-time intervention that modifies router scores to suppress selected experts, particularly planning-associated experts.

Critical strengths

None

Critical weaknesses
  1. This paper seems very likely to be ai-generated and does not look polished. Even Figure 1 is clearly ai-generated and is void of format control.

  2. The baselines do not establish an advantage over the closest existing reasoning-efficiency methods. The evaluation compares against unmodified decoding, random expert edits, and one brevity prompt, but does not include SEAL or a competitive reasoning-budget intervention under matched token usage.

  3. The claim that steering removes redundant computation while preserving useful reasoning is not yet sufficiently established. The detailed first-answer diagnostics are concentrated on Qwen3.6 and the first 100 MATH500 problems. In Figure 5, the improvement from 84% to 97% accuracy accompanies a reduction from 17 token-capped traces to zero, so this result does not by itself demonstrate preservation of useful verification.

Questions for authors
  1. How are the final layer–expert sets constructed for each of the four models? Please specify the annotation corpora, development-set sizes, allocation of selected experts across layers, and configuration-selection objective. Were all data used to construct the expert rankings disjoint from the steering test problems? In particular, how do the 330 MATH500 problems in Table 5 relate to the rankings used for the MATH500 steering evaluation?

  2. Under the same models, decoding settings, and comparable development-tuning budgets, how does the method compare with SEAL and a simple reasoning-budget intervention at matched average token usage?

  3. What are the paired, problem-level confidence intervals for the accuracy changes in Table 7, accounting for repeated generations of the same problems?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.