Steering AI’s internal routers to cut repetitive reasoning

Reading reasoning through routing

In the preprint Routes of Reasoning: Reading and Reshaping Test-Time Compute in MoE Models, Shengjie Wang, Mohsen Fayyaz and Nanyun Peng investigate an internal interface for understanding and steering AI reasoning.

Mixture-of-experts, or MoE, models route each token through a small selection of expert neural networks at each relevant layer. The researchers ask whether those routing decisions reflect the role a passage plays in solving a problem, rather than merely its vocabulary or subject matter.

They divide written reasoning into sentence- or clause-level segments and use GPT-5.5 to assign six labels: planning, execution, restatement, fact, reflection and conclusion. These distinguish proposing an approach, deriving something new, repeating information, invoking established knowledge, checking earlier work and stating a result. The labels describe the function of the written reasoning, not whether it is correct.

Using Qwen3-30B-A3B, the team trained simple classifiers to predict these labels from internal routing scores. Combining scores from nine layers achieved a macro-F1 score of 0.659, compared with 0.658 for a classifier using the model’s residual-stream representation, a standard target for this kind of analysis. The router representation used 1,152 features rather than 2,048.

That near-match required combining layers: a single router layer reached 0.608. Comparisons with text-based features and analyses excluding segments’ opening tokens also supported the authors’ interpretation that routing carried information beyond obvious introductory phrases.

Associations across different subjects

The researchers next examined 1,980 reasoning traces from six datasets spanning mathematics, code, science, knowledge and commonsense. They identified experts selected more frequently during particular behaviors relative to each dataset’s overall expert usage.

Expert sets selected on one dataset could help identify the same behavior on another. At the reported layer, their mean transfer AUC—a measure of how well the scores rank relevant segments—was 0.768. Reflection and conclusion transferred more strongly than execution.

These are behavioral associations, not proof that individual experts exclusively implement particular reasoning functions. The paper explicitly distinguishes identifying useful intervention targets from establishing what an expert causally controls.

Shorter responses without retraining model weights

For steering, the team changed router scores during generation so that selected experts would be excluded from the routed computation. The intervention did not update model parameters, although choosing the expert sets required analysis and development-set evaluation.

Tests covered Gemma-4-26B-A4B-it, Qwen3-30B-A3B, Qwen3.6-35B-A3B and GPT-OSS-120B. Each was evaluated on MATH500, AIME 2026, GPQA-Diamond and AICrypto, covering mathematics, scientific questions and cryptographic reasoning. Intervention configurations were selected on development data disjoint from the test benchmarks.

Across the sixteen model–benchmark settings, the authors report a median reduction of 30.4% in generated tokens. The largest reduction, 63.3%, occurred with Qwen3.6 on AICrypto. Accuracy was unchanged or higher in fourteen settings; the other two showed decreases of 0.3 and 0.5 percentage points.

The targeted interventions saved more tokens than a system prompt requesting concise reasoning in every setting. Equally sized random expert interventions generally did not reproduce the benefit: they increased token generation in fourteen settings and often reduced accuracy.

Why suppress planning?

Every selected steering configuration suppressed planning-associated experts, sometimes alongside other behavioral targets. That might sound counterproductive, but the authors’ follow-up experiments suggest the intervention changes what the model plans next rather than eliminating planning.

Repeated checks often begin with a fresh proposal to try another approach. In Qwen3.6 experiments on the first 100 MATH500 problems, applying suppression only after the first answer reproduced the effect of full-trace steering. Suppression before that answer left the initial reasoning largely unchanged. Separate continuation experiments found more plans directed toward wrapping up and fewer toward opening another angle.

The authors interpret these results as evidence that steering primarily curbs post-answer repetition, rather than simply removing essential derivation steps.

What remains uncertain

All behavior labels came from one LLM annotator, with no reported human agreement. Ambiguous boundaries—especially between planning and reflection—could affect the resulting associations.

The largest tested model had 120 billion parameters, leaving effectiveness on substantially larger MoE systems unresolved. The authors also frame the contribution as efficiency rather than a reliable increase in reasoning capability: boosting behavior-associated experts did not consistently improve accuracy. Longer-horizon and agentic tasks remain directions for future testing.