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).

The 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.

The router is a natural instrument panel for reasoning state

The 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.

Behaviors 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.

On 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.

What 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.

Behavior-to-expert bindings survive a change of domain

If 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.

At 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.

The 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.

Editing the router really does save tokens

The 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.

Four 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.

All 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.

Two 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.

The counterintuitive part: cut planning, not reflection

The 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%.

The 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.

The 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%.

Going 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.

Three unresolved problems

Annotation 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.

This 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.

Scale 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/reasoning-moe), but the repository was created on November 4 and still has zero stars.

So what

The 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.

The 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.

That 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.

For 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.