acceptodds
Under review as a conference paper at ICLR 2027

RouteVerify: Routing Interventions for Step-Level Verification in Mixture-of-Experts Models

Abstract

Routing states in mixture-of-experts (MoE) models can distinguish correct from erroneous reasoning steps, but such information does not by itself show whether changing routing changes a model's step-level verification judgment. We introduce RouteVerify, a controlled intervention study that replaces selected internal activations from one solution trace with activations from another, while keeping the evaluated input unchanged. Across nine MoE checkpoints, we compare three intervention interfaces: router logits, MoE outputs, and post-MoE residual states, and vary the layers and token positions at which activations are replaced. We measure changes in a verification margin that quantifies the model's preference for INCORRECT over CORRECT. We also evaluate the same input with activations from correct versus erroneous steps to test whether the resulting judgment depends on the source step's correctness. At the same layer and token positions, post-MoE residual replacements produce larger paired margin effects than router-logit or MoE-output replacements. Router-logit interventions become substantially stronger when extended across multiple routed layers and tokens, and intervening on equal numbers of tokens at different positions produces systematically different effects. However, large paired effects do not necessarily imply consistent differences between correct and erroneous source steps. With the evaluated input fixed, erroneous-step activations produce more INCORRECT-favoring margins than correct-step activations under local residual interventions in most checkpoints; this ordering is less consistent under broad routing interventions. These findings distinguish changing a verification judgment from changing it according to the source step's correctness, and show that both depend on the intervention interface and its layer and token coverage.

Primary area:
interpretability and explainable AI
Keywords:
mixture-of-experts, routing, activation patching, causal interventions, step-level verification, mechanistic interpretability

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

This paper asks when and where routing interventions change step-level verification in mixture-of-experts (MoE) models. The model sees a problem and a solution up to one step and judges that step CORRECT or INCORRECT. In nine MoE checkpoints, the authors replace activations in a receiver trace with those from a donor trace, keeping the receiver's input text unchanged. They compare three sites (router logits, MoE outputs, and post-MoE residual states) and vary layer and token coverage.

The outcome is a verification margin, the length-normalized log-probability of INCORRECT minus that of CORRECT. T, the paired directional effect, averages the margin shifts when the two steps of a correct/error pair swap activations. D, the fixed-receiver donor contrast, compares error and correct donors from another problem on the same receiver. The main data are 290 PRMBench step pairs.

At one layer, residual replacements give larger T than the other two sites in all 36 model-by-window comparisons (Figure 2B). At the last token, residual D is positive in eight of nine models, while no single-layer router D is resolved (Table 1). Router interventions across all routed layers and the full step are much stronger and exceed the residual effect in GPT-OSS-120B. Their D is clearly positive only in GPT-OSS-20B (Figure 3). The authors conclude that a large judgment change does not imply dependence on donor correctness.

Critical strengths
  1. The design of T and D metrics addresses an methodological limitation of activation-patching studies: large output changes do not necessarily indicate that the intervened representations causally encode task-relevant correctness information.
  2. Experiments are comprehensive and cover 9 MoE checkpoints, 3 intervention interfaces, multiple layer configurations, and 4 token windows. The inclusion of norm-matched random perturbations, matched external donors, and bootstrap confidence intervals strengthens the empirical investigation.
  3. The results demonstrate that local residual interventions generally outperform router-level interventions, while routing effects become substantially stronger when applied across multiple layers and tokens. The position-dependent effects provide potentially useful insights into MoE verification mechanisms.
Critical weaknesses
  1. The paper primarily applies established activation-patching techniques to MoE verification. While the experiments characterize how intervention effects vary across interfaces, layers, and token positions, they provide limited insight into which experts or routing decisions are responsible for correctness-dependent behavior. Targeted expert-level interventions could strengthen the mechanistic contribution.
  2. The fixed-receiver contrast D is intended to distinguish correctness-dependent effects from intervention magnitude, but correct and erroneous donors also differ in lexical and semantic content. The authors' TF–IDF baseline achieves 0.743 AUROC using boundary-step text alone, highlighting potential surface-level confounds (Appendix G). Without placebo contrasts between different donors with the same correctness label, it remains unclear whether positive D reflects reasoning correctness or unrelated differences between donor representations.
  3. The comparison between residual and router interventions is asymmetric: residual patching replaces the entire hidden representation, whereas router patching only modifies expert selection. Thus, the consistently larger residual effects may primarily reflect intervention scope rather than the relative importance of routing. Moreover, matched external donors produce larger broad-routing effects than paired donors in eight of nine models, raising concerns that these effects reflect off-distribution perturbations rather than correctness-related information transfer. Norm-matched Gaussian controls do not adequately rule out this explanation because they do not preserve realistic routing structure.
Questions for authors
  1. Can the authors identify specific experts or routing decisions that causally influence verification judgments, rather than only comparing aggregate interventions across interfaces and layers?
  2. Could the authors evaluate placebo D contrasts using two distinct donors with the same correctness label, or use lexically matched donors, to determine whether the observed effects are specifically attributable to reasoning correctness?
  3. Can the authors introduce more comparable intervention controls, such as matched activation perturbation strengths or realistic routing perturbations, to distinguish genuine routing-mediated verification effects from differences in intervention scope and off-distribution disruption?
2 comments
AuthorsOpenReview ↗

Response to Reviewer RHfM

Thank you for the careful and accurate review.

Summary. We agree with one point and have tested it. On the other three, we would like to clarify what the paper claims.

  • We agree: D needed a placebo test. We ran yours, and D passes in six of eight models.
  • To clarify: the paper does not rank routing below the residual.
  • To clarify: the matched-donor result is why the paper separates T from D, and D is the control you ask for.
  • To clarify: the paper works at the interface level by design, and its results show why.

Weakness 1 and Question 1: which experts or routing decisions.

The patching operator is standard. What the paper adds is the separation of T and D, which your review lists as a strength, and the interface-level findings.

Can we identify specific routing decisions? Yes. Individual experts? Not in this paper. A single-layer router replacement is weak, while replacing router logits in every layer over the whole step changes the judgment in all nine models (§4.3). For individual experts, the results show where to look. A single-layer router replacement changes the selected experts in 35% to 93% of cases. Yet its D is indistinguishable from zero in all nine models. So at one layer, there is no correctness-dependent routing effect to attribute to them. The effect appears in the experts' output and under broad routing in GPT-OSS-20B (§4.4).

Weakness 2 and Question 2: is D a lexical effect?

We agree that D needed this test, and D passes it. For 195 of the 290 pairs, two of our existing runs had picked different external donors with the same label. That is your placebo: the same receiver, two donors, the same correctness label. An undirected placebo averages to zero, so we direct it by surface content. It is the margin under the donor whose step looks more error-like to TF-IDF, minus the margin under the other.

We fixed the rule before computing the numbers. A model passes if three things hold: real D is above zero; the placebo interval includes zero, or stays below a quarter of real D; real D minus placebo is above zero. Six of eight pass.

modelreal Dplacebo
GLM-4.7-Flash0.183 [0.144, 0.222]−0.004 [−0.036, 0.029]
Nano0.075 [0.063, 0.088]0.004 [−0.006, 0.014]
Lightning0.140 [0.117, 0.164]0.016 [−0.003, 0.035]
Gemma0.805 [0.558, 1.068]−0.157 [−0.418, 0.105]
GPT-OSS-20B0.130 [0.095, 0.167]0.030 [−0.012, 0.070]
GPT-OSS-120B0.213 [0.175, 0.254]0.019 [−0.015, 0.054]
OLMoE0.169 [0.102, 0.241]0.111 [0.062, 0.163]
Qwen3.60.008 [−0.014, 0.031]−0.013 [−0.039, 0.013]

Two exceptions. Qwen3.6 has no residual D, as in Table 1. OLMoE is the only model that verifies worse than the TF-IDF baseline (AUROC 0.602 against 0.743). Its placebo is about two thirds of its real D, and the difference is not distinguishable from zero (0.057 [−0.009, 0.124]). A substantial part of its D is therefore associated with surface content. Sensitivity: six with character-length orientation, five with word count or the total-logprob margin (GPT-OSS-20B: placebo positive, below its real D).

Lexically matched donors: on 80 pairs, TF-IDF rates the correct donor step at least as error-like as the erroneous one. There, D stays positive in seven of eight models. It vanishes in OLMoE (0.001 [−0.022, 0.024]).

We claim dependence on donor correctness beyond this lexical proxy, not pure correctness. The full analysis is in a new appendix of the revised paper, and the OLMoE result is qualified wherever it is cited.

Weakness 3 and Question 3: scope, off-distribution effects, and controls.

Scope. The paper does not conclude that routing matters less. It states that the interfaces replace different amounts of computation (§2.3), and it shows that broad routing exceeds the residual effect in GPT-OSS-120B (§4.3). The comparison with matched scope is router against MoE output, since routing acts only through the routed update. Replacing that update gives a positive D in five of nine models. Replacing the router logits gives none (Table 1). The revised paper states this explicitly.

Off-distribution effects. The paper makes the same point. §5 reports the same eight-of-nine result. It concludes that this should not be read as more correctness-related information. That is why the paper introduces D and does not read T as correctness transfer. New diagnostics, now in the appendix, quantify it: matched donors perturb routing 1.6 to 2.3 times as much as paired donors.

The control you ask for. D is that control. Its two donors are real routing states from the same external problem, so they keep realistic routing structure. They also perturb routing by the same amount (erroneous-to-correct ratio 0.90 to 1.09). Under this control, a correctness-dependent routing effect is resolved under broad routing in GPT-OSS-20B (§4.4), and at a single layer in no model.

Thanks again for your review. Your suggestions led to new analyses in the paper. Here is how the revision addresses each point you raised.

  1. Is D a lexical effect? (Weakness 2, Question 2) We ran the same-label placebo you proposed, and D passes it in six of eight models. On the 80 pairs where the correct donor looks at least as error-like to TF-IDF, D stays positive in seven of eight. §4.2 reports the result, and Appendix L.1 and L.2 give the full analysis.

  2. Scope and off-distribution effects. (Weakness 3, Question 3) §4.2 now states the comparison with matched scope: router against MoE output. Appendix L.3 shows that the erroneous and correct donors used in D perturb routing by similar amounts (KL ratio 0.90 to 1.09).

  3. Which experts are responsible? (Weakness 1, Question 1) The paper works at the interface level, and §8 states this limit. Appendix L.3 adds one result at the expert level. A one-layer router replacement changes the selected experts in 35% to 93% of cases, yet its D is indistinguishable from zero.

The placebo test is now one of the main controls in the paper, and we are grateful for the suggestion. We hope the revision resolves your concerns.

Thank you for your time and effort.

Reviewer_toNcOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the effects of different activation interventions on step-level verification in MoE models. It compares residual-state, router-logit, and MoE-output interventions across multiple models and analyzes their effects on verification judgments.

Critical strengths

The paper evaluates multiple intervention strategies across several MoE models and includes various control experiments.

Critical weaknesses
  1. Limited novelty. This work primarily applies existing activation patch techniques to MoE routing and step-level verification, with limited methodological innovation [1, 2].
  2. Poor presentation. The chart design is inadequate, and there is an excessive emphasis on numerical estimates and confidence intervals, making the paper unnecessarily difficult to read.

[1] Zhang and Nanda. Towards Best Practices of Activation Patching in Language Models: Metrics and Methods.

[2] Zhou et al. The Hidden Signal of Verifier Strictness: Controlling and Improving Step-wise Verification via Selective Latent Steering.

Questions for authors
  1. What is the key technical novelty beyond applying existing activation-patching techniques to MoE models?
  2. Do the observed changes in verification margins translate into meaningful improvements in verification accuracy?
  3. How do the reported uncertainty intervals support the main conclusions, given the inconsistent effects across models and intervention settings?
1 comment
AuthorsOpenReview ↗

Response to Reviewer toNc

Summary. We disagree with both weaknesses, and none of the three questions is about what the paper actually does.

  • Novelty: the review cites two papers as prior work. Both are already in our reference list, and neither does what our paper does.
  • Presentation: the review does not name a chart, and it does not name a problem.
  • Question 1 asks for a new technique. The paper contributes new findings instead.
  • Question 2 asks about improving accuracy. That was never the goal of the paper.
  • Question 3 calls the effects inconsistent, but does not say which ones.

Weakness 1 and Question 1: novelty.

The paper does not propose a new patching technique, and it does not claim to. It is an interpretability study, and its contribution is what it finds. The ICLR 2027 Reviewer Guide counts "new, relevant, impactful knowledge (incl., empirical, theoretical, for practitioners, etc)" as value.

What is new:

  1. It separates two questions: how much an intervention changes the model's judgment (T), and whether that change depends on the donor step being correct (D) (§2.2). Reviewer RHfM describes this as addressing a "methodological limitation of activation-patching studies".
  2. It compares three intervention points in MoE models (router logits, MoE output, residual state) at the same layer and tokens, in nine checkpoints (§2.3).
  3. It shows that a large judgment change does not imply a response to correctness (Table 1; §4.4).

Neither [1] nor [2] does any of these. [1] is a methodology paper on small dense models. [2] steers one residual layer in mainly dense verifiers. Our paper already cites both ([1] in §7 and §8; [2] in §1 and §7).

Weakness 2: presentation.

This weakness has two parts.

First, the review says the "chart design is inadequate". The main text has three figures and one table. The review does not say which one it means, or what is wrong with it.

Second, the review says there is "an excessive emphasis on numerical estimates and confidence intervals". The intervals are the evidence for each claim: an interval tells the reader whether a difference is distinguishable from zero. Question 3 of the same review asks how these intervals support the conclusions. The review treats them as excessive here and as central there.

The review calls the presentation poor, yet its own Presentation score is 2, "generally understandable".

Question 2: "... meaningful improvements in verification accuracy?"

Improving accuracy is not a goal of this paper, and the paper never claims it. The paper is an intervention study. We replace internal activations and measure two things: whether the model's judgment of a step changes, and whether that change depends on the donor step being correct (abstract). The question asks about something the paper does not do.

Question 3: "... inconsistent effects across models and intervention settings?"

The question does not say which effects are inconsistent. The main results hold in every model: the residual intervention has the larger effect in all 36 comparisons (§4.1), and the position effect appears in all nine checkpoints (§4.3). One result does vary across models, the broad routing intervention. The paper reports that variation as a finding, in the abstract and in the title of §4.4. The question treats a finding the paper states as if it were a flaw.

Soundness.

The review gives Soundness 1, defined on the form as "significant concerns about the paper's soundness that do not appear easy to address". The review states no soundness concern. It does not refer to any figure, table, section, checkpoint, dataset, or number in the paper.

We ask the reviewer three questions:

  1. Which passage of [1] or [2] compares these three intervention points, or contrasts correct and erroneous donors on a fixed receiver?
  2. Which chart is inadequate, and in what way?
  3. Which result in the paper is unsound, and why?
Reviewer_rKsGOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper investigates how you can transfer a model's understanding purely by transferring values from the MoEs. It outlines this method of investigation but does not give clear conclusions.

Critical strengths

The core intervention is very interesting. I read this paper and it makes me curious about what the intervention will do.

Critical weaknesses

An interesting investigation is done, but it seems that no clear results were found.

You find that " At the same layer and token positions, post-MoE residual replacements produce larger paired margin effects than router-logit or MoE-output replacements." -- but this seems mechanically true, because of the way that the math works. I may be misunderstanding the implications of this.

Figure 1 appears AI generated. It does not really clarify what's going on. I suggest that you replace Figure 1 with a figure that actually depicts an MoE element within a model, and points directly to the intervention points.

The conclusion says "whether that change follows the correctness of the replacement activation should be evaluated separately" and "equal-length windows reveal systematic dependence on token position", but doesn't really give more detail. How should they be evaluated? What is the systematic dependence?

The paper does not appear to investigate dependence on token position.

This paper should investigate different key points in the reasoning trace for intervention. For example, do filler words like "the" cause any degradation when the MoE are messed with? What about key intermediate results?

Questions for authors

Could you elaborate on what you found most interesting or surprising about the results of this paper? I find the intervention to be intriguing, but I don't feel like I learned very much about the outcome. I would be much more likely to support this paper if you can articulate a clear conclusion.

9 comments
AuthorsOpenReview ↗

Response to Reviewer rKsG

Thank you for the open and constructive review.

Summary. We agree with two of your points. On three others, we would like to clarify what the paper already contains.

  • We agree: the conclusion was too compressed. We state it plainly below and have rewritten it in the revised paper.
  • We agree: the three interventions replace different amounts of computation. The revised paper now says so directly.
  • To clarify: the paper does study token position.
  • To clarify: the residual result is expected, but the math does not force it.
  • To clarify: Figure 1 is not an AI-generated image, and it does mark the intervention points. We have made them easier to see.

Your question: a clear conclusion.

What surprised us most: a large change in the model's judgment is not the same as a change that follows correctness. Three findings:

  1. At a single layer, routing is the weakest lever. Replacing the residual state changes the judgment more than replacing router logits or the MoE output, in all 36 comparisons (§4.1). The residual effect depends on whether the donor step is correct in eight of nine models. For the router, that dependence is indistinguishable from zero in all nine (Table 1).
  2. Routing becomes strong when the intervention is broad. Replacing router logits in every layer over the whole step increases the effect in all nine models, and in GPT-OSS-120B it exceeds the residual effect (§4.3).
  3. A bigger change is not a more correctness-sensitive one. Under broad routing, GPT-OSS-120B changes its judgment more than GPT-OSS-20B, but only the change in GPT-OSS-20B depends on donor correctness (§4.4, Fig. 3).

We have rebuilt the conclusion around these findings and added a contribution list to §1.

Token position.

Token position is one of the things the paper varies (§2.3, §4.3). We replace eight tokens at two positions: the end of the step, and the point where the correct and erroneous solutions first differ (Fig. 2A). Moving to the second position shrinks the gap between the residual and the routing effect in all nine models (Fig. 2D). This is the "systematic dependence on token position" in the conclusion. You are right that the conclusion did not spell it out. The revised conclusion does.

"Evaluated separately" means measured with a separate quantity, D (§2.2). We keep the receiver fixed, give it activations from a correct and from an erroneous step, and compare the two judgments (Table 1, Fig. 3).

"Mechanically true."

You are partly right. A residual replacement swaps more of the computation than a router replacement (§2.3), so a larger effect at one layer is expected. But the math does not force it, and two results go beyond it. First, the order flips: broad routing exceeds the residual effect in GPT-OSS-120B (§4.3). Second, a router replacement does change which experts are selected, in 35% to 93% of cases depending on the model (a new diagnostic, now in the appendix). Yet its dependence on donor correctness stays indistinguishable from zero in all nine models.

Figure 1.

Figure 1 is not an AI-generated image. We built and revised it ourselves as vector artwork in Inkscape, and checked each element against §2. AI assistance in figure preparation is disclosed in the AI-use statement. The submitted figure does mark the intervention points. Its middle panel, "Intervention sites", draws the router, the experts and the residual addition. Each of the three points has a pin (router logits, MoE output, residual state). We take your point that this was hard to see at that size. In the revised paper, we have enlarged this panel into a diagram of one MoE layer, with the three intervention points marked on it.

Filler words and key intermediate results.

Our intervention does not disturb the MoE at a chosen word. It replaces activations with those from the other solution of the same pair, and we measure how the judgment changes. Before the point where the two solutions first differ, the model has seen exactly the same text in both, so the replacement changes nothing there. This is why we define positions by the structure of the step and not by word type. We compare two positions (§4.3). One is the point where the erroneous solution first departs from the correct one, which is the closest thing in our design to your "key intermediate results". The other is the end of the step. A split by word type would be a different study.

Reviewer_rKsGOpenReview ↗

Paper is clearer, but results are still weak

Thank you. Figure 1 is clearer now.

I understand your conclusion. I don't find it very surprising. But, it is communicated clearly and the investigation makes sense.

The results are very modest.

Replacing router logits in every layer over the whole step increases the effect in all nine models, and in GPT-OSS-120B it exceeds the residual effect

That is to say; replacing router logits in every layer produces less effect than replacing the residual in a single layer, in all models except one.

You call out the one outlier, but without a reason to think that GPT-OSS-120B behaves differently than the other models, this appears to be an outlier. I assume you aren't claiming that that model behaves differently and that you have the statistical power to provide evidence for that claim.

We replace eight tokens at two positions: the end of the step, and the point where the correct and erroneous solutions first differ (Fig. 2A). Moving to the second position shrinks the gap between the residual and the routing effect in all nine models

I read "the end of the step" to mean the last token. It seems unsurprising that moving the intervention earlier the text would produce a stronger effect, for path dependence reasons.

A residual replacement swaps more of the computation than a router replacement (§2.3), so a larger effect at one layer is expected. But the math does not force it

You repeat that the math doesn't force it, but AFAICT it does, because the residual is strictly downstream of the router. So there is a greater-than-or-equal relationship. You then bring up the claim again that swapping the routers at every position is stronger than swapping the residual at a single position, which does not seem like a fair comparison.

I continue to think that this paper should be more ambitious in what it is trying to show. For example:

  • Run the intervention at all positions n, then show how the results vary by n
  • Assess other interventions like randomizing the MoEs or choosing the worst expert
  • Present MoE interventions as a percentage of residual interventions on one or all layers
AuthorsOpenReview ↗

The results are not weak: the main result is about correctness, not effect size

Thank you for the follow-up. From your review and this comment, we think the paper has not been fully understood. We go through your points one by one, also for other readers.

  1. "The results are very modest." Both comments look at how large each effect is. The main result is about something else: whether a change follows correctness. Do activations from an erroneous step push the model toward "incorrect" more than activations from a correct step? For routing, no in all nine models at one layer, and yes in only one across all layers. For the residual state, yes in eight (Table 1, §4.4). So routing can change the judgment a lot and still not follow correctness. This result is not weak.
  2. "The math forces it." At one layer, a larger residual effect is expected, and we agree. But it is not forced. Being downstream means a residual replacement overrides the router at that layer. It does not limit how far a router replacement can move the output. And the argument is about size, so it says nothing about whether a change follows correctness.
  3. "An outlier." GPT-OSS-120B is not the only case. Its interval excludes zero, and routing and the residual effect are indistinguishable in two more models. With eight tokens at the first divergence, routing is larger in three models (§4.3).
  4. "Not a fair comparison." We agree it is not like-for-like, and the paper does not present it as one (§2.3, §4.4). It shows that a weak single-layer effect is not the whole story for routing.
  5. "The end of the step." It is the last eight tokens, not the last token (§2.3). Moving the window earlier makes the effects weaker, not stronger: the residual effect shrinks in eight of nine models. Path dependence would give the opposite.
  6. Your suggestions. The percentages are a good idea. Over the full step, routing at one layer is at most 6% of the residual effect, and routing in all layers is 25% to 134%. We will add the table to the appendix. A random control is already in the paper (§2.4, §5). Forcing the worst expert tests whether breaking the routing breaks the model, which is a different question. A sweep over every position does not fit paired traces, because the two traces are identical before the first divergence.

We hope this helps, and we are happy to explain further.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Why this paper matters

Dear Area Chair, reviewers, and readers,

Without this paper, three things would be missing.

1. A warning about how to read routing. A probe can read step correctness from the router of an MoE model, and swapping routing in every layer moves the model's judgment a lot. It is natural to conclude that routing passes correctness on to the judgment. In nine MoE models, it rarely does: the change follows correctness in no model at one layer, and in only one across all layers. People who study or steer MoE models through routing would keep drawing that conclusion.

2. A test that tells the two apart. A changed output does not show that the thing you changed carries the information. Our test compares states from a correct step and from an erroneous step on the same input. Any interpretability study can reuse it, for any property: refusal, hallucination, honesty. Reviewer RHfM describes it as addressing a "methodological limitation of activation-patching studies".

3. A negative result that others would have to find again. "Routing does not pass it on" is the kind of result that rarely gets published. Here it is measured in nine open MoE models, at three points inside the layer, with the controls needed to trust it. Without it, each group working on MoE routing starts from the opposite assumption.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.