- 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
- 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.
- 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.
- 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
- 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.
- 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.
- 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
- Can the authors identify specific experts or routing decisions that causally influence verification judgments, rather than only comparing aggregate interventions across interfaces and layers?
- 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?
- 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
Thanks again for your review. Your suggestions led to new analyses in the paper. Here is how the revision addresses each point you raised.
-
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.
-
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).
-
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.
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.
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.
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.