acceptodds
Under review as a conference paper at ICLR 2027

From Routing Differences to Candidate Selection in MoE Reasoning

Abstract

Mixture-of-Experts (MoE) routing carries meaningful structure and is increasingly treated as an actionable interface. Targeted expert activation or deactivation can steer model behavior, motivating us to ask whether naturally occurring routing variation also reflects answer correctness. Under stochastic sampling, the same question can yield both correct and incorrect solutions, creating a practical need to identify which candidates are worth selecting or retaining. Using a paired-trajectory design, we compare correct and incorrect candidates for the same question across five MoE models and five reasoning datasets, connecting routing differences, cross-question readout, and candidate selection. Routing differences exceed a matched random-label baseline in 21 of 24 settings after Holm correction. Supervised probes recover correctness-related information on unseen questions, with readout strength varying across models and tasks. Adding routing to candidate scoring improves both Top-1 accuracy and Coverage@2 for four of five DAPO-Math models, including a 7.3 percentage-point Coverage@2 gain for GLM. Retention gains broadly follow relative probe advantages across models and tasks. Mean routing distance offers little additional benefit over text diversity or existing scores and lowers rare-correct retention in all nine evaluated settings. These results establish routing as a source of learnable candidate quality information whose selection value depends on both signal strength and how it is used.

Primary area:
interpretability and explainable AI
Keywords:
MoE, Routing, Empirical Analysis, 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_MC3oOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates whether natural MoE routing variation encodes information about answer correctness and aids candidate selection. Using paired trajectories, it quantifies routing gaps between correct and incorrect rollouts, trains linear probes on routing occupancy features, and evaluates two selection strategies: learned scoring and routing-distance diversity.

Critical strengths

The research question is meaningful, and the experimental coverage is relatively broad. This paper investigates whether naturally occurring routing variation in Mixture-of-Experts models is associated with answer correctness and whether routing information can improve candidate selection. Using a paired-trajectory design, the authors compare correct and incorrect candidates for the same question across five MoE models and five reasoning datasets. The evaluation connects routing difference (RD), a matched permutation baseline, linear occupancy probes and nonlinear sensitivity analyses, learned candidate scoring, and routing-distance selection.

Critical weaknesses
  1. The candidate selection baselines are relatively limited. The base scorer uses rollout length and four simple text statistics, while the routing-augmented scorer adds the routing feature vector. Comparisons with stronger baselines, such as majority voting where applicable, would make the practical benefits of routing information more convincing.

  2. The paper preserves phase-specific routing features but mainly reports aggregate routing differences and overall prediction and selection metrics, without corresponding phase-level analyses. Additional results for the early, middle, and late phases would help readers understand how correctness-related information is distributed across generation progress.

Questions for authors

See above

3 comments
AuthorsOpenReview ↗

Response to MC3o

Thank you for recognizing the significance of the research question and the breadth of the evaluation. We address both concerns with additional selection and phase-specific readout experiments.

Q1. Does routing improve candidate selection beyond a stronger baseline such as majority voting?

We compare random selection, majority voting (MV), and routing-augmented voting (MV+R) across five models on DAPO-Math under matched supervision, using the same evaluation questions and candidate pools for the three methods within each model. The method, numerical results, and paired intervals are provided in #82tE Q1.

Q2. How is correctness-related routing information distributed across early, middle, and late generation phases?

We added separate linear probes for Early, Middle, and Late routing occupancy, alongside All, which concatenates the three phases, and the original Length baseline. We retain the paper's phase definitions, original candidates and labels, question-level folds, and correct–incorrect pairs. Each probe is trained independently, with standardization and regularization selection restricted to the training folds.

We report pair-order AUC (%) below, with 95% question-level bootstrap intervals.

BenchmarkModelLengthEarlyMiddleLateAll
DAPO-MathQwen3-30B-A3B57.36 [52.48, 61.97]55.06 [50.58, 59.51]53.11 [48.43, 57.72]59.92 [55.36, 64.35]57.74 [53.26, 62.30]
DAPO-MathGLM-4.7-Flash50.33 [44.00, 56.79]50.33 [44.58, 56.29]59.00 [53.42, 64.54]58.50 [52.75, 64.25]59.50 [53.75, 65.17]
DAPO-MathQwen3-235B-A22B57.20 [51.42, 62.74]58.74 [52.97, 64.36]56.92 [51.38, 62.34]56.37 [50.47, 62.18]58.62 [53.05, 64.00]
DAPO-MathGemma-4-26B-A4B-it40.92 [32.33, 49.67]46.33 [38.67, 53.92]50.75 [42.25, 59.00]49.33 [41.33, 57.59]50.00 [42.17, 57.92]
DAPO-MathQwen3.5-35B-A3B37.91 [25.80, 50.59]51.95 [37.30, 66.41]50.98 [37.70, 64.45]50.98 [37.11, 64.84]57.03 [43.16, 71.09]
HotpotQAQwen3-30B-A3B43.85 [34.27, 53.53]51.76 [43.15, 60.58]55.29 [45.77, 64.57]56.75 [48.19, 65.32]57.26 [47.58, 66.83]
HumanEvalQwen3-30B-A3B52.72 [36.14, 69.02]53.80 [39.94, 66.85]57.34 [40.49, 73.37]56.25 [42.39, 70.11]66.85 [52.99, 79.89]

On DAPO-Math, Qwen3-235B supports above-chance readout in all three phases, Qwen3-30B has its highest AUC in Late, and GLM has similar Middle and Late scores. For Qwen3-30B on HotpotQA and HumanEval, All yields the highest AUC.

These findings extend the cross-question readout results in Table 2 by showing where correctness information is readable across generation progress. In some models, early and middle expert-use patterns already distinguish correct from incorrect candidates. The differing phase profiles reveal structure not visible in aggregate metrics and motivate retaining phase information so that a learned readout can use features from different stages according to the model and task.

Thanks for the response. Most of my concerns are addressed; I decide to increase my rating.

AuthorsOpenReview ↗

Comment by Author

Thank you for your time and the constructive feedback throughout the review process. We are glad the clarifications addressed your concerns, and we will ensure all discussed revisions are fully incorporated into the final version.

Reviewer_7QsHOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether naturally occurring MoE routing variation carries information about answer correctness and whether it helps candidate selection. Through a paired-trajectory design, it measures routing differences between correct and incorrect rollouts, trains linear probes on routing occupancy features, and evaluates two selection strategies: learned candidate scoring and routing-distance diversity.

Critical strengths
  1. The overall question arc, from detection to cross-question readout to selection, is a clean and logical decomposition.
  2. The writing is clear and easy to follow.
Critical weaknesses
  1. Limited contribution. The main experiments are linear probes on routing occupancy features, which verify a fairly intuitive conclusion: internal model states contain information related to answer correctness.
  2. The selection gains are unstable and sometimes negative. In Table 3, 4 of 11 settings show negative Coverage@2 increments.
  3. No random-selection baseline in the main comparison (Table 3), and Figure 5 shows it matters. In most cases, random selection performs no worse than SBS_B and SBRS_{BR}, and in some settings it even exceeds them.
  4. Distance-based selection in §5.2 yields no benefit and consistent harm.
  5. Taken all together, the downstream section does not demonstrate stable practical value, which leaves the paper's contribution at the detection level.
Questions for authors
  1. The number of settings differs across sections: §3 and §4 use 24 settings, while §5 uses only 11.
3 comments
AuthorsOpenReview ↗

Response to 7QsH

Q1. Is the contribution limited to the intuitive observation that internal states contain correctness information?

Natural expert occupancy supports cross-question correctness readout in multiple settings, while mean routing distance yields little consistent gain over text diversity for complementary selection (§§4–5). For the tested occupancy representation and distance rules, these results distinguish readable correctness information from geometry useful for retaining complementary answers.

The underlying association is established independently of the probes. Within-question comparisons against a matched permutation null survive Holm correction in 21 of 24 settings (Tables 1/8), with length and segmentation controls reported in Tables 9/10. Together with the selection results, this provides specific evidence about natural expert allocation and the utility of its readout and distance-based uses.

Q2. Do unstable gains, the random baseline, and negative distance results undermine the selection contribution?

Variation across settings. The original linear scorer has variable selection gains, with four negative Coverage@2 increments out of eleven in Table 3. On GLM/DAPO-Math, coverage increases from 76.67% to 84.00%, a gain of 7.33 pp over the base scorer with a paired 95% interval of [0.67, 14.00]. The direct comparison with Random below places these gains in context.

Comparison with random selection. We extend the comparison in Figure 5 by reporting Random alongside both scorers for all eleven Table 3 settings below. The values are Coverage@2 percentages computed from the original saved predictions on the same eligible mixed questions. Random is the expected coverage of two candidates sampled uniformly without replacement.

BenchmarkModelRandomBase scorer S_BRouting scorer S_BR
DAPO-MathQwen3-30B82.9482.5985.02
DAPO-MathQwen3.570.5175.0084.38
DAPO-MathGemma77.3072.0070.67
DAPO-MathGLM80.5076.6784.00
DAPO-MathQwen3-235B81.5080.3881.01
HotpotQAQwen3-30B63.9358.0667.74
HotpotQAQwen3.570.6777.0868.75
HotpotQAGemma67.4369.0963.64
HotpotQAGLM73.2168.1870.00
HotpotQAQwen3-235B73.4370.7372.36
HumanEvalGLM95.34100.0095.65

The routing scorer exceeds Random in five settings and falls below it in six. For GLM-DAPO, the routing scorer exceeds Random by 3.50 pp.

The additional DAPO-Math experiment in #82tE Q1 tests routing combined with majority voting under matched supervision on the same evaluation questions and candidate pools. On Qwen3-30B/DAPO-Math, majority voting exceeds Random by 13.32 pp [10.37, 16.21], and adding routing raises Top-1 accuracy from 81.36% to 83.55%, a further 2.19 pp [0.44, 4.39]. This provides a concrete instance of selection value beyond an informative answer-agreement baseline using a different rule from the original linear scorer.

Scope of the distance result. The description “no benefit and consistent harm” is too broad for §5.2 as a whole. Tables 4/20 report a median Coverage@4 gain of 2.47 pp for Routing-Diverse over Random across 24 settings, but a median difference of −0.32 pp relative to Text-Diverse. The consistent decline in all nine settings concerns the specific rare-correct retention experiment under a shared synthetic score provider. The finding is that mean routing distance does not provide reliable additional value beyond text diversity or available scores and can discard rare correct candidates under the tested rule.

Q3. Why do the numbers of evaluated settings differ between §§3–4 and §5?

The eleven settings correspond to learned candidate selection in §5.1, rather than to §5 as a whole. This evaluation retains complete five-model comparisons on DAPO-Math and HotpotQA, supplemented by GLM-HumanEval as a code setting. DAPO-Math and HotpotQA provide 32–247 and 48–123 mixed questions per model, respectively, whereas the available MATH-500 and GSM8K settings provide only 3–31 and 7–38. On HumanEval, GLM has 46 mixed questions, compared with 2–23 for the other models. The association and readout analyses in §§3–4 and the rule-based distance comparisons in §5.2 still cover 24 settings.

Both the original and supplementary results show unstable gains, so I maintain my score.

Thank you for considering our rebuttal and additional results. As discussed, we acknowledge the variability in selection gains, while maintaining that characterizing both the utility and limitations of routing information is a substantive contribution. Could you clarify any remaining concerns that we could address through further discussion? In light of the additional evidence and clarified scope, we respectfully ask you to reconsider your score.

Reviewer_82tEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether the expert routing of an MoE model, as it naturally varies across sampled solutions, says anything about which solutions are correct. The authors compare correct and incorrect rollouts of the same question, show that their expert-occupancy patterns differ more than a label-permutation baseline would predict in most model–dataset settings, train probes that read correctness from routing on held-out questions, and finally plug routing features into a candidate scorer and a pair selector. The headline is that routing helps candidate scoring for most models on DAPO-Math, while using routing diversity to pick complementary candidates does not help and can hurt.

Critical strengths
  1. The core association finding is carefully established. Comparing correct and incorrect rollouts within the same question, checking against a matched permutation null, correcting for multiple settings, and controlling for length and generation phase together make me fairly convinced that routing really does differ with correctness, and that this is not just a length artifact. This is a cleaner result than simply reporting raw routing differences.
  2. Honest reporting of where the idea does not work. The paper explicitly tests the intuitive idea that routing distance could be used to select diverse candidates, and shows that it does not help. Together with the cross-task results, this gives a fairly frank picture of where the signal does and does not transfer, which I found more useful than a uniformly positive story.
Critical weaknesses
  1. The evidence does not show that the useful signal is specific to routing. The positive selection result is measured against a deliberately weak baseline (length plus a few surface text statistics). When a stronger pair-selection baseline is used, adding routing mostly stops helping. In the paper's own representation comparison, router logits and occupancy do no better than hidden states or even a random projection of hidden states, and none of them beats length alone. A simpler explanation is that routing is one more view of the hidden state, and any correctness information it carries is generic rather than routing-specific. If so, the conclusion that "routing is a source of learnable candidate quality information" is true but not very informative, since the same holds for almost any internal representation.
  2. The correlation may reflect what the solution does rather than how good it is. Correct and incorrect solutions to the same question often take different solution strategies, and different strategies may plausibly use different experts. The length and phase controls do not address this. The paper's interpretation, that routing reflects candidate quality, needs some argument or experiment that separates it from "routing reflects which approach was taken, and some approaches happen to be right more often for this question." This matters for the paper's motivation: a content proxy would not be expected to transfer across tasks, which is roughly what the cross-task results show.
Questions for authors
  1. Can you show that routing improves selection on top of a non-routing quality signal that is itself informative on these splits (e.g., a hidden-state correctness probe or an answer-agreement feature), under matched supervision and question splits? Relatedly, how do you interpret the result that occupancy does not outperform a random projection of the hidden state?
  2. Is there any evidence that the routing difference reflects quality rather than solution strategy? For example, does it persist when comparing correct and incorrect rollouts that follow the same approach, or that agree up to a late point and then diverge?
  3. What does selection performance look like on the full question set, including questions where all candidates are right, or all are wrong, and on a DAPO-Math sample that was not screened with Qwen3-30B?
3 comments
AuthorsOpenReview ↗

Response for 82tE

Q1. Does routing add value beyond an informative non-routing quality signal?

We compare Random, majority voting (MV), and MV+R across five DAPO-Math models on identical evaluation questions and candidate pools. MV selects the most frequent answer. MV+R combines votes with within-group routing coherence when the runner-up has at least 80% of the leading answer's votes and both groups have at least two candidates, allowing narrow margins to be overturned. Coherence uses PCA-reduced expert-use frequencies adjusted for text similarity and length. Transformations are fitted on training questions only. Supervision is matched as neither method uses correctness labels to determine decision rules or parameters, which are fixed before evaluation.

Accuracy is in %, with gains in pp and paired 95% intervals. This experiment tests answer-agreement fusion alongside the distance rules in §5.2.

ModelRandomMVMV+RMV − Random [95% CI]MV+R − MV [95% CI]
Qwen3-30B-A3B68.0481.3683.55+13.32 [10.37, 16.21]+2.19 [+0.44, +4.39]
GLM-4.7-Flash69.5374.8077.24+5.27 [1.92, 8.54]+2.44 [0.00, +5.69]
Qwen3-235B-A22B69.6479.7182.61+10.07 [6.67, 13.53]+2.90 [−0.72, +7.25]
Gemma-4-26B-A4B-it64.6069.5374.22+4.93 [−0.21, 9.93]+4.69 [0.00, +10.94]
Qwen3.5-35B-A3B52.2260.4960.49+8.28 [1.45, 14.96]0.00 [0.00, 0.00]

On Qwen3-30B, MV gains 13.32 pp over Random and routing adds 2.19 pp, demonstrating an increment over informative answer agreement. Four models have positive incremental point estimates and one is unchanged, with the clearest evidence on Qwen3-30B.

Q2. How should the random-projection and length comparisons be interpreted?

The appendix reports an occupancy-minus-random-projection AUC difference of −1.25 pp [−7.99, 4.87]. All four representation-augmented predictors have lower AUC point estimates than length alone. This comparison is consistent with correctness information being accessible through multiple representations, without establishing an advantage or unique correctness information for routing occupancy.

Tables 1–4 establish permutation-calibrated association under length and phase controls, cross-question readout, and contrasting results for learned scoring and routing-distance selection. These findings distinguish readable correctness information from geometry useful for complementary selection. Smaller gains under the stronger pair-selection baseline further show that incremental value depends on the baseline and decision rule. Q1 tests a further use with answer agreement.

Q3. Does the association persist within the same solution strategy?

We compare correct and incorrect candidates within the same question and annotated coarse-grained strategy.

Model / taskSupportObserved RDPermutation RDΔRD [95% CI]One-sided p
Qwen3-30B / DAPO-Math13 questions, 13 groups, 191 candidates1.14340.9096+0.2338 [+0.0833, +0.4010]0.0010

RD units are 10⁻³. ΔRD subtracts the permutation baseline. From 20 fixed sampled questions, we exclude missing-figure or strategy-ambiguous cases and retain groups with at least two correct and two incorrect candidates. The paper's RD estimator averages three phases with equal question weights. Labels are permuted within question–strategy groups, preserving correct/incorrect counts. Intervals use question bootstrap.

The positive excess RD provides evidence that differences between coarse-grained strategies alone do not fully explain the association in this sample, extending the paper's within-question result to within-strategy comparisons.

The paper trains and tests readouts separately within each task. These results measure task-dependent readout strength, while cross-task transfer requires evaluating the same trained readout on another task.

Q4. How do mixed-only evaluation and Qwen screening affect the results?

For fixed candidate pools and prescribed selection sizes, all-correct and all-incorrect questions give identical Top-1 and Coverage@k outcomes across selectors. With equal question weights, Δall=(Nmixed/Nall)Δmixed\Delta_{\mathrm{all}}=(N_{\mathrm{mixed}}/N_{\mathrm{all}})\Delta_{\mathrm{mixed}}. Including these questions scales the gain and preserves its direction.

Qwen screening concentrates collection on questions with selection headroom (§2.3, Appendix A.2), with mixed membership determined per model. The estimates characterize this screened pool. Quantifying unscreened DAPO performance requires an independently sampled evaluation.

Thank you for the response.

Q1. The MV comparison resolves part of our concern: routing adds a measurable increment over answer agreement on Qwen3-30B. For the other models the gain is not clearly distinguishable from zero. It also does not resolve our main concern, because MV+R is a new unsupervised rule, not the learned routing score the paper's selection claims rest on. In Table 16, adding routing scores to the stronger P0 baseline lowers Coverage@2 for three of five DAPO models, and the response does not explain this.

Q2. This resolves our question about the random-projection result. We now understand that the paper supports readable correctness information, not routing-specific information. We recommend that the revision state this explicitly.

Q3. The within-strategy analysis resolves part of our concern, since the association persists after controlling for coarse strategy. It covers one model and one task, with 13 questions after manual exclusions, so it does not yet generalize. The annotation protocol is also unclear.

Q4. This clarifies that including all-correct and all-incorrect questions only scales the gain down. It does not resolve how large the overall gain is, since the mixed-question ratios are missing. It also does not resolve the effect of the Qwen3-30B screening, because no unscreened evaluation was added.

Overall, the response clarifies the paper's claims and adds some evidence, but the incremental value of routing scores over the stronger baseline and the stability of the gains on unscreened pools remain open. I maintain the rating of 2.

Q1. How do the Table 16 results affect the value of learned routing scores?

Table 16 shows that the current score-fusion method provides no consistent increment over P0P_0, with Coverage@2 decreasing in three of five DAPO models. Here, P1P_1 adds summaries of scores trained for individual candidate correctness to a selector trained for joint failure. These results show that improvements in individual candidate ranking do not necessarily translate into additional gains over a stronger selector trained directly for joint failure.

This is compatible with Table 3, where routing improves candidate ranking over SBS_B for four of five DAPO models. The new MV+R results provide complementary evidence for combining routing with answer agreement. Together, these findings motivate examining how routing representations and fusion rules align with the selection objective.

We will clarify that the demonstrated gains concern specific selection methods and baselines. The distinction between readable correctness information and effective selection remains central to the paper, alongside the negative result for mean routing-distance diversity.

Q2. Readable correctness information and routing specificity.

Thank you for confirming this clarification. We will make the distinction between readable correctness information and routing-specific information explicit in the revised manuscript.

Q3. How were strategies annotated and the 13 questions selected?

We will document the annotation protocol and selection steps. From the eligible pool, we randomly sampled 20 questions. AI-assisted grouping used the core mathematical formulation and solution approach, keeping calculation and execution errors within the same strategy. Annotation materials omitted correctness labels, routing information, and standalone answer sections. The authors confirmed and froze the groups before joining correctness labels or computing RD. Five questions lacked necessary figures, and two additional candidates had ambiguous strategies. Applying the requirement of at least two correct and two incorrect candidates per group then yielded 13 groups across 13 questions.

Strategy checking requires question-specific inspection of reasoning trajectories, so this analysis focused on Qwen3-30B/DAPO-Math. We will state its model, task, and sample scope explicitly in the revision.

Q4.1. How large are the gains when all-correct and all-incorrect questions are included?

We extend the original held-out results to the archived screened DAPO pools by including all-correct and all-incorrect questions, on which all selectors receive identical outcomes. Original mixed-question predictions and candidate sets are retained. Scores are percentages and gains are percentage points, with 95% question-paired bootstrap intervals.

ModelMetricFull-pool SBS_BFull-pool SBRS_{BR}Mixed gainFull-pool gain [95% CI]
Qwen3-30B-A3BTop-176.4678.64+3.64+2.18 [−1.46, +5.83]
Qwen3-30B-A3BCoverage@285.4086.86+2.43+1.46 [−1.95, +4.87]
GLM-4.7-FlashTop-179.4082.63+8.67+3.23 [−0.25, +6.70]
GLM-4.7-FlashCoverage@284.7587.50+7.33+2.75 [+0.50, +5.25]

Evaluation uses judgeable candidates, with eligibility determined separately for Top-1 and Coverage@2. Mixed questions account for 59.95% and 60.10% of the eligible Qwen3-30B pools for Top-1 and Coverage@2, respectively, and 37.22% and 37.50% for GLM. Full-pool P1−P0P_1-P_0 Coverage@2 increments are −0.73 pp [−2.43, +0.73] on Qwen3-30B and +0.25 pp [−1.75, +2.25] on GLM.

The trends agree with Tables 3 and 16 for both models. Candidate-scoring gains remain positive, while the stronger pair selector shows no consistent increment. Including homogeneous-outcome questions scales the gains by the mixed-question share without changing their direction.


Q4.2. Does the selection benefit persist without Qwen-based screening?

The Qwen3-30B screening was used to identify questions with both successful and unsuccessful solutions, providing informative comparisons under a limited generation budget. The full-pool results in Q4.1 quantify gains within the archived screened sample. Their applicability to an independently sampled DAPO pool remains to be established. We will make this sampling condition explicit in the revised selection claims.

We would be grateful if you would reconsider your rating in light of the additional evidence and clarifications provided.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.