acceptodds
Under review as a conference paper at ICLR 2027

Why This Expert? A Feature-Level Analysis of Routing in Mixture-of-Experts with Sparse Autoencoders

Abstract

Understanding Mixture-of-Experts (MoE) routing requires identifying which features of a token's representation favor selected experts over their competitors. We present a framework for explaining individual routing decisions through sparse autoencoder (SAE) features. Our framework decomposes routing logits into feature contributions, a decoder bias, and an explicit reconstruction residual, allowing us to quantify how features contribute to routing logits and expert selection. We formulate the search for a minimum-cardinality feature subset preserving the selected expert set as a mixed-integer linear program, distinguishing explanations of reconstructed routing from explanations of original routing conditional on the residual. Across four MoE models, we examine routing fidelity, feature-expert affinities, and the number of features required to preserve selection. On OLMoE, an average of 8.79 of 32 active features suffices to preserve the reconstructed expert set; preserving the original set with the residual fixed requires 9.57 features. SAE reconstructions recover 81–88 of selected experts across OLMoE layers and 73–86 across DeepSeek-V2-Lite layers. Case studies connect these quantitative contributions to interpretable token and contextual features, while steering experiments demonstrate control over expert selection. Together, these results provide a quantitative account of routing through compact feature subsets while explicitly separating their contributions from information the SAE leaves unexplained.

Primary area:
interpretability and explainable AI
Keywords:
Mixture-of-Experts, sparse autoencoders, routing mechanism, interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zKAUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes a feature-level framework for interpreting routing decisions in Mixture-of-Experts (MoE) models using Sparse Autoencoders (SAEs). The authors decompose routing logits into additive contributions from SAE features, a decoder bias, and a reconstruction residual, enabling quantitative analysis of how individual features influence expert selection. They further formulate a mixed-integer linear programming (MILP) problem to identify the minimum subset of active SAE features sufficient to preserve the selected experts. Experiments across four MoE models (OLMoE, DeepSeek-V2-Lite, gpt-oss-20B, and Qwen3-30B-A3B) investigate feature–expert affinities, routing reconstruction fidelity, and feature sparsity. The results show that SAE reconstructions recover a substantial proportion of selected experts and that relatively small feature subsets can preserve routing decisions. The authors also provide qualitative case studies linking SAE features to interpretable linguistic concepts and steering experiments demonstrating that selected feature directions can influence expert routing.

Critical strengths
  1. An explicit and transparent decomposition. The paper offers an easily verifiable decomposition of routing logits into SAE features, a bias term, and a reconstruction residual. Keeping the residual explicit is valuable because it avoids falsely claiming that SAE features perfectly reconstruct the input.

  2. Broad empirical evaluation and a well-defined optimization objective. The experiments cover four MoE architectures, and the minimum-cardinality feature subset formulation provides a concrete way of quantifying routing sparsity. These are useful empirical tools, even if their explanatory significance is not yet established.

Critical weaknesses
  1. The central attribution methodology lacks a sound justification for explaining routing decisions.

Although the additive decomposition of routing logits is mathematically correct, the proposed metrics are not necessarily aligned with the top-K selection mechanism. In particular, a positive or negative contribution to an individual expert's logit does not imply that a feature promotes or suppresses that expert relative to its competitors. Similarly, the L1 norm of feature-contributed logits can be arbitrarily large while having no effect on expert selection. These are fundamental problems because the paper uses these quantities to characterize influential features and interprets their semantic meanings as explanations of routing decisions. The authors need to establish why their attribution measures are valid, preferably using relative routing margins and causal interventions.

  1. The experimental evidence does not establish faithful explanations, and the novelty relative to prior work is limited.

The experiments demonstrate routing reconstruction and the ability to alter routing logits using selected feature directions, but neither establishes that the identified SAE features correspond to the semantic causes of routing decisions. The work lacks controls against alternative decompositions, quantitative validation of the semantic explanations, and an evaluation of their stability. Moreover, the central feature–expert affinity analysis overlaps with RouterInterp [1], which already quantifies the alignment between SAE features and router vectors. The paper's distinct contribution lies primarily in activation-weighted decomposition and minimum-subset optimization, whose explanatory usefulness remains insufficiently demonstrated.

[1] Lasy, I., Cai, N. Y., & Ayonrinde, K. (2026). RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing. In Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), Proceedings of Machine Learning Research (PMLR), Vol. 306, pp. 62858–62883.

Questions for authors
  1. Why are signed individual routing-logit contributions appropriate for identifying promotive and suppressive features? Since selection is determined by relative logits, how do the authors handle features that reduce a target expert's logit but reduce competing experts' logits even more?

  2. Why is the L1 norm a meaningful measure of routing importance? Could the authors replace it with a routing-margin-based metric and compare the resulting feature rankings and interpretations?

  3. How do the authors validate the faithfulness of their semantic explanations? Can they demonstrate that removing or intervening on a feature associated with a particular concept changes the predicted expert selection in held-out contexts, rather than merely showing that adding its router-projected direction changes logits?

  4. How robust are the minimum-cardinality feature subsets? Do independently trained SAEs yield similar subsets or semantic explanations for identical routing decisions? How do the results compare with random or non-SAE decompositions of comparable reconstruction quality?

  5. What is the precise novelty over RouterInterp? In particular, how does the paper justify its first-to-quantify-affinity claim given that Lasy et al. already explicitly calculate expert–feature affinities?

3 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

Weakness 1

We adopt the L1-norm as a proxy to measure the “contribution” of a feature to routing logits of all experts, regardless of whether they are selected or not. It does not directly measure the impact of a feature on the routing decisions.

To the best of our knowledge, we imagine a causal way to determine the “contribution” of a feature to the decisions: measuring the ranking variance of top-K experts or all experts when the feature is removed. Such a metric relies on the ambient features; in other words, the contribution depends not only on the feature itself but also on other features from this viewpoint. We want to find a metric that can measure the “contribution” of a feature and isolate it from the influence of other features. We adopt L1-norm because it is simple, and essentially it is a sum of the routing logits assigned to each expert by the feature. It is also deterministic – even if we do not know what other features are, at least we can always use it as a proxy to estimate its influence.

Here we use Spearman’s rank correlation coefficient ρ\rho to measure the correlation between a feature’s L1 norm and its causal influence (i.e., influence caused by ablating the feature) on 5M active features. We use the following four metrics as a proxy of causal influence: (1) Absolute Variation of ranks of top-k experts: AV topk_rank = ∑e∈E∣Δrank(e)∣\sum_{e \in E} |\Delta rank(e)|, where ee’s are selected experts; (2) Absolute Variation of logits of top-k experts: AV topk_logit = ∑e∈E∣Δlogit(e)∣\sum_{e \in E} |\Delta logit(e)|. Similarly, we have (3) the AV of ranks of all experts AV all_rank and (4) the AV of logits of all experts AV all_logit, this is mathematically equal to the definition of L1-norm of routing logits assigned by the feature.

Model, LayerAV topk_rankAV topk_logitAV all_rankAV all_logit
OLMoE, Layer 00.6580.9080.9381.000
OLMoE, Layer 70.4020.7470.6891.000
OLMoE, Layer 150.3950.7540.7211.000
DeepSeek-V2-Lite, Layer 10.6030.9280.9441.000
DeepSeek-V2-Lite, Layer 130.4070.8110.7851.000
DeepSeek-V2-Lite, Layer 260.3680.7730.6911.000

The results indicate L1-norm has a high correlation with the absolute variation of ranks of all experts and the logits of top-K experts. Although it may not be a suitable proxy for measuring how a feature influences the ranks of Top-K experts, this limitation does not affect our analysis, as we are more interested in how features contribute to the routing logits of experts in this work, regardless of whether they are selected or not.

Weakness 2

We address this issue in two parts.

Regarding semantic causes of routing decisions with SAE features: SAE is a tool for linearly decomposing a given representation into feature vectors. We identify SAE feature explanations based on their contexts, whose reliability has been demonstrated in prior work, such as [1]. We have applied feature steering to show that these features can causally trigger experts to get a higher/lower rank.

We provide the results of feature ablation as follows. We zero out the identified promotive/suppressive features when they are active and measure the resulting difference of L1-E9 of OLMoE.

After ablating the promotive feature "a", the average rank of L1-E9 increases and its logit has a large decrease, indicating that the feature has a prominent promotive influence on L1-E9. Ablating the suppressive feature "and" yields a decrease in the rank and an increase in the logit of L1-E9. We note that ablating the promotive feature "the"-12162 yields a relatively small increase in the rank while the logit change is large, which hints that there can be some concurrent features that overshadow its contribution to the rank of L1-E9.

Actually, we identify two features related to "the" (Features 12162 and 29054). Their abstract is in Appendix E.1, Table 7. Previously, we only focused on Feature 12162 in the main text. These two features are typically active concurrently. (Occurrence rate: Feature 12162=4.07%, Feature 29054=3.89%, both=3.55%) We repeat the feature ablation experiment and find that ablating either feature yields a relatively small change in the rank of L1-E9. However, ablating both features jointly results in a greater rank change than ablating either one individually. The results suggest that these two features typically collaborate in influencing the rank of L1-E9 and provide a reason for why the rank change is relatively small for ablating either feature.

Feature(s)Average rank before ablationAverage rank after ablationAverage $\Delta$ RankAverage $\Delta$ logit
"a"4.4512.938.48-0.72
"the"-121623.365.011.65-0.55
"and''48.4537.31-11.140.32
a random feature34.4536.742.29-0.02
"the"-290543.544.080.54-0.43
"the"-12162 & "the"-290541.747.435.69-1.05
AuthorsOpenReview ↗

Official Comment by Authors (2/3)

Weakness 2 (cont.)

Regarding RouterInterp: The “feature-expert affinity analysis” in our work is totally different from RouterInterp. (NOTE: the following content is the same as General response to all reviewers 2.)

In this work, we are dissecting routing decisions by studying how features assign logits to each expert, no matter whether the experts are selected or not.

In RouterInterp (Lasy et al.), "A feature ff is ρ\rho-useful for expert EiE_i if the expected activation of ff is higher when expert EiE_i is selected than when it is not". ρ\rho-usefulness is defined as ρ(f,Ei)=E[f(x)∣Ei∈T(x)]–E[f(x)∣Ei∉T(x)]\rho(f, E_i) = E[f(x) | E_i \in \mathcal{T}(x)] – E[f(x) | E_i \notin \mathcal{T}(x)], where T(x)\mathcal{T}(x) is the set of selected experts for input xx.

We discuss the difference here:

  1. Essentially, Lasy et al. use the difference between the expected activations of a feature as a metric. It is an overall metric, not a metric for a specific scenario (it cannot be case-by-case): it cannot tell, given that feature ff has an activation of aa, what the contribution of feature ff to expert ii is. However, given the SAE is trained on MoE layer input, we formulate that as a∗(fTe)a * (f^T e), where aa is the feature activation, ff is the SAE feature vector, and ee is the expert vector in the router matrix WrW_r.
  2. Although \rho-usefulness is a causal metric, it is still indirect – it uses feature activation as a proxy to measure “usefulness”. As we formulated above, feature activation is only a factor of assigned logits. Imagine that two features may assign all experts the same logits without equal feature activation. They theoretically have the same influence on the routing logits and routing decisions, but they get a different \rho-usefulness just because they have a different feature activation.
  3. ρ\rho-usefulness relies on a threshold: whether expert ii is selected or not. However, feature ii may not be the one that decides whether expert ii is selected/unselected. ρ\rho-usefulness cannot eliminate the influence of other coexisting features when studying the “usefulness” of feature ff to expert ii, even though they use expectation to alleviate this issue.
  4. ρ\rho-usefulness is rank-related. In other words, it is not logit-sensitive. Imagine two features that have the same “usefulness” to Expert ii, they may still assign similar but different logits to the expert. In our work, we aim to measure the logits assigned by a feature to an expert ii. ρ\rho-usefulness is not a suitable metric for our task.
  5. One can cheaply predict whether a feature promotes or suppresses experts with our method arithmetically (I(fTe>0)I(f^T e > 0)); the feature activation is just a scale factor that influences the magnitude of promotion/suppression, while the dot product fTef^T e is an inherent characteristic of the feature and the expert. However, one has to run the models to obtain ρ\rho-usefulness for each feature and each expert to get such predictions.
  6. One can quantify the interaction among features mathematically with our proposed method, i.e., we can learn how active features f1,f2,…fnf_1, f_2, …f_n jointly contribute to the routing logits of experts. In this way, we can decompose the logit of an expert as they were assigned by these features. However, ρ\rho-usefulness does not support such an analysis.
  7. RouterInterp identifies "most predictive" features for an expert to be selected. However, there can be multiple features that contribute to the routing, or there may be no “most predictive” feature at all. In such scenarios, we need to analyze the "competition" among features. Our method is straightforward for those scenarios.

Question 1

Promotive and suppressive are descriptive terms for features that have logits of (relatively large) positive or negative magnitude. They are not exclusive to specific expert(s). One feature can assign large negative logits to different experts; in such a situation, this feature is regarded as a suppressive feature for those experts. We aim to demonstrate that routing logits can be decomposed, and we can study how features’ ``attitude’’ towards experts from that perspective. We do not indicate that a promotive/suppressive feature for a specific expert guarantees the expert to be selected/unselected.

Question 2

Please refer to Weakness 1.

AuthorsOpenReview ↗

Official Comment by Authors (3/3)

Question 3

"adding its router-projected direction changes logits" is essentially equivalent to injecting the feature into the MoE layer input. Given a steering feature vector $f$, a steered MoE input vector $m$, and the router weight matrix $W_r$, the routing logits $W_r(f + m) = W_r f + W_r m$. We add $W_r f$ to $W_r m$, which is equivalent to directly adding $f$ to $m$ in terms of their effect on the routing. Hence, it is still a kind of feature steering.

A feature ablation analysis is provided in Weakness 2.

Question 4

Robustness of minimum-cardinality feature subset: We test on SAEs trained on three different seeds, and the variance of each result is less than 0.05.

We speculate that different SAEs may yield different explanations to a certain degree for identical routing decisions, given that SAEs may learn different features from the same training data [2]. We are trying to run a test for this question now.

We show that our methods have better recall@K for the binary classification (selected/unselected) for experts with a logistic regression trained on SAE latent, as proposed in RouterInterp(Lasy et al.). They show that logistic regression performs better than other traditional methods, such as PCA-based probes.

Recall (model, layer)OursLogistic regression on SAE latent
OLMoE, Layer 00.8460.786
OLMoE, Layer 70.8400.795
OLMoE, Layer 150.8710.834
DeepSeek-v2-Lite, Layer 10.8610.815
DeepSeek-v2-Lite, Layer 130.7650.692
DeepSeek-v2-Lite, Layer 260.7890.721

Question 5

Essentially, we use linear decomposition in the sparse autoencoder space to quantify feature(s) contribution to the routing logits. While RouterInterp identifies features "most predictive" for an expert to be selected with the metric $\rho$-usefulness.

The novelty includes:

  1. Our method can study feature contribution to the routing logits for each token (case-by-case); while RouterInterp is a statistical method for finding features "most predictive" for an expert to be selected; they cannot have such a fine-grained study.

  2. Our method actually creates a linear model for how routing logits are decided, so one can study any combinations of features in an arithmetic and cheap way. However, in RouterInterp, one needs to run models to collect data for specific feature combinations; and some feature combinations may be rare that they are difficult to collect.

  3. (Duplicate of response to Weakness 2 (7)): RouterInterp identifies "most predictive" features for an expert to be selected. However, there can be multiple features that contribute to the routing, or there may be \textbf{no} “most predictive” feature at all. In such scenarios, we need to analyze the “competition” among features. Our method is straightforward for those scenarios.

Regarding affinity: Actually, RouterInterp does not call the metric (ρ\rho-usefulness) an affinity. It is a statistical metric to measure the difference between the expected activations of a feature when a given expert is selected/unselected. The affinity we call is essentially the logit assigned from a feature to an expert. It is not a statistical metric; it is case-by-case (when the token is different, the affinity is changed). We have argued why ρ\rho-usefulness is not a suitable metric when studying the features’ contributions to routing logits in the response to Weakness 2.

[1] Gonc¸alo Santos Paulo, Alex Troy Mallen, Caden Juang, and Nora Belrose. Automatically interpreting millions of features in large language models. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=EemtbhJOXc.

[2] Gonc¸alo Paulo and Nora Belrose. Sparse autoencoders trained on the same data learn different features. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=EjInprGpk9.

Reviewer_rLU4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

They use SAEs to split MoE routing logits into feature contributions, decoder bias, and reconstruction residual. A MILP then finds the smallest feature subset that keeps the same expert selection. They test on four MoE models and show that some feature interventions shift which experts get picked.

Critical strengths

The decomposition is clean, the MILP gives a concrete minimal-subset result, and the suppressive features are a useful addition. Four models is more than one architecture.

Critical weaknesses

There are no real baselines. No alternative SAE configs, no simpler feature-selection method, no prior MoE routing interpretation work. I can’t tell what the framework adds over a trivial alternative. Would a simple top-contribution selection or a differently configured SAE find subsets of similar size?

The explanations are never tested. The qualitative analysis covers a handful of experts and tokens. Do the feature explanations predict expert selection on held-out inputs, across many experts? Until they do, this is anecdotal.

The minimal-subset result is measured against the reconstructed Top-K, not the model’s actual routing. With a residual term in the decomposition, those can differ. How often does the reconstructed selection disagree with the real one?

The intervention results have no null control. Do random or size-matched feature interventions shift expert selection as much as the identified features do?

Robustness is thin. The SAEs are trained and tested on the same distribution, with little on seeds, SAE configurations, or distribution shift. Do the identified features and subsets survive a different seed or a different dataset?

Questions for authors

Please answer the questions in Weaknesses, starting with the baselines and the held-out predictive test.

2 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

We treat each paragraph as a weakness and answer them here.

Weakness 1

No real baselines: We agree that there is a lack of real baselines for an explanation of a routing. It is difficult to define a “real” baseline: We may (partially) attribute a routing decision to a set of features, but we do not have a ground truth on what features exactly contribute to the routing decision and the quantification of that. In this work, we attempt to provide a decomposition method to linearly attribute routing logits to features, which can be regarded as a possible way to analyze and interpret routing.

No alternative SAE configs: Our work focuses on dissecting routing logits and attributing them to SAE features, rather than optimizing SAE performance for interpretation. In our preliminary experiments, we explored several SAE architectures and hyperparameter configurations. We adopted Top-K SAE because it maintains a manageable number of active features, facilitating our subsequent analysis. Vanilla SAE and JumpReLU SAE tend to have a large number of active features and suffer from the dead-feature issue, hindering the analysis. We also tried adjusting some hyperparameters, e.g., the expansion factor, and tried adjusting the loss function of SAE. However, these modifications yielded slight improvements, and thus we do not discuss them in the paper.

No simpler feature-selection method: MILP is mathematically guaranteed to find the minimum subset size, given the problems we design in the paper have feasible solutions. However, a simple top-contribution selection cannot find subsets of similar size. We now provide the results of two heuristic selection methods (feature L1-norm, feature activation), one greedy algorithm (select one active feature that maximizes the minimal logits gap between the selected experts and the unselected experts at each time), and one random baseline for searching subsets. For the random baseline, we randomly select features until the selected feature set can recover the specific target, and repeat this process 50 times for each sample to get the minimal results.

The following table shows the size of the feature subset identified by different methods that maintain the ground-truth Top-K expert set without order (gt_set) or with order (gt_order), and the reconstruction Top-K expert set without order (recon_set) or with order (recon_order). MILP achieves better performance than the compared methods. The results are obtained on 10000 randomly selected tokens.

OLMoE (Layer 0)MILPHeuristic (L1-norm)Heuristic (Activation)GreedyRandom (Best results of 50 trials)
gt_set9.1722.0524.5517.5320.21
gt_order13.7628.0329.4130.4127.67
recon_set9.2122.4124.9018.9020.67
recon_order13.8528.0929.5030.9227.84
OLMoE (Layer 7)MILPHeuristic (L1-norm)Heuristic (Activation)GreedyRandom (Best results of 50 trials)
gt_set10.7921.7121.9812.2017.81
gt_order15.0329.0229.3722.8525.67
recon_set9.5121.7221.9011.2316.68
recon_order13.2128.9829.2522.7925.09
OLMoE (Layer 15)MILPHeuristic (L1-norm)Heuristic (Activation)GreedyRandom (Best results of 50 trials)
gt_set8.6118.5318.859.8314.87
gt_order13.4828.3128.7922.1524.73
recon_set7.5418.3418.559.2113.84
recon_order11.8627.9728.4321.7823.88

We now train SAEs with a larger vocabulary size (expansion factor=32, previously we used 16 in the paper) and top-32 features for OLMoE, and find that they yield a similar size of feature subset:

OLMoEgt_setgt_orderrecon_setrecon_order
Layer 0+0.50+0.57+0.58+0.59
Layer 7+0.40+0.39+0.36+0.39
Layer 15+0.42+0.50+0.44+0.54

No prior MoE routing interpretation work: There is little work on MoE routing interpretation directly related to this work. We consider RouterInterp [1] to be the most similar work to our work, as both use SAE to find features; however, their motivations are still quite different.

[1] Ilya Lasy, Nora Yinuo Cai, and Kola Ayonrinde. RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id=rDNCWfRd69.

Weakness 2

We examine the selection rate of L1-E9 of OLMoE on a held-out set of 5M tokens and provide its selection rate when the specific feature is active as follows:

FeatureSelection rate of L1-E9
"a'' (promotive)0.984
"the'' (promotive)0.963
"and'' (suppressive)0.008

Note that a random selection yields a selection rate of 8 / 64 = 0.125 . Hence, these features can be used to predict whether an expert is selected or not.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

Weakness 3

If you mean the exact reconstruction of all the selected experts (the Top-K experts): We now include the exact recovery rate in Appendix I.

We also provide the distribution of the number of recovered top-K experts below:

#recovered experts@Top-8 (OLMoE)<=2345678
Layer 0<0.0010.0010.0120.0690.2580.4640.196
Layer 7<0.0010.0010.0100.0700.2780.4710.170
Layer 15<0.0010.0010.0100.0500.1860.4610.291
#recovered experts@Top-6 (DeepSeek-V2-Lite)0123456
Layer 1<0.001<0.0010.0030.0240.1400.4710.361
Layer 13<0.0010.0020.0180.0930.3020.4420.143
Layer 26<0.0010.0020.0160.0800.2560.4380.207

We emphasize that an exact recovery (all top-K experts are recovered) is usually difficult and possibly not so meaningful because the borderline experts may be swapped with other experts without a significant influence on the model output. The exact recovery rate is low because the middle-ranked experts, even those borderline experts (e.g., ranked K-1 or K), tend to have a close routing logit distribution with their neighbor experts, compared with the top-ranked or bottom-ranked experts (Page 9, Fig. 7), making them hard to distinguish from their neighbors.

Weakness 4

We provided results of random feature interventions in the paper. For example, in Figure 3, we steer routing decisions with promotive features and random features.

Weakness 5

The results of reconstruction fidelity and minimal subset (Appendix J, Table 13) are obtained on SAEs trained with three different seeds. Their variance is very small, and thus it is not obvious in most figures; however, the feature death rate is an exception (Appendix G, Figure 16c). We are trying to run a related experiment to observe whether or how different seeds change the captured features. We speculate that the identified features and subsets may change given that SAE can learn different knowledge given the same training data [2].

[2] Gonc¸alo Paulo and Nora Belrose. Sparse autoencoders trained on the same data learn different features. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=EjInprGpk9.

Reviewer_sv2VOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper attributes MoE routing logits to Top-K SAE features, the decoder bias, and the reconstruction residual, then uses a mixed-integer program to find a small feature subset that preserves the selected experts. The decomposition follows from a linear router, and the experiments do not yet show that the new machinery explains routing beyond existing SAE-based analyses.

Critical strengths

The decomposition keeps the decoder bias and the reconstruction residual separate from the SAE features. The MILP is stated both for the reconstructed expert set and for the original set with the residual fixed. The authors also note that the minimal subset need not be unique and that cancelling features can be dropped.

Reconstruction fidelity is reported across layers.

Critical weaknesses
  1. Eq6 follows from substituting the SAE reconstruction into a linear router, so the decomposition itself is not a substantial contribution. RouterInterp (Lasy et al., 2026) already uses SAE features to explain expert selection. This paper adds suppressive features, per-token values, and the MILP, but it does not show that these additions change the conclusions one would draw from that prior method.

  2. The SAE recovers 81–88% of the selected experts on OLMoE and 73–86% on DeepSeek-V2-Lite, so part of each decision remains in the residual. Figure 4b shows the residual’s L1 norm, but the case studies did not say whether it changes the selected expert set. Recall@K is an average overlap, and the paper does not report how often the entire Top-K set is recovered exactly.

  3. A promotive feature is selected because it assigns a large positive logit to the target expert, and steering adds that same vector back to the router logits. The expert is therefore expected to move up almost by construction. The experiment also covers only L1-E9 of OLMoE, with two promotive features and one suppressive feature.

  4. The qualitative analysis uses three experts and one short sentence, and the early-layer features are mostly simple token detectors. The deeper-layer labels are automatically generated, and some are quite vague. The author said they were manually examined, but it reports neither the number of features checked nor any agreement or quality score.

  5. The entropy in Section 5.5 is computed from W_r(z_i f_i), so it mixes activation magnitude with direction. The claim that the bottom and top layers have a stronger expert preference does not follow from this plot alone.

  6. The applications in Section 6 are only suggested.

Questions for authors
  1. What can this method show that RouterInterp or direct logit attribution cannot?

  2. Is the subset of 8.79 features smaller than the nine largest-L1 features, or than a greedy subset of the same size? Does this fraction stay near one quarter of the active set if the SAE sparsity or dictionary size changes?

  3. Would the steering results hold if the intervention were applied to the representation rather than added to the router logits?

  4. gpt-oss has a per-expert router bias, but Eq6 omits it. Are these biases included in the reported logits, fidelity, and MILP? If not, how much do the subset sizes and Recall@K change?

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

Weakness 1

In this work, we propose that routing logits can be attributed to SAE features linearly and arithmetically. We show that the decomposition is a straightforward and practical way to understand how features assign logits to experts. To the best of our knowledge, this has not been discussed in previous studies.

RouterInterp aims to identify features that are most predictive of the selection of a given expert. However, expert selection is not usually driven by only one feature; it is a result of the "competition" among active features (Page 7, Figure 4a). RouterInterp does not consider such a "competition" among features and attributes selection only to the most predictive features because it aims to summarize which features tend to activate a specific expert. This is a fundamental difference between RouterInterp and our work: We are quantifying logits assigned by features to the experts; in this way, we are not identifying “most predictive” features; we are dissecting the routing to gain knowledge about it. More differences between RouterInterp and our work are summarized in General response to all reviewers.

We use SAE because it is a suitable method to find features from representations systematically. The fundamental motivations of RouterInterp and our work are different, and thus we have no intention to challenge their conclusions.

Weakness 2

In the case of Figure 4, 7 of the top-8 experts are recovered (Rank [1 2 3 4 5 6 7 8]→[1 3 2 6 4 7 5 11]). We now include the exact recovery rate in Appendix I.

We put the distribution of recall@Top-8 of OLMoE and recall@Top-6 of DeepSeek-V2-Lite here, which we consider a better reflection of the performance of SAE (The related illustration is now included on page 8, Figure 6):

(NOTE: here we show the number of recovered experts actually, Recall = #recovered_experts / K)

#recovered experts@Top-8 (OLMoE)<=2345678
Layer 0<0.0010.0010.0120.0690.2580.4640.196
Layer 7<0.0010.0010.0100.0700.2780.4710.170
Layer 15<0.0010.0010.0100.0500.1860.4610.291
#recovered experts@Top-6 (DeepSeek-V2-Lite)0123456
Layer 1<0.001<0.0010.0030.0240.1400.4710.361
Layer 13<0.0010.0020.0180.0930.3020.4420.143
Layer 26<0.0010.0020.0160.0800.2560.4380.207

We emphasize that an exact recovery (all top-K experts are recovered) is usually difficult and possibly not so meaningful because the borderline experts may be swapped with other experts without a significant influence on the model output. The exact recovery rate is low because the middle-ranked experts, even those borderline experts (e.g., ranked K-1 or K), tend to have a close routing logit distribution with their neighbor experts, compared with the top-ranked or bottom-ranked experts (Page 9, Fig. 7), making them hard to distinguish from their neighbors.

Weakness 3

We agree that the increase is usually expected by construction. However, our intention is to empirically verify that the identified features do promote/suppress the corresponding experts, and show that they can be used to steer routing decisions. From this perspective, we believe that the experiments serve the intended purpose.

Weakness 4

We use simple tokens to demonstrate in the main paper for simplicity. We now include a quantitative result regarding human evaluation in Appendix C4.

Weakness 5

We thank you for identifying the inconsistency between the figures and the text. We apologize for a bug in the implementation code that led to incorrect data being used to plot results in Figure 8 (previously Figure 6). We have now fixed the bug and uploaded the corrected figures.

Weakness 6

We agree that these are potential applications of our methods and findings. We leave their practical implementation to future work.

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

Question 1

Regarding RouterInterp: (This part is a duplicate of General response to all reviewers 1)

RouterInterp is a method that uses statistical correlation between features and routing decisions; i.e., it gains interpretability by identifying features that are "most predictive'' for a given expert to be selected.

Our proposed method uses linear decomposition in the sparse autoencoder space to quantify feature(s) contribution to the MoE's routing decision (more specifically, "routing logits'').

This main distinction between using statistical correlation vs. linear decomposition in the SAE space leads to the following differences in interpretation capability:

  1. On a per-token basis, with our method we can identify when a feature (or any arbitrary combination of features) can promote or suppress each expert. RouterInterp, being a coarse-grained statistical method, needs a) numerous tokens, and b) computing the statistical correlation for every single arbitrary combination of features (rendering the method impractical for fine-grained interpretability at the level of a single token with particular features).

  2. RouterInterp, being a statistical method, focuses on the statistical correlation of the most important features. With our method, since we can compute the contribution for any arbitrary combination of features, we could identify features that individually appear to be of low contribution, yet when combined, they have a joint influence on the routing decision.

  3. RouterInterp could not give interpretation token-by-token (since it averages the statistics of tokens).

  4. RouterInterp considers the statistical correlation, on the experts' side, to be a binary decision: i.e., an expert is either triggered by the feature or not. In our method, since the linear decomposition in the SAEs space is linked directly to the logits of the experts, we get a magnitude of "by how much" each expert was triggered, leading to better fine-grained interpretability capability.

Regarding direct logit attribution: Given a representation, direct logit attribution (DLA) is to project it onto the unembedding space and interpret its semantic information or contribution to the model output. However, we study routing at MoE layers in this work. Furthermore, although DLA may be applied to interpret the semantic information of representations, it may not be suitable to apply to the early and middle layers, as they are less aligned with the unembedding space.

Question 2

MILP is mathematically guaranteed to find the minimum subset size, given the problems we design in the paper have feasible solutions. We now provide the results of two heuristic selection methods (feature L1-norm, feature activation), one greedy algorithm (select one active feature that maximizes the minimal logits gap between the selected experts and the unselected experts at each time), and one random baseline for searching subsets. For the random baseline, we randomly select features until the selected feature set can recover the specific target, and repeat this process 50 times for each sample to get the minimal results.

The following table shows the size of the feature subset identified by different methods that maintain the ground-truth Top-K expert set without order (gt_set) or with order (gt_order), and the reconstruction Top-K expert set without order (recon_set) or with order (recon_order). MILP achieves better performance than the compared methods. The results are obtained on 10000 randomly selected tokens.

OLMoE (Layer 0)MILPHeuristic (L1-norm)Heuristic (Activation)GreedyRandom (Best results of 50 trials)
gt_set9.1722.0524.5517.5320.21
gt_order13.7628.0329.4130.4127.67
recon_set9.2122.4124.9018.9020.67
recon_order13.8528.0929.5030.9227.84
OLMoE (Layer 7)MILPHeuristic (L1-norm)Heuristic (Activation)GreedyRandom (Best results of 50 trials)
gt_set10.7921.7121.9812.2017.81
gt_order15.0329.0229.3722.8525.67
recon_set9.5121.7221.9011.2316.68
recon_order13.2128.9829.2522.7925.09
OLMoE (Layer 15)MILPHeuristic (L1-norm)Heuristic (Activation)GreedyRandom (Best results of 50 trials)
gt_set8.6118.5318.859.8314.87
gt_order13.4828.3128.7922.1524.73
recon_set7.5418.3418.559.2113.84
recon_order11.8627.9728.4321.7823.88

We have now trained a larger dictionary with an expansion factor of 32 (previously we used 16 in the paper). We show the variation below, which is fairly small:

OLMoEgt_setgt_orderrecon_setrecon_order
Layer 0+0.50+0.57+0.58+0.59
Layer 7+0.40+0.39+0.36+0.39
Layer 15+0.42+0.50+0.44+0.54
AuthorsOpenReview ↗

Official Comment by Authors (3/3)

Question 3

The steering results still hold because it is mathematically equivalent to adding the steering feature vector to the steered representation when the SAE is trained on the MoE layer input. Given a steering feature vector ff, a steered MoE input vector mm, and the router weight matrix WrW_r, the routing logits Wr(f+m)=Wrf+WrmW_r(f + m) = W_r f + W_r m. We add WrfW_r f to WrmW_r m, which is equivalent to directly adding ff to mm in terms of their effect on the routing.

Question 4

We now update some results of gpt-oss-20b to include two variants in the appendix: bias-included (Fig. 18) and bias-excluded (Fig. 19). In the bias-excluded results, the router bias is set to zero when computing the ground-truth routing logits and recovered routing logits, to eliminate the router bias influence when we only care about the performance of the recovery from SAEs. We find that including or excluding the router bias has a rather weak impact and does not affect our conclusions. Other results of gpt-oss-20b are bias-included.

The rebuttal does not convince me to change my rating.

W1: The distinction from RouterInterp is still conceptual. Per-token logit contributions and competition among features are a reasonable difference in goal, but the response gives no example where this decomposition reaches a conclusion that RouterInterp would miss.

W2: I was worried about the recovery rate, and the rebuttal shows that it is indeed low. The concern is therefore no longer that this rate was unreported, but that the full expert set is not recovered for most tokens. The further claim that these borderline swaps have little effect on the model output is not tested either.

W3: “We agree that the increase is usually expected by construction.” That agreement already gives up the point. These features were chosen because they assign a large positive logit to the target expert, and the experiment then adds that same logit back. After that, the expert moving up is just what the setup does.

“our intention is to empirically verify that the identified features do promote/suppress the corresponding experts.” This does not follow. To verify that claim, the experiment would need an outcome that was not already fixed by how the features were chosen and how the intervention was made.

Similarly, you want to “show that they can be used to steer routing decisions.” Adding h(f) to the router logits and showing that the selected expert changes only shows that editing the logits changes the ranking (which is, of course, what editing the logits does).

Restating the purpose does not make the experiment a test of that purpose.

Q1 raises the same issue as W1, and Q3 is again about whether the steering experiment is circular. I am not convinced on either. The other points are largely addressed, but the remaining ones are more important, so I will keep my score.

I would suggest that the authors try to convince the AC, who may then challenge me during the discussion period, as I will not be persuaded by another restatement.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers (1/2)

We appreciate the reviewers’ efforts in carefully reading our paper and providing valuable feedback. Here, we provide a summary response to the critical concerns about our methodology.

1. Novelty over RouterInterp

RouterInterp is a method that uses statistical correlation between features and routing decisions; i.e., it gains interpretability by identifying features that are "most predictive'' for a given expert to be selected.

Our proposed method uses linear decomposition in the sparse autoencoder space to quantify feature(s) contribution to the MoE's routing decision (more specifically, "routing logits'').

This main distinction between using statistical correlation vs. linear decomposition in the SAE space leads to the following differences in interpretation capability:

  1. On a per-token basis, with our method we can identify when a feature (or any arbitrary combination of features) can promote or suppress each expert. RouterInterp, being a coarse-grained statistical method, needs a) numerous tokens, and b) computing the statistical correlation for every single arbitrary combination of features (rendering the method impractical for fine-grained interpretability at the level of a single token with particular features).

  2. RouterInterp, being a statistical method, focuses on the statistical correlation of the most important features. With our method, since we can compute the contribution for any arbitrary combination of features, we could identify features that individually appear to be of low contribution, yet when combined, they have a joint influence on the routing decision.

  3. RouterInterp could not give interpretation token-by-token (since it averages the statistics of tokens).

  4. RouterInterp considers the statistical correlation, on the experts' side, to be a binary decision: i.e., an expert is either triggered by the feature or not. In our method, since the linear decomposition in the SAEs space is linked directly to the logits of the experts, we get a magnitude of "by how much" each expert was triggered, leading to better fine-grained interpretability capability.

AuthorsOpenReview ↗

General response to all reviewers (2/2)

2. Comparison between ρ\rho-usefulness (the metric used in RouterInterp) and logits assigned from features to experts (Ours)

In this work, we are dissecting routing decisions by studying how features assign logits to each expert, no matter whether the experts are selected or not.

In RouterInterp (Lasy et al.), "A feature ff is ρ\rho-useful for expert EiE_i if the expected activation of ff is higher when expert EiE_i is selected than when it is not". ρ\rho-usefulness is defined as ρ(f,Ei)=E[f(x)∣Ei∈T(x)]–E[f(x)∣Ei∉T(x)]\rho(f, E_i) = E[f(x) | E_i \in \mathcal{T}(x)] – E[f(x) | E_i \notin \mathcal{T}(x)], where T(x)\mathcal{T}(x) is the set of selected experts for input xx.

We discuss the difference here:

  1. Essentially, Lasy et al. use the difference between the expected activations of a feature as a metric. It is an overall metric, not a metric for a specific scenario (it cannot be case-by-case): it cannot tell, given that feature ff has an activation of aa, what the contribution of feature ff to expert ii is. However, given the SAE is trained on MoE layer input, we formulate that as a∗(fTe)a * (f^T e), where aa is the feature activation, ff is the SAE feature vector, and ee is the expert vector in the router matrix WrW_r.
  2. Although \rho-usefulness is a causal metric, it is still indirect – it uses feature activation as a proxy to measure “usefulness”. As we formulated above, feature activation is only a factor of assigned logits. Imagine that two features may assign all experts the same logits without equal feature activation. They theoretically have the same influence on the routing logits and routing decisions, but they get a different \rho-usefulness just because they have a different feature activation.
  3. ρ\rho-usefulness relies on a threshold: whether expert ii is selected or not. However, feature ii may not be the one that decides whether expert ii is selected/unselected. ρ\rho-usefulness cannot eliminate the influence of other coexisting features when studying the “usefulness” of feature ff to expert ii, even though they use expectation to alleviate this issue.
  4. ρ\rho-usefulness is rank-related. In other words, it is not logit-sensitive. Imagine two features that have the same “usefulness” to Expert ii, they may still assign similar but different logits to the expert. In our work, we aim to measure the logits assigned by a feature to an expert ii. ρ\rho-usefulness is not a suitable metric for our task.
  5. One can cheaply predict whether a feature promotes or suppresses experts with our method arithmetically (I(fTe>0)I(f^T e > 0)); the feature activation is just a scale factor that influences the magnitude of promotion/suppression, while the dot product fTef^T e is an inherent characteristic of the feature and the expert. However, one has to run the models to obtain ρ\rho-usefulness for each feature and each expert to get such predictions.
  6. One can quantify the interaction among features mathematically with our proposed method, i.e., we can learn how active features f1,f2,…fnf_1, f_2, …f_n jointly contribute to the routing logits of experts. In this way, we can decompose the logit of an expert as they were assigned by these features. However, ρ\rho-usefulness does not support such an analysis.
  7. RouterInterp identifies "most predictive" features for an expert to be selected. However, there can be multiple features that contribute to the routing, or there may be no strong "most predictive" features at all. In such scenarios, we need to analyze the "competition" among features. Our method is straightforward for those scenarios.

Reference for RouterInterp:

Ilya Lasy, Nora Yinuo Cai, and Kola Ayonrinde. RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing. In Forty-third International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id=rDNCWfRd69.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.