- 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
-
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.
-
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
- 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.
- 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
-
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?
-
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?
-
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?
-
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?
-
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
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 is -useful for expert if the expected activation of is higher when expert is selected than when it is not". -usefulness is defined as , where is the set of selected experts for input .
We discuss the difference here:
- 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 has an activation of , what the contribution of feature to expert is. However, given the SAE is trained on MoE layer input, we formulate that as , where is the feature activation, is the SAE feature vector, and is the expert vector in the router matrix .
- 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.
- -usefulness relies on a threshold: whether expert is selected or not. However, feature may not be the one that decides whether expert is selected/unselected. -usefulness cannot eliminate the influence of other coexisting features when studying the “usefulness” of feature to expert , even though they use expectation to alleviate this issue.
- -usefulness is rank-related. In other words, it is not logit-sensitive. Imagine two features that have the same “usefulness” to Expert , 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 . -usefulness is not a suitable metric for our task.
- One can cheaply predict whether a feature promotes or suppresses experts with our method arithmetically (); the feature activation is just a scale factor that influences the magnitude of promotion/suppression, while the dot product is an inherent characteristic of the feature and the expert. However, one has to run the models to obtain -usefulness for each feature and each expert to get such predictions.
- One can quantify the interaction among features mathematically with our proposed method, i.e., we can learn how active features 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, -usefulness does not support such an analysis.
- 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.
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) | Ours | Logistic regression on SAE latent |
|---|---|---|
| OLMoE, Layer 0 | 0.846 | 0.786 |
| OLMoE, Layer 7 | 0.840 | 0.795 |
| OLMoE, Layer 15 | 0.871 | 0.834 |
| DeepSeek-v2-Lite, Layer 1 | 0.861 | 0.815 |
| DeepSeek-v2-Lite, Layer 13 | 0.765 | 0.692 |
| DeepSeek-v2-Lite, Layer 26 | 0.789 | 0.721 |
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:
-
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.
-
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.
-
(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 (-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 -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.
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 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)∣, where e’s are selected experts; (2) Absolute Variation of logits of top-k experts: AV topk_logit = ∑e∈E∣Δlogit(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.
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.