acceptodds
Under review as a conference paper at ICLR 2027

Intervening Before Selection: Counterfactual Effects in Adaptive Sparse Models

Abstract

Adaptive sparse models select which components participate in a computation. A component's contribution to the selected computation can therefore differ from the model's dependence on its availability. We formalize this distinction through suppression after selection and exclusion before selection, which admits eligible alternatives. Across mixtures of experts, dynamic token selection, and fitted sparse semantic fields, the two interventions yield opposite effect signs and different maximizing components, even where rankings by representation change nearly coincide, and recomputing subsequent expert assignments can reverse their loss ordering. We derive a sufficient condition for invariance of the maximizing component across interventions and use suppression ranking margins to allocate additional measurements under a reference intervention. On inputs separate from threshold calibration, this allocation reduces disagreement with the reference maximizer relative to uniform random allocation at the same number of additionally measured groups, and motivates calibration objectives that also account for the effect gaps of the remaining errors.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8Nz1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Importance estimation guided component selection is essential for improving the efficiency of adaptive sparse models. However, there is an important but easily overlooked point is that the component's contribution to the model can differ from the model's dependence on its availability. This paper is therefore motivated to formalize the distinction between component ablated after routing (suppression) and that of before the routing (exclusion), which helps to establish the stability of the resulting attribution across different removal implementations. Specifically, this paper provide empirical evidence across mixtures of experts (MoE), dynamic token selection, and fitted sparse semantic fields to answer 3 questions: (1) Which conclusions about component function remain reliable across these responses? (2) When do changes in those conclusions affect component choices (3) How can uncertain choices be checked efficiently?

Critical strengths
  1. Different from existing works focusing on examining substitution, renormalization, or compensation for a specific intervention, this work focus on the resulting attribution stability across the suppression and exclusion interventions, which is meaningful to guide model expert pruning.
  2. The introduction and methodology sections are well written and easy to follow.
Critical weaknesses
  1. For the abstract part (line 18-24), it is hard to understand the motivation behind the empirical investigations as claimed "We derive a sufficient condition,...,that also account for the effect gaps of the remaining errors". The abstract needs some modifications to be reader friendly, which should concisely state the questions that the introduction section aims to answer.
  2. The experiment didn't consider the interactions between different components. For example, if we have two components A and B, they can serve as good substitutes for each another, however, the model is highly dependent on the presence of at least one of them. In this case, exclusion testing on only one component concluded that neither A nor B were the important components, so they were both removed during model pruning and results in collapsing system. I know the table 4 includes joint exclusion results but it take different components individually and still doesn't consider the component interaction.
Questions for authors
  1. The tables and figures in the experiment section should generally be placed at the top or the bottom.
  2. In table 1, will the conclusions still hold under larger-scale MoE model?
  3. In figure1 a, the spearman correlation for representation change is close to 1 and that of the signed loss effect is much lower, could you provide more analysis of the reason?
1 comment

Thank you for the constructive and very helpful suggestions. Following them, we added experiments on the larger Qwen3-30B-A3B model and on pairs of experts that can substitute for each other, revised the paper and the figure layout, and added figures to make it easier to follow. We address each of your points below.

W1. We rewrote the abstract and added diagrams of the two interventions (Figures 1 and 2) and two MMLU examples (Figure 4) in which removing the same expert keeps the correct answer under one procedure but makes it wrong under the other, so the expert looks unnecessary in one case and important in the other.

W2. Component interactions and substitute pairs. Conclusion: Interactions between two removed experts are common at early layers, especially when one expert serves as a replacement for the other. Removing two individually minor experts rarely produces a large joint effect. Individual scores can therefore form candidate pruning sets, and the complete set should then be evaluated together.

We evaluated all pairs of 16 frequently routed experts in OLMoE and Qwen3 (Section 4.1, Figure 6, Appendix C.5). Under exclusion, pairs in which one expert replaces the other during individual removal interact more often than other pairs at every depth where interactions occur (Figure 6b), consistent with the replacement mechanism you describe. Two experts with small individual effects produce a large joint effect in fewer than 1% of evaluated pair removals (Appendix C.5). This is consistent with our pruning results, where candidate sets built from individual removal effects perform well under joint exclusion (Section 5.3, Table 1).

Q1. Figure and table placement. All figures and tables in the main text (Figures 1 to 8 and Table 1) are now placed at the top or bottom of the page.

Q2. Scaling to a larger MoE model. Conclusion: We added Qwen3-30B-A3B (48 layers, 128 experts per layer), and our main findings hold at this scale.

Within the rebuttal period, we evaluated five evenly spaced layers (0, 12, 24, 36, and 47) that span the full depth of the model. Suppression and exclusion give opposite effect signs in 24.5% of removals with a measurable effect and identify different top experts in 62.0% of groups (Figure 3, Table C.2). Later routing and pair interactions show the same qualitative patterns as in OLMoE (Figures 5 and 6, Appendix C.4). We also pruned the model at those five layers and evaluated it on 500 questions (Section 5.3, Table 1, Appendix G). Both criteria based on measured removal effects preserve the loss of the unpruned model more closely than REAP at both pruning budgets. We also report that routing frequency gives the lowest loss deviation when 16 experts are pruned per layer.

Q3. Why representation change correlations are near 1 while signed loss correlations are much lower. Conclusion: Representation change measures only the size of the output change, whereas the loss also depends on its direction.

  1. Size of the change: When the original routing weights of the remaining experts are kept, as in OLMoE, exclusion adds one replacement expert to the suppression output. By the triangle inequality, the two output change norms differ by at most the norm of this replacement contribution (Appendix A.4). The high rank correlations (0.997 in OLMoE and 0.985 in Qwen3) show that adding the replacement rarely changes the ordering of experts by output change size (Section 4.1).
  2. Direction of the change: Even a small replacement contribution can move the prediction toward or away from the correct answer, which can reverse whether the loss rises or falls (Appendix A.4). In language models, the effects of size and direction cannot be separated exactly. Time-MoE uses squared error, where the loss difference splits exactly into a direction term and a size term (Section 4.2, Equation 3). The direction term dominates almost everywhere, so whether the loss rises or falls is decided by the direction of the change rather than its size (Figure D.2).
Reviewer_JLgvOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The most recent models (e.g., LLMs based on mixtures of experts architectures) don’t use all their parts (i.e., a set of model parameters) to compute every input. They first select a few parts to do the work: a router picks a handful of experts for each token, or a scoring module picks which image patches to keep. In this setting, the paper asks: if we want to know how important one of those parts is, does it matter when we take it away? There are two ways to take a part away: 1) After selection (suppression): the model has already made its picks, and we silently switch off the one we’re testing. 2) Before selection (exclusion): we take the part out of the pool before the model picks. The model then picks the next-best part to fill the gap.

The paper assumes that these two methods should give the same answer, but the paper shows they often don’t. The same part can look helpful under one method and harmful under the other, and the “most important” part can differ between them.

Critical strengths
  • The experimental setting is tested across selection mechanisms in different models and modalities: language MoEs, a time-series MoE, token dropping in a ViT

  • The paper proposes a constructive (but preliminary) next step from how the disagreements can be used, i.e., proposes a concrete rule for where to spend extra measurements

Critical weaknesses
  • The paper is hard to understand and to follow: Altough I am quite familiar with the component attribution literature, I found this paper really hard to read and to follow. A few examples: 1) I had to reread the abstract multiple times to roughly understand what the paper is trying to accomplish, and yet from the abstract alone I wouldn't be able to tell what the exact research question is that the authors want to address. I think this has to do with the heavy use of jargon and multiple key terms being undefined (e.g., selection). 2) What is also confusing is that certain words are used to mean mutliple things. To give a another concrete example: the paper uses the word “selection” mostly for the model’s own step of picking, for each input, which of its many candidate components actually take part in the computation. The paper also uses the word for a second, different thing in Sections 5 and 6, where it refers to the researcher choosing which components to treat as most important or to prune. 3) The term "controlled replay" was introduced in the introduction without a clear definition of what this means in this context.

  • There is a possible numerical-precision problem that may affect the language-model results. App. B.2 suggests logits are stored in bf16. Then the the loss effects that define sign disagreement may be below the quantization step of the correct-answer logit. The opposite signs and top-1 disagreements could then be due to rounding. Additionally, near-ties in the router can also flip under bf16. Can the authos also report their finding of Table 1 and Fig. 1b,c in fp32? The authors themselves say the bf16 case “does not establish robustness to higher precision”.

  • I am worried that some of the key claims do not hold or that some of the claims made are overstated: For example, the authors state that the two interventions considered in their work (exlusion and suppression) “yield opposite effect signs and different maximizing components” in all settings. In table 1, we see 23.8-27.2% sign disagreement for the language MoEs, but in in table 3 (LangSplatV2) we have 0.79-3.92% and in Table 2 (Time-MoE) it is 1.4-9.9%. In addition, Table 1 counts only conditions where both effects are above the threshold, while Table 2 and Fig. 2b count all conditions. So the rates cannot be compared across settings! Further, the abstract says “We derive a sufficient condition for invariance of the maximizing component” which overstates this as a theoretical contribution (compare with Eq. 9).

  • Very close prior works are not cited and categorized, which makes it difficult to tell apart the work's contribution to the filed (see my questions below).

I very strongly recommend the authors substantially re-write this paper to 1) make it human understandable and 2) to match the claims with the conducted experimental results!

Questions for authors
  • There seems to be very similar works [1,2] to this work. Can the authors comment on how their work differs from these works?
  • Can the authos also report their findings of Table 1 and Fig. 1b,c in fp32?
  • I dont quite see how the exclusion operator corresponds to real pruning. Exclusion keeps the full router weights. Can the authors comment on this?

References

[1] Liu et al (2026), "How to Score Experts for One-Shot MoE Expert Pruning: A Unified Formulation and Selection Principle", EMNLP 2026 (https://arxiv.org/pdf/2606.15716)

[2] Engmann et al (2026), "From Observation to Intervention: A Causal Audit of Expert Importance in Mixture-of-Experts Models", ICML 2026 Workshop (https://arxiv.org/pdf/2606.10703)

2 comments

Thank you for the thorough review. We have revised the paper and added figures to make it easier to follow.

W1. Research questions. We rewrote the abstract and Section 1 to state our three research questions and main findings in plain language, and added schematic diagrams (Figures 1 and 2) and concrete MMLU examples in Qwen3 (Figure 4). In the first example, removing an expert after selection slightly lowers the loss on the correct answer and the model still answers correctly, so the expert looks unnecessary, whereas removing it before selection raises the loss and the answer becomes wrong, so it looks important. In the second example, the pattern reverses. The same expert can therefore be judged unnecessary under one procedure and important under the other. We describe the operation directly, running exclusion while keeping subsequent expert assignments fixed to those recorded under suppression (Section 4.1).

W2 & Q2. Numerical precision. Because we study how importance judgments behave in practical use, we evaluate the models in the precision in which they are released and deployed. Qwen1.5-MoE and Qwen3 are released in bfloat16, and we run them in their original configuration. We run OLMoE in bfloat16 as well, so that all three models share the same setting. Qwen3 is newly added in this revision.

  • In bfloat16, suppression and exclusion give opposite effect signs in 23.8% (OLMoE), 27.2% (Qwen1.5-MoE), and 24.5% (Qwen3) of removals, and identify different most important experts in 36.4%, 56.8%, and 62.0% of groups (Figure 3).
  • In float32, the disagreement persists: opposite effect signs in 11.8% (OLMoE) and 10.9% (Qwen1.5-MoE) of removals, different most important experts in 20.8% and 24.0% of groups, and recomputing later expert assignments still reverses the loss comparison (Appendix C.2, Table C.1).

Q1. Comparison with Liu et al. (2026) and Engmann et al. (2026). Both appeared on arXiv in June 2026, within four months of the ICLR deadline, which ICLR treats as contemporaneous work. This is why they were not cited in our submission.

  • Liu et al. state the assumption behind activation pruning explicitly: they decompose the exact pruning perturbation into the removed expert's contribution and a rerouting residual from the replacement and weight renormalization, and drop the residual because rankings by the size of the layer output change are nearly the same with and without it. Our results agree at that level (median correlations 0.997 in OLMoE and 0.985 in Qwen3, Section 4.1), but show that the conclusion about the loss can still reverse: opposite loss effect signs in about one quarter of removals and different top experts in 36% to 62% of groups (Figure 3). Measured removal effects also preserve the loss of a jointly pruned model more closely than an activation score in every setting (Section 5.3, Table 1, Figure G.1).
  • Engmann et al. compare observational activation metrics against zeroing selected experts after routing, which corresponds to our suppression. We show that removing the same expert before selection can reverse that conclusion (Figure 3), and we separate the effects of admitting a replacement, subsequent routing, candidate scoring, and weight normalization.

Q3. How exclusion relates to physical pruning and router weights. In models that normalize weights over the selected experts (such as Qwen3), removing an expert from the router softmax gives exactly the weights of our exclusion operator, because the common denominator cancels (Appendix A.2). In models that retain unnormalized probabilities (such as OLMoE and Qwen1.5-MoE), router deletion both admits a replacement and rescales the remaining weights. Our exclusion operator retains the original weights to separate these two effects, and we evaluate router deletion directly: in OLMoE, suppression and exclusion each disagree with deletion about as often as with each other (Figures 2, 5a,b, and 6a).

Dear authors,

Thank for the detailed response.

After reviewing the rebuttal and the points raised, I have decided to maintain my initial rating. My core concerns regarding the paper’s clarity, presentation, framing remain unresolved.

Best wishes,

Reviewer JLgv

Reviewer_nP4FOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies the attribution of individual components of MoE models by comparing two ways of removing a component. Suppression is removing a component after the model has selected it, without replacing it, while exclusion is removing an expert before selection and asking another to replace it. The authors quantify the disagreement across multiple architectures.

Critical strengths
  1. The authors test multiple architectures, including language MoEs, Time-MoE, DynamicViT, and sparse semantic fields. This shows that intervention-dependent attribution is not limited to one particular model.

  2. The paper reports disagreement between interventions, including opposite loss-effect signs and different most-important experts. Although the underlying observation is intuitive, quantifying its prevalence in real models is useful.

Critical weaknesses
  1. Comparing attribution methods and quantifying their disagreements is interesting, but not new. These interventions change the computation differently. It is therefore unsurprising that they produce different attribution scores, rankings, and even loss-effect signs.
  2. The paper does not establish which attribution is actually meaningful. For example, which measurement should we trust and how should we utilise it in practice?
  3. Some existing works are in the field. e.g, Explaining by Removing: A Unified Framework for Model Explanation; When Are Experts Misrouted? Counterfactual Routing Analysis in Mixture-of-Experts Language Models
Questions for authors

See above.

1 comment

Thank you for the review. We have revised the manuscript to clarify our contributions and address your concerns.

W1. "It is unsurprising that different interventions produce different scores."

Conclusion: The finding concerns the conclusions drawn from the scores, and it contradicts an assumption on which current practice relies. Existing methods measure component importance without a replacement, either from the activations of the intact model or by zeroing selected components under suppression, and apply that score to prune the model, after which the router selects replacements under exclusion. This practice assumes that importance is a property of the component alone and transfers across removal procedures (Abstract, Section 1). We show that the conclusion can reverse between the two procedures:

  • Opposite conclusions despite nearly identical representation rankings: In the evaluated language MoEs, suppression and exclusion rank experts almost identically by the magnitude of the change in the layer output (median correlations 0.997 and 0.985), yet reach opposite conclusions about whether removing an expert increases or decreases the loss on the correct answer in about one quarter of removals and identify different most important experts in 36% to 62% of groups (Section 4.1, Figure 3). These opposite loss changes shift the relative probabilities of the answer options and cause the two interventions to predict different final answers: admitting a replacement expert under exclusion can recover a correct answer that suppression loses or overturn a correct answer that suppression preserves (Section 4.1, Figure 4).
  • Contributing mechanisms: Through controlled comparisons across four model families, we show that replacement, subsequent routing, candidate scoring, and weight normalization each contribute to these differences (Sections 4.1 to 4.3, Figures 5 to 7).

W2. "Which measurement should we trust, and how should we use it in practice?"

Conclusion: The measurement should match the modification being evaluated. Suppression measures a component's contribution to the selected computation. Exclusion measures the model's dependence on its availability, which is the quantity relevant to pruning. This yields three practical rules (Sections 5 and 6):

  1. Score by measured removal effects and evaluate the pruned set jointly under exclusion. Directly measured removal effects preserve the loss of the unpruned model more closely than REAP in all three models and both pruning budgets, while loss preservation and accuracy can favor different criteria (Section 5.3, Table 1). Because removing multiple experts eliminates replacements that were available during individual removal and introduces pair interactions (Figure 6), candidate sets should be evaluated jointly.
  2. Rank by effect magnitude rather than signed effect. Whether removing a component increases or decreases the error varies across inputs and can offset when averaged, whereas rankings by effect magnitude transfer consistently to later inputs (Section 5.2, Figure D.1).
  3. Use the suppression margin to decide which choices to verify. When the top candidate under suppression leads the second by a large margin, the choice is likely to hold under exclusion. Prioritizing groups with small margins for exclusion measurement leaves fewer disagreements than random allocation at the same number of measured groups in every evaluated configuration (Sections 5.1 and 6, Figure 8).

W3. Relationship to existing work.

  • Covert et al. ("Explaining by Removing") formalize explanations based on feature removal (cited in Sections 1, 2.2, and 3). We extend this perspective to internal components of adaptive sparse models, where removing a component before selection admits a replacement and alters subsequent routing.
  • Yoon et al. ("When Are Experts Misrouted?") address a different question. They evaluate router quality by comparing selected routes with sampled alternatives at equal compute. We keep the router parameters unchanged and test whether the attribution of a given component transfers across removal procedures and how the disagreement affects component choices. Their paper is available only as an arXiv preprint (May 2026), which is why it does not appear in our submission.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.