acceptodds
Under review as a conference paper at ICLR 2027

Co-Activation Is Not Collaboration: An Interventional Study of Expert Coalitions in Mixture-of-Experts Routing

Abstract

Sparse Mixture-of-Experts routers score each expert independently and execute the Top-, so the routed set maximizes a modular set function, yet everything downstream acts on the selected experts' outputs jointly. Whether set utility is additive cannot be read from co-activation, a statistic of the router's own policy. We measure it interventionally on three frozen MoE language models of 1.3B to 14.3B parameters, swapping coalition members at fixed active compute against a candidate pool that includes uniformly random experts, and decomposing pair utility into unary and interaction terms. Non-additivity is real, shallow and modest in cost. On OLMoE the interaction-energy ratio falls from 0.302 at the first MoE layer to 0.0005 at layer 14, about half the non-additive energy at shallow layers lies beyond any two-slot account, and interaction accounts for about a tenth of the router's regret against the best measured equal-compute coalition; the shallow concentration replicates on all three models. Within-layer pairwise co-activation does not track signed interventional synergy, because a pair's synergy barely reproduces. Co-activation is reliable at the sampling that routing logs supply, but a pair's mean synergy on one set of documents does not predict its mean on a disjoint set ( over eleven layers of two models), and no co-activation–synergy correlation survives FDR in forty-four tests. The oracle gain is heavy-tailed, and a linear per-token head recovers no detectable share of the reweighting gain at our scale, although the same pipeline recovers a third of a planted signal of the same size.

Primary area:
interpretability and explainable AI
Keywords:
mixture of experts, sparse routing, expert interaction, co-activation, interventional analysis, interpretability, measurement reliability

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_i2U2OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether expert co-activation in sparse MoE models reflects actual functional collaboration between experts. The authors introduce an interventional framework that compares expert coalitions under the same active-compute budget, includes randomly sampled experts in the candidate pool, and decomposes coalition utility into unary and interaction effects. Across three frozen MoE language models, they find that non-additive expert interactions are strongest in shallow layers and become very small in deeper layers, while a substantial part of the shallow interaction cannot be explained by pairwise effects alone. They further show that pairwise co-activation does not reliably predict signed interventional synergy, as pairwise synergy transfers poorly across documents and no co-activation–synergy correlation survives multiple-testing correction. Finally, they test whether the observed routing headroom can be exploited with simple learned predictors, but find no reliable gains at their experimental scale, while positive controls confirm that the evaluation pipeline can recover sufficiently strong signals.

Critical strengths

Strength 1: The paper clearly distinguishes expert co-activation from actual functional interaction through controlled interventions.

Strength 2: The experiments are carefully designed, with random expert baselines, equal-compute comparisons, and explicit numerical noise controls.

Strength 3: The analysis is thorough across multiple MoE models and reveals a clear pattern that non-additive effects are strongest in shallow layers.

Critical weaknesses

Weakness 1: The paper measures collaboration mainly through pairwise counterfactual synergy under a fixed compute budget. This does not cover context-dependent complementarity, interactions among more than two experts, or effects that emerge across layers. The results therefore show that co-activation is not a reliable indicator of the pairwise synergy measured in this paper, but they do not fully support the broader claim that co-activation does not reflect expert collaboration.

Weakness 2: The measured interaction is not necessarily generated by the experts at the intervened layer. For some of the studied models, expert outputs at that layer are combined additively, so the observed non-additive effect may instead arise from downstream nonlinearities or later routing changes. The experiments show that changing an expert coalition can produce a non-additive effect on final utility, but they do not isolate where this interaction is actually generated.

Weakness 3: Unstable average pairwise synergy across documents does not by itself show that co-activation is unrelated to collaboration. If collaboration depends on the hidden state or local context, the same expert pair may help in some cases and have little or even negative effect in others, making the average synergy unstable. The paper also finds that a substantial part of the shallow-layer non-additivity is not explained by pairwise effects. This leaves open the possibility that co-activation is related to context-dependent or higher-order collaboration.

Questions for authors

Question 1: Can the authors clarify how far the conclusion is intended to generalize beyond the pairwise counterfactual synergy studied here? In particular, do the results extend to context-dependent, higher-order, or cross-layer forms of expert collaboration?

Question 2: Can the authors separate interaction generated at the intervened layer from non-additive effects introduced later by downstream nonlinearities or subsequent routing changes?

Questino 3: Can the authors test whether co-activation becomes predictive after conditioning on the hidden state or finer-grained context, and whether group-level co-activation is related to higher-order synergy?

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

The paper reads like AI generated content without any human oversight. I feel like I spent more time trying to understand the content of this paper then the authors did writing it. It requires a serious rewrite before reviewing it on its scientific merit.

Critical strengths

NA

Critical weaknesses
  • The paper is very poorly written and has many sentences that are very hard to parse what they mean and try to convey, for example (many more exist) "A no-op exercises no coalition the sweep measures, so we also check the engine against full recomputation under a non-natural coalition" or "A defender had two openings." defender against what attack? Reads like AI writing without any human guidance or editing.
Questions for authors

NA

Reviewer_91uDOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines whether expert co-activation reflects functional collaboration in sparse Mixture-of-Experts models. The authors introduce an equal-active-compute intervention protocol that swaps experts and decomposes coalition utility into additive and interaction components. Across three frozen MoE models, they find that non-additivity is concentrated in shallow layers, while pairwise co-activation does not reliably predict signed interventional synergy. They also examine the cross-document stability of pair-level synergy and report limited exploitation of oracle gains under the tested predictors. The main contribution is an empirical measurement framework and evidence clarifying the interpretation of routing statistics.

Critical strengths
  1. The paper tests whether co-activation actually reflects functional collaboration. By intervening on expert combinations at fixed active compute, it goes beyond observational routing statistics and directly measures the effects of changing the selected experts.

  2. Numerical precision checks, full-recomputation validation, alternative gate weighting, and fresh-document tests strengthen the credibility of the findings.

Critical weaknesses
  1. A routing decision can affect subsequent tokens, while the training objective aggregates losses across positions. Current-token likelihood therefore does not represent the router’s complete training objective. The reported 6.2% agreement between current-token and continuation-window oracle coalitions shows that utility choice substantially changes which coalition is considered best. The paper should clarify which central findings remain consistent across these utility choices.

  2. Appendix D describes the two-slot protocol as a lower bound on total interaction. Please clarify whether this refers to omitted interaction components or to the interaction-energy ratio. The former does not establish the latter, particularly because the two protocols use different supports and utility windows.

Questions for authors
  1. Which findings remain consistent when continuation-window utility replaces current-token utility?

  2. What establishes the claimed lower bound on non-additivity across the two different intervention protocols?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.