acceptodds
Under review as a conference paper at ICLR 2027

Experts Work Together: Learning MoE Routing from Cooperative Games

Abstract

MoE models select a small number of experts for each token, and an expert's contribution depends on which experts are activated together. We model MoE expert interactions through cooperative games: under a fixed Top-k constraint, we treat two compatible expert replacement actions as players and define coalition utility as the loss reduction of a complete expert combination relative to natural routing. Marginal contributions and the pairwise interaction identify complementarity and conflict missed by per-action evaluation. We then propose GameMoE, which constructs an interaction-weighted reward from coalition utilities. Router updates account for single-action utility and pairwise interaction, with the interaction scaled by the other replacement's current selection probability. Training combines this coalition supervision with cross-entropy on the natural forward pass and constrains relative routing score changes among experts with respect to the initial router; inference keeps the native Top-k without search or utility evaluation. Theoretical analysis characterizes how loss curvature gives rise to interactions and how they enter routing updates. GameMoE achieves the best greedy-decoding results among the evaluated methods on social commonsense reasoning and both multi-hop and single-paragraph reading comprehension. These results support using cooperative games to guide expert selection and make better use of existing expert capacity under a fixed activation budget.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Mixture-of-Experts, Expert Routing, Cooperative Game Theory, Expert Interactions, Task Adaptation

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 4 reviews, above 27% of the others.
Reviewer_5Tf3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GameMoE, a task-adaptation method that uses cooperative games to supervise MoE routing. Under a fixed Top-k budget, it evaluates individual and joint expert replacements through reductions in reference-token loss, using their utilities and pairwise interactions to train the final-layer router. Training combines this reward with cross-entropy and a centered routing anchor that constrains relative score changes. Experiments on OLMoE across three tasks show improved greedy-decoding performance while retaining native Top-8 inference.

Critical strengths
  1. Interpretable interaction-aware routing formulation: The paper explicitly distinguishes individual replacement utilities from pairwise interactions. Equation (10) shows how each action’s update combines its single-action utility with an interaction term weighted by the other action’s selection probability, providing a clear link between coalition evaluation and router optimization.

  2. Effective adaptation with native inference: GameMoE achieves the best greedy-decoding results among the evaluated methods across all three tasks, without test-time search. The parameter-matched CE-only comparison on SQuAD supports the benefit of the added training objective beyond adaptation capacity alone.

Critical weaknesses
  1. The sensitivity to interaction upweighting needs stronger explanation: In Figure 2(a), SQuAD EM drops from 85.0 with the default interaction weight ρ\rho=2 to 75.4 with ρ\rho=1, which uses the original coalition utility. Both this variant and the no-interaction variant perform substantially worse than CE-only training (82.4). Equation (10) explains why upweighting can strengthen the interaction signal, but does not fully explain the large performance gap. Results across nearby interaction weights and multiple random seeds would help establish whether the improvement is stable.

  2. The ablations do not isolate the contribution of coalition supervision from CRA: The CE-only baseline removes both coalition supervision and CRA, whereas the interaction-weight variants retain CRA. These comparisons support the combined training objective, but do not establish how much of the improvement comes from coalition supervision rather than the routing anchor. A CE + CRA baseline would help clarify this contribution.

Questions for authors
  1. Why does using the original coalition utility (ρ\rho=1) perform substantially worse than CE-only training? Could you report results across multiple random seeds and nearby interaction weights to clarify the stability of this behavior?

  2. How does CE + CRA perform without coalition supervision? Conversely, how does removing CRA affect the full method?

  3. In the fixed-model diagnostic, additive selection recovers 98.9% of the loss reduction achieved by joint selection. How should we interpret this result alongside the downstream gains from interaction-aware training? Is there additional evidence connecting the measured pairwise interactions to these gains?

  4. What is the additional training time and peak memory usage relative to the matched CE-only baseline? Although inference retains native routing, equal training epochs do not imply equal computational cost.

Reviewer_zRdSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GameMoE, which evaluates single and paired expert replacements using reference-token loss, then trains routing probabilities with an interaction-weighted coalition reward. It combines this supervision with natural cross-entropy and a routing regularizer. Coalition supervision updates the final-layer router, while inference retains native Top-8 routing. Experiments adapt OLMoE to three QA tasks.

Critical strengths
  • The replacement-game formulation gives a precise account of pairwise interactions under a fixed activation budget. Equation 10 explicitly connects these interactions to routing updates, and the intervention experiments demonstrate individually harmful replacements that become beneficial together.
  • The method improves greedy performance across three tasks while preserving native inference.
Critical weaknesses
  • The ablations leave the mechanism insufficiently established. CE-only removes both coalition supervision and CRA, so their contributions remain confounded. Moreover, ρ = 0 and ρ = 1 reach 76.3 and 75.4 EM, below CE-only at 82.4, while ρ = 2 reaches 85.0. The cause of these drops is unestablished. A CRA-only control, interaction-weight sensitivity analysis, and repeated runs are needed to support the claimed mechanism.
  • Results use one backbone without reported training-seed variation or confidence intervals; gains over the strongest baseline on SocialIQA and HotpotQA EM are only 0.5 points. Adaptation scopes differ across methods (Table 3), and EPO, a directly relevant alternative-set supervision method discussed in Section 2.1, is absent from the comparison.
Questions for authors
  1. What are the results for CE+CRA and coalition supervision without CRA? Across seeds and nearby ρ values, does the gain persist, and what accounts for the drops at ρ = 0 and ρ = 1?
  2. How does GameMoE compare with CE-only and EPO at matched adaptation scope on all three tasks? It would be better to include uncertainty and measured training-time overhead.
Reviewer_NZnYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that MoE routers miss something by scoring experts one at a time, because an expert’s usefulness depends on which other experts it’s paired with. The authors model expert swaps as a cooperative game. They measure whether two swaps help more together than apart, and they turn that interaction into a reward that trains the last-layer router. Inference uses standard Top-k routing with no added cost. Experiments on OLMoE-1B-7B across SocialIQA, HotpotQA and SQuAD show the best greedy-decoding results among six baselines. The paper also includes an ablation on the interaction weight and diagnostics of how often experts complement each other.

Critical strengths

This is a novel framing that is well motivated. Treating expert swaps as players in a cooperative game gives a concrete way to measure when experts complement or conflict with each other. Standard per-expert routing ignores this, which is why I think the work is worth publishing in some form.

Critical weaknesses

Experiments don’t support claims. The method only beats plain fine-tuning at one setting (rho = 2). At the other two settings it’s actually 6–7 points worse than not using the reward at all, which makes it look like a lucky setting rather than the interaction doing the work. The paper’s own diagnostics also show almost no interaction at the layer they train. On top of that, everything is a single run, and most of the gains are about half a point, which is within noise.

The paper never shows how routing actually changed. The reward only touches the last-layer router, but there are no numbers on how often the chosen experts differ after training, or whether the new picks are the complementary pairs the method is supposed to find. Without that, it’s hard to tell whether the gain comes from better expert combinations or from some side effect of the extra loss terms.

Questions for authors
  1. Can you report results over a few different seeds?
  2. Why does the method do worse than plain fine-tuning at rho = 0 and rho = 1? A wider sweep over rho can show that rho = 2 was not just lucky.
  3. How were rho and the other hyperparameters picked?
  4. How often do the chosen experts at the last layer change after training? This would show whether gains come from routing vs LoRA fine-tuning.
Reviewer_fxS9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GameMoE, which trains an MoE router using the losses of alternative expert combinations. Two compatible expert replacements define a local cooperative game, and their individual and joint effects are combined into an interaction-weighted reward. This supervision updates the final-layer router alongside ordinary cross-entropy and a routing anchor. Inference retains native Top-8 routing. Experiments on OLMoE show improvements in greedy decoding on three tasks, with ablations and intervention experiments examining expert interactions.

Critical strengths
  1. The replacement-based construction keeps the activation budget fixed and provides a concrete way to evaluate expert interactions. Restricting the additional evaluations to the final layer allows earlier computation to be reused, and inference requires no extra search.
  2. The SQuAD experiment includes a CE-only control with the same trainable parameters. The improvement from 82.4 to 85.0 EM provides useful initial evidence that the additional objective can help.
Critical weaknesses
  1. The effect of interaction weighting is not adequately explained. In Figure 2a, CE-only reaches 82.4 EM, while rho = 0 and rho = 1 reach 76.3 and 75.4. Only rho = 2 improves over CE-only, reaching 85.0. The particularly poor result at rho = 1, which uses the actual joint utility, deserves more attention. Increasing rho can change the preferred coalition in ways that do not agree with the task loss. For example, B_a = B_b = -0.10 and B_ab = -0.02 give J = 0.18. At rho = 2, the joint reward becomes -0.10 - 0.10 + 2(0.18) = 0.16. All replacements worsen the actual loss, yet the joint replacement receives the highest reward. This is an illustrative example rather than an observed failure. Still, it raises a practical question: how often does this disagreement occur, and why is the altered objective more effective than the original utility?

  2. The experiments do not yet isolate the contribution of interaction supervision. Figure 2b shows that additive selection already recovers 98.9% of the NLL reduction from joint selection. Figures 2b–d use full-answer losses in a fixed model and different candidate protocols, whereas training uses current-token losses with changing model parameters. These diagnostics establish that complementary replacements exist, but do not explain the large performance change in Figure 2a. The CE-only control also removes both coalition supervision and CRA, leaving the effect of anchoring unresolved. This matters because the cited RPB paper already studies centered logit-L2 anchoring in Section 3.3. CRA uses a different reference target, but the current discussion does not explain that distinction sufficiently. A CE+CRA control and a comparison with the existing anchoring formulation would help establish what the interaction supervision contributes.

Questions for authors
  1. How was rho = 2 chosen, and does its advantage over rho = 1 persist across repeated runs? Among the pairs sampled during training, how often does the weighted reward prefer a coalition whose actual utility is negative?

  2. Can you report CE+CRA and GameMoE without CRA under the same setup? Please also clarify how CRA differs from the centered logit-L2 anchor evaluated in RPB.

  3. For the replacement pairs actually sampled during training, how often does the interaction term change the update direction? Is there evidence that these updates lead to lower-loss routes under native Top-8 selection?

  4. A smaller reproducibility point: Appendix D.3 reports 69,274 training and 5,901 test examples for HotpotQA, while Table 6 of Router Lens lists 69,281 and 5,904. Please specify any additional filtering or dataset-version changes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.