- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
- This paper investigates whether the performance degradation caused by randomly shuffling learned routing gates is a reliable measure of the benefit of adaptive routing in neural networks, particularly graph neural networks (GNNs).
- In mixture-of-experts (MoE) architectures, different nodes may receive different routing weights that determine how expert outputs are combined. A common experimental practice is to shuffle these routing weights across nodes and observe the resulting performance degradation.
- The usual interpretation is that a large performance drop indicates that adaptive routing is important.
- The paper challenges this interpretation, arguing that shuffling measures sensitivity to the particular assignment of gates, which is fundamentally different from measuring the actual benefit of adaptive routing.
- The authors distinguish three comparisons: Shuffle sensitivity, Frozen constant gap, and Trained constant gap. The central argument is that these three quantities can differ substantially, so a shuffle-induced performance drop should not automatically be interpreted as evidence that adaptive routing improves predictive performance.
Critical strengths
The paper is overall easy to follow.
Critical weaknesses
W1. Limited novelty of the central observation.
The paper's central argument—that performance degradation under shuffled routing does not directly measure the benefit of adaptive routing—appears conceptually straightforward and closely related to established limitations of permutation-based interventions. Prior work has already emphasized that perturbation-based performance changes depend on the specific intervention and should not automatically be interpreted as evidence of a component's necessity or learning benefit (Jain & Wallace, 2019; Wiegreffe & Pinter, 2019; Hooker et al., 2021; Covert et al., 2021). More directly, Hassid et al. (2022) [not cited] investigated replacing input-dependent attention with constant attention, while Roller et al. (2021), Dikkala et al. (2023), and Fan et al. (2024) examined the benefits of learned versus non-learned routing.
- 1. Jain, S., & Wallace, B. C. (2019). Attention is not Explanation. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), 3543–3556.
- 2. Wiegreffe, S., & Pinter, Y. (2019). Attention is not not Explanation. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 11–20.
- 3. Hooker, G., Mentch, L., & Zhou, S. (2021). Unrestricted permutation forces extrapolation: Variable importance requires at least one more model, or there is no free variable importance. Statistics and Computing, 31, Article 82.
- 4. Covert, I., Lundberg, S., & Lee, S.-I. (2021). Explaining by removing: A unified framework for model explanation. Journal of Machine Learning Research, 22(209), 1–90.
- 5. Hassid, M., Peng, H., Rotem, D., Kasai, J., Montero, I., Smith, N. A., & Schwartz, R. (2022). How much does attention actually attend? Questioning the importance of attention in pretrained transformers. Findings of the Association for Computational Linguistics: EMNLP 2022, 1403–1416. [Not cited]
- 6. Roller, S., Sukhbaatar, S., Szlam, A., & Weston, J. (2021). Hash layers for large sparse models. Advances in Neural Information Processing Systems (NeurIPS), 34, 17555–17566.
- 7. Dikkala, N., Ghosh, N., Meka, R., Panigrahy, R., Vyas, N., & Wang, X. (2023). On the benefits of learning to route in mixture-of-experts models. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 9376–9396.
- 8. Fan, D., Messmer, B., & Jaggi, M. (2024). Towards an empirical understanding of MoE design choices. ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models (ME-FoMo).
W2. The theoretical contributions appear relatively elementary.
Theorem 1 follows directly from separability, uniform permutation marginals, and the observation that a maximum is no smaller than an average. Theorem 2 establishes that either ordering is possible under coupled routing, but its counterexamples rely on highly constructed settings, including all-positive labels, complete graphs, and arbitrarily weak coupling. The remaining propositions largely apply standard algebraic, combinatorial, and concentration arguments.
W3. Insufficient evidence of the practical significance of the identified evaluation problem.
The practical significance of this issue is not sufficiently demonstrated. In particular, the authors do not identify representative studies in which shuffle-induced performance drops have actually been used to make unjustified claims about adaptive routing, nor do they demonstrate that applying their proposed evaluation controls would materially change conclusions from prior work. I encourage the authors to identify concrete examples of potentially misleading shuffle-based conclusions in existing GNN/MoE literature and demonstrate how the proposed controls would alter their interpretation.
W4. Limited empirical validation of the main theoretical findings.
Although the paper presents an extensive experimental study in terms of training runs, the empirical evidence supporting its central claims is relatively narrow. The substantial discrepancy between shuffle sensitivity and constant-routing performance is largely concentrated on Roman-empire, where the joint shuffle drop is 11.59 percentage points. In contrast, the other four datasets exhibit joint shuffle drops of only 0.06–0.12 points.
More importantly, Theorem 2 establishes that coupled routing can exhibit either ordering between shuffle sensitivity and the advantage over the best constant. However, the experiments only certify the direction in which shuffle sensitivity overstates the frozen advantage, and only on Roman-empire. This behavior is already possible under separable routing, as established by Theorem 1. The opposite phenomenon, in which shuffled heterogeneous routing outperforms every constant, is demonstrated only through synthetic constructions.
Furthermore, although Section 4 introduces coupling-specific diagnostics, including the decomposition into heterogeneous reassignment effects and constant-gate suboptimality, these quantities are not systematically investigated in the main experiments. I encourage the authors to provide direct measurements of coupling effects in trained networks and investigate whether the theoretically distinctive behavior of coupled routing arises in practice. Extending the analysis beyond ACM-GCN would also strengthen the empirical contribution.
Questions for authors
See weaknesses above.
2 comments
Thank you for the detailed rebuttal and revision. I have reviewed the newly added analyses in the revised PDF.
The additional experiments meaningfully strengthen the empirical contribution, particularly the Penn94 comparison, the direct coupling measurements, and the certified positional-routing analysis.
The expanded theoretical treatment in Appendix J also provides a useful clarification of the conditions under which the separable-routing bound applies.
My remaining reservation primarily concerns the novelty and broader practical significance of the contribution. The authors have not established that prior studies have actually drawn misleading conclusions from these comparisons.
Furthermore, the distinctive positive coupled composition gap remains supported only by synthetic constructions, while the new Mowst experiments primarily concern separable interfaces. The Penn94 result is interesting, but the nearly node-constant learned gates also suggest that its training-performance reversal should not be interpreted as direct evidence for the benefit of input-dependent adaptation.
Overall, I would like to keep my initial rating. I hope the authors can take more time to improve this work and submit it to a future venue.
Thank you for separating the validity of the distinction from its significance. We agree that the paper must demonstrate a useful inference, beyond observing that interventions differ.
Prior work and theoretical scope (Appendix J). Intervention-dependent explanations, constant attention and fixed routing precede our work. Hassid et al. (2022) is a relevant omitted precedent; Roller et al. (2021) and Dikkala et al. (2023) also study alternatives to learned routing. Our narrower result concerns additive scores at a separable intervened interface: the expected shuffled score equals an average of broadcast-constant scores, so the shuffle drop upper-bounds the optimal frozen advantage. Coupling removes this guarantee. The proof is short; its value lies in the inference and applicability conditions. The constructions establish possibility, not prevalence. An extension characterizes the replacement laws preserving this guarantee for every bounded additive separable score on a finite gate bank: exactly those with common node-wise marginals. For scores in [0,1], the sharp allowance for unequal marginals is their minimum average total-variation distance from a common law.
A consequential example beyond Roman (Table 3; Appendix M). We completed 60 native Mowst*-GCN fits on Penn94 and ogbn-arxiv with the published 20,000-update joint budget and validation-selected checkpoints. On Penn94, adaptive accuracy is 82.25%. The shuffle drop is 0.01 points, while the learned-minus-frozen-constant gap is −1.08 [−1.70, −0.51] and the separately trained adaptive-minus-constant gap is +0.70 [0.14, 1.23]. Frozen replacement and training therefore give opposite rankings. Brackets are descriptive 95% split/seed bootstrap intervals, not population coverage. This evidence is separable and does not demonstrate the positive coupled extreme.
Direct measurement of coupling (Appendix I). On the original Roman checkpoints, paired validation interventions keep the evaluated node's gate fixed and change the surrounding field. Correctness gains and losses are 1.940 and 1.836 points, leaving a net 0.104 with conditional Monte Carlo SE 0.154. The average broadcast score falls 7.261 points below the best-found constant. Thus cross-node effects occur but largely cancel; the large measured discrepancy principally reflects poor average broadcast policies. This is a fitted-model explanation, not a universal graph-level cause.
A stronger static comparison (Appendix N; Tables 68-69). At a pretrained Switch Transformer's final decoder router, a certificate covers all declared soft positional gates repeated across documents without content access, with eight experts and capacity 64. On 512 calibration documents, learned accuracy is 68.878%, versus a certified positional optimum of 68.483%. This teacher-forced span-reconstruction certificate concerns real-arithmetic evaluation of stored binary32 states, not every native floating-point gate execution. The attaining policy passes native replay and uses the same 114 expert calls per document, without drops. On 512 fresh documents, the preselected policy scores 66.709% versus 68.108%; this does not certify the fresh class optimum.
We have not demonstrated an incorrect empirical conclusion in a prior paper or a competitive natural-task positive coupling gap. The Switch broadcast-comparator coupling gap remains negative. The stronger case is that these tests answer different practical questions, now illustrated beyond one graph and supported by a complete comparator-class analysis at one specified interface.