- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies how routed experts interact within the same MoE layer. The authors measure this by zeroing out the output of one or two experts. They keep the other routing weights fixed. They define interaction as the loss change from removing both experts minus the sum of individual loss changes.
The authors test every expert pair in six open weight MoE models. They use MMLU subjects as evaluation domains. The analysis yields three main findings.
First, the effects are nearly additive. Removing two experts costs about the sum of removing them individually. The regression slopes are between 0.97 and 1.00. Second, domain overlap increases interaction. Experts active in the same domains interact more. Generalist experts interact more than specialists. Third, interactions weaken in deeper layers. They are lowest in the final MoE layer.
The second part uses a simplified model. It has one MoE layer with softmax routing and linear experts. The inputs are Gaussian mixtures. The teachers are orthogonal and linear. The loss is squared error. Here, interaction is exactly the routing weighted inner product of expert outputs. SGD reduces both normalized interaction and output cosine similarity. The authors also prove local convergence to a teacher equivalent solution.
Critical strengths
S1. A large scale analysis with consistent results.
The empirical study is extensive. It covers all routed expert pairs across every MoE layer in six models. The knockout intervention is simple and clearly defined. The near additive pattern appears consistently across models. It also holds under two different loss definitions.
This is the main strength of the paper. The finding is highly relevant to expert pruning. Many pruning methods estimate expert importance individually. They assume the effects of removing multiple experts are additive. The results show this approximation works well at the aggregate level.
S2. Sound theoretical analysis with careful limitations.
The proofs for the main lemmas appear correct. This includes the case of rank deficient teachers. The authors are also very clear about what their results do not prove. They acknowledge that orthogonality does not imply additivity under cross entropy. They note the theorem does not cover the experimental setting. They also admit that absolute interaction increases during training.
These details make the claims easier to assess. The authors avoid overstating their theoretical results. The mathematical arguments are solid despite the gap between the simplified model and real LLMs.
Critical weaknesses
W1. The empirical results lack a proper baseline and overstate additivity.
This is my main concern. The large scale knockout analysis is the central contribution. However, it has three major issues.
First, the paper lacks a proper empirical baseline. Removing one expert from six is a very small perturbation. Single knockout effects are around 0.001 nats per token. Approximate additivity is expected for such small changes. A second order Taylor expansion predicts this. The key question is whether expert perturbations are more additive than other perturbations of the same size. The paper does not test this. I request a clear comparison against random directions with matched norms. You could also compare against groups of neurons in a dense network. Without this baseline, we do not know if the results are specific to MoE structure.
Second, the headline regression results hide the true pair level interactions. Figure 1 averages signed effects across domains. Positive and negative interactions cancel out. Figure 2 tells a different story. The binned median of absolute interaction is often 10 to 50 percent of the single knockout effect. This ratio is even higher for small effects. This contradicts the narrative of perfect additivity. Individual interactions are actually quite substantial. The paper needs to reconcile this framing with the data in Figure 2. Reporting normalized interaction for the LLMs would help clarify this.
Third, the domain overlap result is likely just coactivation. If two experts are never active in the same prompt, their interaction is exactly zero. Domain overlap just measures the chance for experts to interact. The current analysis does not prove that overlap explains anything beyond simple coactivation frequency.
W2. The simplified model does not explain the LLM results.
Section 4 tries to explain why additivity emerges. The connection to the empirical findings remains speculative. I see three main problems.
First, the theoretical result relies heavily on model assumptions. Interaction reduces to an inner product because of the squared loss. Orthogonality is just an assumption. The local convergence rate depends on unquantified variables. The math is fine, but it does not explain why real LLMs should behave this way.
Second, the proposed mechanism is never tested in the LLMs. The authors admit the theorem does not apply to deep networks. They do not check if expert output geometry relates to interaction in real models. Figure 5 offers weak support. The observed cosine similarities are only slightly above chance level. This does not prove the outputs are orthogonal enough to explain additivity.
Third, the toy experiment does not isolate output geometry. Normalized interaction is already very small at initialization. The model is nearly additive before training even starts. The routing factor is never measured. We do not know if routing changes or output geometry changes cause the interaction drop. The experiment fails to track convergence to the teacher.
Questions for authors
- What are the normalized interaction values for each LLM? Please break this down by layer and domain. Also, is the R squared in Figure 1 centered or uncentered? Figure 2 shows substantial absolute interactions. This might require qualifying the claim of near additivity.
- What numerical precision was used for the knockout experiments? How did you estimate the noise floor? Absolute interaction averages absolute values. Numerical noise can inflate this estimate for small effects. A noise floor analysis is necessary.
- How do the routing weights and distance to the teacher evolve during training in the toy experiment? Randomly placed centers would make routing nearly one hot. This alone could suppress interactions. Tracking these quantities would separate routing effects from output alignment effects.
- What is the random direction baseline for Figure 5 at each actual hidden size? Please also report the cosine similarity of expert pairs that are not co selected. This will show if the output geometry is truly unusual.
Dear Reviewer KF6N,
Thank you for your review. The authors have posted their response. Please add a brief follow-up indicating whether your concerns have been addressed and noting any remaining questions.
The discussion phase closes on November 18 at 11:59 pm AoE.
Thank you for your attention and contribution.
Best regards,
Area Chair