- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3(+1)
Read review
Paper summary
The paper investigates the causal mechanics of routing in Sparse Mixture-of-Experts (MoE) LLMs. Whereas prior interpretability work has largely taken a token-local perspective, the authors demonstrate that an expert's computation on a single token can causally influence downstream routing decisions across both tokens and layers. The study identifies two causal pathways: a direct same-token residual channel and an attention-mediated inter-token context channel. Across five open MoE models, ablating an expert changes the routing of a median 52-81% of tokens that never selected it. Before the first discrete Top-K routing change, these pathways combine approximately linearly in routing-margin space; the direct channel acts immediately and fades with depth, while context-mediated effects propagate to later tokens. Either channel alone is typically sufficient to trigger a routing change. To formalize this behavior, the paper introduces margin-normalized routing leverage () and three first-order estimators, , , , which progressively account for broader causal pathways to predict the first layer at which routing diverges from the clean forward pass. The authors support their findings with controlled causal interventions and attention key-value patching across five models, evaluating the routing-leverage estimators on four.
Critical strengths
-
The paper's main strength is that it moves beyond a token-local view of MoE interpretability and provides causal evidence that an expert's computation on one token can affect the routing of other tokens in later layers. In Section 4.1 (Table 1), ablating an expert changes the routing of a median 52-81% of tokens that never selected it. Unlike prior work that studies how context affects routing or decomposes routing scores into upstream contributions (Section 2), the authors intervene directly on an internal component and trace its influence through attention, which they term the context-mediated inter-token channel. In Section 4.3, restoring the clean keys and values eliminates these routing changes, providing further evidence that attention mediates the effect. Together, these experiments offer a mechanistic explanation for how an expert's influence extends beyond the tokens routed to it. Section 6 complements this finding with a routing-leverage measure that captures how the perturbation's alignment with the contrast direction between competing experts, relative to the margin separating them, determines whether a routing boundary is crossed. Finally, the findings suggest a potential application to expert pruning and quantization, although this practical benefit remains unevaluated.
-
The claims are supported by a range of dedicated causal experiments across five architecturally diverse MoE models, including direct, context, and joint ablations (Section 4), the additivity of the two channels before the first routing divergence (Section 5.1), and the depth dependence of direct and context-mediated influence under load-matched ablations (Section 5.2). Because the models differ in normalizer, attention variant, routing rule, number of experts and Top-K, the consistency of these results across them suggests that the observed mechanisms are not specific to a single architecture.
Critical weaknesses
-
The paper reports that ablating an expert changes the routing of a median 52-81% of tokens that never selected it (Section 4.1), but this number is not compared against a control. Top-K routing can contain near-tied boundaries, and a perturbation propagated through many layers may cross some of them regardless of its source. A comparison with a random perturbation of matched norm, removed at the same positions, would show how much of this effect is specific to the ablated expert and how much reflects the general sensitivity of Top-K routing. In addition, every change in the Top-K set is counted as a routing divergence, but the effect of these divergences on the model output is not measured, and swapping two nearly tied experts may change little downstream. In OLMoE, for example, ablations change the routing of a median 60.0% of tokens that never selected the expert while the NLL remains essentially unchanged (Tables 1 and 3). Without a baseline and an analysis of the effect on model output, the significance of these routing changes, and the suggested implications for pruning and quantization, are unclear.
-
When the true perturbation is used, simply restates that the routing margin became negative (Equation 5), so the contribution of the predictor lies in the first-order estimate of the perturbation. With , this estimator predicts the exact layer of the first routing divergence with a median recall of 30.2-51.3% (Section 6.2). From the fifth hop on, the margin ranker alone, which uses no perturbation estimate, overtakes (Section 6.3). outperforms the margin again only when evaluation is restricted to the clean-path regime, that is, to tokens whose context has not yet diverged, and identifying these tokens requires the ablated forward pass. Moreover, computing requires propagating the perturbation through all layers and tokens, at a cost comparable to running the ablation directly. It is therefore unclear in which setting the predictor would be preferred over simply running the intervention.
Questions for authors
- How does the routing flip rate under expert ablation (Section 4.1, Table 1) compare to a control intervention that removes a matched-norm random perturbation in place of the expert's output, at the same positions?
- What is the effect of the observed routing flips on the model output, for example the change in NLL at tokens that diverge? Given that OLMoE ablations change the routing of a median 60.0% of tokens that never selected the expert while leaving NLL nearly unchanged (Tables 1 and 3), do these flips reflect functional redundancy among experts?
- What is the computational cost of (Section 6.1) compared with running the ablated forward pass directly? In which practical setting do the authors expect predicting routing divergences with to be preferred over observing them through running the ablation?
3 comments
We would be very glad to answer any further questions or clarify any remaining points.
We thank the authors for the thorough response and the new experiments, which directly address our main concerns. The matched-norm control (G2) and the output analysis with the causal routing rescue (G1) are the analyses we requested.
The control result is informative: a random perturbation of the same norm flips nearly as many tokens as the ablation (47-78% vs. 52-81%), and largely the same ones. We appreciate that the authors revised their interpretation accordingly. We therefore ask the authors to ensure that the relevant parts of the paper (e.g., the introduction, the discussion of Section 4.1) reflect this revised interpretation, so that the flip rate is presented as a general property of Top-K routing under perturbation rather than as a measure of the influence of a specific expert.
The new cost comparison (Appendix L.8) confirms that a single costs about as much as the ablated pass. The multi-perturbation setting is a reasonable motivation for the predictor, although the proposed use for mixed-precision quantization has not yet been demonstrated.
Given these revisions, we raise our soundness score and keep our overall rating.
W1, Q1 (baseline; output effect; implications for compression). We implemented the proposed matched-norm control (see G2) and added both an output analysis and a causal routing rescue (see G1). Routing flips carry part of the output change causally (Table 10), and how much depends on the expert: small per token for typical experts, but for coalition experts the top-1 prediction changes for up to 24% of tokens (Table 9). Compressing an expert therefore affects tokens that never selected it. Routing leverage applies to such errors directly, and our influence graph identifies the experts for which this reaches the output, including some that are individually safe to prune but not jointly (see our response to Reviewer HXiP, Q3).
Q2 (do flips reflect functional redundancy?). Partly: the experts that replace each other after a flip are functionally close, but not redundant. We report the KL divergence KL(p∥p~) of the next-token distribution, which is the expected NLL increase of a next token drawn from the clean model. In OLMoE, aggregate NLL is essentially unchanged (Table 6), yet flips redistribute probability among the candidate tokens: the top-10 set changes for 34% of flipped tokens, against 16% of the others (Table 9). Restoring a flipped token's own route removes about a third of its output change (Table 10), so the substituted experts compute something measurably different. For coalition experts, the changes are large (G1).
W2, Q3 (value and cost of J(∞)). On what J adds beyond restating a negative margin, see bmYt W1; on its accuracy at depth, see bmYt W2. We added the cost comparison (Appendix J.8). On OLMoE, given the clean pass that both require, a single J(∞) takes about as long as the ablated pass (0.84 s against 0.82 s per batch), while J(0) and J(1) take 1% and 22% of that time. Prediction becomes preferable when many perturbations are compared on the same tokens: because routing leverage is linear in the perturbation, all three estimators for our three interventions are computed in one propagation (1.29 s), whereas the ablation requires a forward pass per intervention (2.47 s). We expect this advantage to carry over to mixed-precision quantization of the experts. Quantizing an expert perturbs its output at every token that selected it, and J(∞) propagates this perturbation like a removed write. A search over bit-width assignments then needs one propagation per expert and bit-width, after which each assignment is scored by summing the propagated perturbations, whereas running the quantized model requires a forward pass per assignment considered.