acceptodds
Under review as a conference paper at ICLR 2027

From Routing Scores to Routing Circuits in Cross-Layer Shared-Expert Transformers

Abstract

Post-hoc circuit tracing can require trained surrogate models whose reconstruction error complicates mechanistic analysis. Building on router-score decompositions and top- selection margins, we study an alternative using cross-layer shared-expert Transformers whose linear routers read the raw residual stream. The difference in linear-router scores between the weakest selected expert and its strongest excluded competitor decomposes exactly over preceding residual writes. This yields a routing-decision graph without a trained explainer. Unlike contributions to individual router scores, these margin contributions are invariant to common shifts of the router weights. We verify the decomposition in models from 4.9M to 81M parameters, with raw-score relative errors of approximately and boundary-margin absolute errors at or below . We then use this accounting to test when routing evidence supports causal conclusions. Routing-related two-dimensional residual subspaces have large effects under ablation, increasing cross-entropy by – nats, but subspaces derived from positive, negative, and total margin contributions largely coincide and have nearly identical effects. Their apparent specificity further weakens under dose-matched controls, and, under the tested interventions, changing expert identity at a single router has little effect on mean cross-entropy. These results establish margin decomposition as an exact, shift-invariant instrument for analyzing expert selection.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, mixture-of-experts, expert routing, circuits, causal interventions.

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 3 reviews, above 27% of the others.
Reviewer_1d3COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper analyzes how upstream residual writes contribute to expert selection in Transformers that employ a linear router. It precisely decomposes the margin between the weakest selected expert and the strongest excluded expert into individual component contributions, while disentangling residual evidence from balancing-policy bias. Unlike individual score contributions, this margin-based contribution is invariant to common router-row shifts. Through interventions, the authors examine whether this accounting method successfully identifies the computations critical to routing and, ultimately, the model's final behavior. The broader objective of the work is to distinguish between exact accounting and causal explanation, while leveraging the model's architectural structure to enable direct observation of specific components of its computation. Consequently, the paper provides tools for analyzing individual routing decisions, along with an assessment of the scope of the explanations they offer.

Critical strengths
  1. The chosen targets for explanation are appropriate. Since expert selection is determined by comparing candidates, the margin between selected and excluded elements serves as a meaningful target for attribution. The common-shift argument highlights a genuine ambiguity that arises when interpreting the sign of a raw score as either promotion or inhibition. Furthermore, distinguishing between residual evidence and balancing policy clarifies exactly what an attribution is explaining.

  2. The framework makes a useful distinction between the approach to computation and the understanding of its effects. Rather than assuming that a faithful reconstruction guarantees a mechanistic understanding, it directly examines whether an exact decomposition can truly support causal conclusions. Comparisons involving signed subspaces, routing interventions, and residual interventions reveal the limitations of such inferences. This perspective holds academic value, even in the absence of new predictive models or improvements in performance.

Critical weaknesses
  1. The additional explanatory contribution is limited. Li et al. [1] already provide a component-level decomposition of router scores, and GRIP [2] addresses routing stability through the margin between selected and excluded experts. While this study correctly refines the interpretation of raw-score signs, it fails to sufficiently establish what significant new insights the boundary graph provides. The problem is not that the structure is simple, but that there is limited evidence regarding how the understanding of MoE computation actually changes beyond the ambiguity of the identified attribution.

  2. The scope of validation is narrow relative to the broader goals of interpretability research. As the authors acknowledge, the graph describes localized routing comparisons based on coarse residual writes, without elucidating the semantics of the experts or the specific computations that generate these writes. Since the comparisons involving surrogate models are also confined to this limited target, they do not demonstrate a broader advantage in terms of interpretability. Furthermore, as it focuses on a self-configured early-training testbed, it is unclear whether it alters important conclusions regarding existing pretrained MoE or predicts when routing evidence becomes behaviorally significant.

  3. There is an operator–result discrepancy in the core intervention. In the local intervention described in Appendix C.1, removing opposing contributions leads to an increase in the scores of selected experts and a decrease in the scores of excluded experts. If all other executed scores at the same router remain constant, the set of selected experts should theoretically stay the same. Therefore, the reported set change requires an explanation beyond reinterpreting the result as a set change rather than a fixed-pair reversal. While this issue weakens the causal validation, it does not invalidate the decomposition identity itself.

  4. As the authors acknowledge, the behavioral findings do not definitively isolate a routing-specific mechanism. Since the subspaces obtained from positive and negative evidence are similar and interpretation varies depending on intervention strength, the general dominant residual direction remains as an alternative. A comparison with unweighted write PCA and residual PCA is necessary. Furthermore, the routing contribution after changing the residual content cannot be determined solely by the fact that the router-only effect is small in the clean residual. A rescue condition would help assess possible routing–content interaction.

  5. The text in the figures becomes pixelated when enlarged, making it difficult to read. Some cross-references to the appendix are also broken and should be corrected.

[1] Wengang Li, Lingqi Zhang, Toshio Endo, and Mohamed Wahib. Understanding Cross-Layer Contributions to Mixture-of-Experts Routing in LLMs. In International Conference on Learning Representations (ICLR), 2026.

[2] Andy Zhu, Rongzhe Wei, Yupu Gu, and Pan Li. GRIP: Algorithm-Agnostic Machine Unlearning for Mixture-of-Experts via Geometric Router Constraints. arXiv preprint arXiv:2601.16905, 2026.

Questions for authors
  1. When the existing component-score analysis is replaced with the proposed boundary accounting, what important conclusions regarding MoE routing change? How can these conclusions be independently verified?
  2. By specifying the intervention and measurement location, and verifying the corresponding expert scores and selected identities, can you explain why a set change occurs after removing opposing edges?
  3. Does the distinct behavioral effect persist when compared against unweighted write PCA and residual PCA, given a comparable intervention strength?
  4. What happens when the baseline expert identities and gates are restored following a residual ablation? How much of the loss increase remains after this rescue?
3 comments
AuthorsOpenReview ↗

Operator Discrepancy (Question 2), What Changes (Question 1), Novelty and Scope

We thank the reviewer for a careful and constructive review. We address your four questions below, with new theory and experiments where needed. This comment covers Questions 2 and 1, novelty and scope. We start with Question 2 because it concerns the soundness of the reported results. The next comment covers Questions 3 and 4 and presentation.

Question 2: why does the set change after opposing edges are removed? It should not, and with the new runner it does not.

The intervention acts on the router of layer l. It changes only two executed scores, each by αa/2 in opposite directions:

  • s_e, the score of the weakest selected expert;
  • s_j, the score of the strongest excluded expert.

We read out the top-k set at that same router. When a ≤ 0, s_e does not decrease, s_j does not increase, and the other scores are fixed, so the set cannot change. The revision proves this as Proposition 3.

The nonzero numbers in the submission came from an earlier implementation that was inconsistent with this operator. We have replaced them.

The new runner checks the property against the executed scores and selected identities at every evaluated position. It finds no set changes for negative edges on any of the five checkpoints or on the released OLMoE-1B-7B. Away from numerical ties, each positive-edge set change is the predicted exchange of e for j; there are 7 near-tie exceptions among about 2.2M positions.

Question 1: which conclusions change, and how can they be checked independently? Some conclusions are withdrawn, others are narrowed, and some become testable predictions.

  1. Conclusions about promotion versus inhibition, read from the signs of raw contributions, depend on the parameterisation. At a fixed token, a function-preserving common shift can make the raw contributions of linearly independent writes all positive or all negative (Proposition 2). In OLMoE, such a shift moves the fraction of positions whose largest raw contribution is negative from 16–88% to 0–3.5%. Decisions and boundary edges are unchanged.
  2. The empirical conclusion becomes narrower. Margin-derived subspaces largely coincide with unweighted write PCA, and restoring clean routing leaves much of their loss effect in place (see Questions 3 and 4). Boundary accounting picks out dominant write directions. It does not establish a routing-specific behavioural mechanism.
  3. The accounting yields predictions that can be tested. For a router call, m/2 is the tight ℓ∞ stability radius in score space (Proposition 1). Proposition 3 predicts from the accounting alone which positions the operator flips and which expert comes in. The runners check this position by position, independently of our attribution code.
  4. Claims about one boundary and claims about the whole route must be tested separately. Preserving one boundary margin need not preserve the selected set (Proposition 4). In deep layers, retaining the 8 most supportive whole writes gives route Jaccard of only 0.14–0.40.

Novelty relative to Li et al. and GRIP.

  • Li et al. decompose individual router scores. Proposition 2 shows that the signs of those scores cannot be identified, so we analyse the boundary edge instead.
  • GRIP constrains the selected–excluded margins so that unlearning updates do not change routing on retained data. Proposition 1 gives the general score-space radius behind constraints of this kind.

What we add is an exact decomposition of that margin over residual writes, and tests of which causal conclusions it supports.

Scope. We agree with this criticism.

  • The paper does not explain the semantics of the experts or the computations that produce individual writes, and the revision states this limitation explicitly. Nor does it yet predict when routing evidence will matter behaviourally; the rescue results in our next comment bear on this only model by model.
  • The comparison with learned surrogates (SAE and transcoder baselines) is limited to the routing target. As the paper states, we do not claim that native margins replace learned features or full transcoder graphs.
  • On pretrained models, the submission already reported observed-forward accounting on the released OLMoE-1B-7B (Appendix A). The revision extends this to seven layers, with top-8 reconstruction Jaccard of at least 0.996, and adds the sign-shift, pair-operator and cross-layer influence tests.
AuthorsOpenReview ↗

PCA Controls (Question 3), Rescue (Question 4) and Presentation

Question 3: write PCA and residual PCA at comparable strength. We compared the margin-derived bases with four kinds of two-dimensional basis:

  • unweighted, centred write PCA;
  • centred residual PCA;
  • an uncentred residual second-moment basis;
  • random bases.

Every basis was tested at unit, margin-matched and energy-matched doses, on disjoint discovery, calibration and test ranges of each model's held-out evaluation text.

The results support the alternative you raised.

  • Write PCA. The margin-derived basis largely coincides with unweighted write PCA (overlap 0.92–0.999 at most layers), and ΔCE is similar. At unit dose, ΔCE is 2.28 for both in the 6-layer model (seed 0, layer 2). In the 81M model (layer 4) it is 1.61 for the margin-derived basis and 1.83 for write PCA.
  • Residual PCA. Centred residual PCA spans different directions. Its effect is larger or smaller than that of the margin-derived basis depending on the layer.
  • Random bases. At unit dose, random projections of the same dimension change loss by at most 0.08 nats. When scaled to the same margin or energy dose, they have effects comparable to the margin-derived bases, and often larger in the 16-layer and 81M models. The submission already noted this.

So no distinctive behavioural effect remains at comparable strength. The revised Section 4.3 says that the load-bearing directions are the dominant directions of the upstream writes, and that a routing-specific effect is not established.

Question 4: rescue. We ablate the subspace in the residual stream and then restore the clean expert identities and gates. We restore them either at the ablated router alone, or at that router and every later one.

ModelΔCE, ablation onlyRescue at ablated routerRescue at all later routers
6-layer, seed 0, layer 22.281.721.21
16-layer WikiText, layer 01.631.581.31
81M FineWeb-Edu, layer 03.223.233.46

Even with every later routing decision restored, ΔCE stays at 1.21 of 2.28 nats in the 6-layer model and 1.31 of 1.63 in the 16-layer model. In the 81M model it rises from 3.22 to 3.46. The effect of rescue depends on the model, and most of the loss increase remains.

Because routing and content can interact, these numbers do not identify a single "routing share". They confirm your point that a small router-only effect on the clean residual does not mean routing is irrelevant. The revision reports this table.

Figures and cross-references. The revised PDF will use vector figures and fix all appendix references.

I thank the authors for their thoughtful comments and analysis. It is clearly a positive step that the authors have retracted the incorrect results and added further verification. The answer to Q1 partially addresses my concern, and the answer to Q2 resolves the operator discrepancy in practice. However, further analyses and experiments seem necessary to fully address Q3 and Q4. Regarding the figures, the equations in them should not become pixelated when zoomed in. In addition, it would be better if the authors mentioned and described the "OLMoE" itself in the main paper. For these reasons, I will keep my rating.

Reviewer_iQz6OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method that does not require additional training to calculate the scores contributed by each layer in front of the model for "which expert to choose", and to verify the accuracy of the calculation on models of different scales. Through intervention experiments, it is found that knowing how experts are selected does not mean knowing how much impact this selection has on the final output.

Critical strengths

1.Without the need for additional training to explain the model, it is possible to accurately calculate how many scores the previous layers contributed to expert selection, and the results are easy to verify;

Critical weaknesses

1.The contribution of the method proposed in this paper is limited, as some works conduct similar analysis. (i.e. https://arxiv.org/html/2412.00099v2) 2.According to the appendix definition, negative contribution intervention should maintain the top-k set of this layer unchanged when other scores remain unchanged, but the experimental results actually show non-zero changes; 3.According to the granularity of the entire layer write described in the article, the first four layers can only have a maximum of 8 sources. Keeping 8 complete writes should keep the routing unchanged, but actual experimental results show that there are still routing changes.

Questions for authors

can weaknesses 2 & 3 be explained in detail?

2 comments
AuthorsOpenReview ↗

Both Operator Inconsistencies Resolved (W2, W3) and Relation to Prior Work

We thank the reviewer for checking the operators against their definitions. Both inconsistencies you found in Weaknesses 2 and 3 were real. We explain each one below.

Weakness 2: a negative edge should not change the set. You are right that this cannot happen under the declared operator. The operator modifies only two executed scores at layer l:

  • s_e ← s_e − αa/2, where e is the weakest selected expert;
  • s_j ← s_j + αa/2, where j is the strongest excluded expert.

When a ≤ 0, s_e does not decrease, s_j does not increase, and all other scores are fixed, so the top-k set cannot change. The revision states and proves this as Proposition 3.

The archived results are therefore incompatible with the declared experiment. We did not keep the source of the archived runner, so we cannot tell whether the error was in how the intervention was executed or in how it was read out. We withdraw those results.

A new runner checks the property at every evaluated position. It finds no set changes for negative edges on any of five checkpoints:

  • the 6-layer model, in three seeds;
  • the 16-layer WikiText model;
  • the 81M FineWeb-Edu model.

The released OLMoE-1B-7B gives the same result.

For the supporting edge at α = 1, the set changes at 69–100% of positions. Away from numerical ties, every change is the predicted exchange of e for j; there are 7 near-tie exceptions among about 2.2M positions. By Proposition 3, this rate equals the frequency of αaᵢ > m, which can be computed from the accounting without a forward pass. The revision therefore presents the rate as an implementation check, not as independent causal evidence.

Weakness 3: retaining 8 whole writes should leave early-layer routing unchanged. This is correct. The router at layer l reads 2l+2 upstream writes:

  • the embedding;
  • the attention and MoE writes of each earlier layer;
  • layer l's own attention write.

For l ≤ 3, retaining N = 8 writes therefore replaces nothing. The new runner checks that whenever N ≥ 2l+2, the logits are bit-identical to the clean forward. On all five checkpoints, route Jaccard is 1.000 in every such case, and |ΔCE| = 0.0.

The archived nonzero changes came from the same runner, and we withdraw them. Retention is informative only in deeper layers, where 8 writes are a strict subset of all writes. There, keeping the 8 most supportive writes gives route Jaccard of 0.14–0.40.

Checking this independently. The new runners will be in the updated supplementary code. A run fails if either invariant is violated, and its output records the check outcomes and counts.

Weakness 1: relation to arXiv:2412.00099. Thank you for the pointer. Skliar et al. ("Mixture of Cache-Conditional Experts for Efficient Mobile Device Inference") is relevant to our behavioural findings. It studies expert dropping and swapping and finds that some substitutions barely change perplexity. This agrees with the small effect of single-router swaps that we observe. Its goal, however, is cache-aware routing for faster on-device inference, and it does not decompose a selection margin over residual-stream writes. We will cite and discuss it.

The closest prior work on the decomposition itself is the component-level router-score analysis of Li et al. (ICLR 2026), which we build on. We add three things.

  1. The signs of raw per-expert contributions are not identifiable from the model's function (Proposition 2). In OLMoE, a function-preserving shift moves the fraction of positions whose largest raw contribution is negative from 16–88% to 0–3.5%. The pairwise boundary edge does not change under this shift.
  2. For a router call, m/2 is the tight ℓ∞ stability radius in score space (Proposition 1). This turns the accounting into checkable conditions for when a residual edit can or cannot change that call's selected set.
  3. Controlled experiments test which causal conclusions the accounting supports.

Thanks for your response to address my concerns, i will keep my rating.

Reviewer_wS5AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies simple Transformer-based LLMs with mixture-of-experts, specifically how earlier writes into the residual stream affects the choice of expert and what effect this ultimately has on the model's performance. The starting point is the observation already made in (Li et al.) that the score given to each expert in the top-k decision can be decomposed into contributions coming from each previous layer. The authors visualize these contributions in a star shaped graph and also note that it is only the relative score of pairs of experts that affects the choice, rather the absolute scores themselves.

The more novel parts of the paper are the sections that discuss using these scores for potential interventions (removing largest contribution, subspace ablation etc), and verify the effect of these interventions on performance. The actual models used in the experiments are quite small and trained by the authors themselves. The conclusions suggest that several of these interventions only hae a modest effect on the model overall.

Critical strengths
  • The paper is very easy to follow and the questions that the authors ask are well motivated. It is great that the authors avoided the temptation to overcomplicate the paper by excessive use of LLMs.
  • The models have a simple structure and the experimental setting is controlled because they were trained by the authors.
  • The reporting on the experiments appears honest, the authors openly report negative results too and do not over-inflate any claims.
Critical weaknesses
  • The idea of decomposing the contribution to the score layer-by-layer is not new, the additional observation of shift invariance, while important, is also quite obvious. So theoretical novelty is limited.
  • What is inferred are really just the relative contributions of the different layers. The authors talk of this as a graph, but a star-shaped graph is not really a very interesting graph in the topological sense. Arguably this is just nit-picking on the terminology though. If they were able to investigate multi-layer interactions, they might be able to construct, real non-trivial graphs showing the flow of information through the network.
  • The causal interventions are good, but stop short of drawing actionable conclusions. The set-change numbers can’t identify which experts swapped, few seeds are used, and one seed didn’t replicate. The large subspace ablation effect is explained away by dose-matched random controls. There are many ways that this type of analysis could be extended, for example using counter-factual edges, incorporating multi-layer effects where the choice of router in one layer affects which router gets selected later, etc.
  • The fact that only a small number of self-trained models (it is not even clear if the models were fully trained) are used limits the extent to which we can draw conclusions from the experiments. I wish the authors repeated the analysis on larger open-srouce models, maybe with more complicated routing mechanisms. Certainly seeing results on multiple models would be important to know to what extent the results generalize rather than being specific to the particular models they trained.
Questions for authors

none

1 comment
AuthorsOpenReview ↗

New Propositions, Layer-to-Layer Influence Matrix, Seeds and OLMoE

Thank you for the positive assessment of clarity and honesty, and for the concrete suggestions. We implemented the main ones.

Theoretical novelty. The score decomposition builds on Li et al., and shift invariance on its own is a simple observation. The revision adds an appendix with four propositions that delimit what the accounting does and does not license.

  1. For a router call, m/2 is the tight ℓ∞ stability radius in score space: no perturbation with ‖δ‖∞ < m/2 changes the selected set. A corollary gives sufficient conditions on residual edits.
  2. The non-identifiability of signs is strong. At a fixed token, a function-preserving common shift can make the raw contributions of any linearly independent set of writes all positive or all negative. In the released OLMoE-1B-7B, such a shift moves the fraction of positions whose largest raw contribution is negative from 16–88% to 0–3.5%.
  3. The local outcome of the boundary operator follows from the accounting.
  4. Preserving one boundary margin need not preserve the selected set.

The star-shaped graph and multi-layer interactions. We will rename the per-decision object "boundary decomposition". Following your suggestion, we added a layer-to-layer influence matrix.

At layer l we intervene on the router only, swapping exactly one expert: the boundary pair. At every later layer m we then measure three quantities:

  • route-set Jaccard with the clean forward;
  • the change in layer m's clean boundary margin;
  • the fraction of layer m's boundary decisions that flip.

A single swap propagates. In the 6-layer model, 21–30% of downstream boundary decisions flip on average. The downstream sets nonetheless stay largely intact (Jaccard 0.96–1.00 in our models, 0.94–0.97 in OLMoE), and ΔCE is at most 1.3×10⁻³. The matrix measures how routing at one layer affects routing at later layers; it does not trace semantic information flow. The revision shows these matrices as heatmaps.

Which experts swapped. We can now identify them. Proposition 3 predicts that every set change is the boundary pair, with e leaving and j entering. The runner checks this at every position. It holds everywhere except at numerical ties: 7 positions among about 2.2M.

Seeds, including the run that did not replicate. That run was a single seed-0 run of one architecture-control configuration. Its layer-2 overlap between the positive and negative subspaces was 0.66; two further seeds gave 0.965 and 0.976 (Appendix, architecture controls). The principal 6-layer analyses now use three seeds. In each seed, the pair-operator reversal rate at α = 1 is 0.81–1.00 and the influence-matrix Jaccard is 0.978–0.989.

Subspace effect and dose-matched random controls. The original claim of specificity was too strong. At unit dose, random projections of the same dimension change loss by at most 0.08 nats, while the routing-derived bases change it by 0.14–3.22. At matched doses, however, random bases have comparable effects, as you point out. The routing-derived basis also largely coincides with unweighted PCA of the upstream writes, with similar ΔCE. The revision therefore attributes the effect to the dominant write directions and states that a routing-specific effect is not established.

Model scale and training. Our models are trained for the fixed budgets listed in Appendix A, and we do not claim they have converged. The submission already reported observed-forward accounting on the released, fully pretrained OLMoE-1B-7B at three layers (Appendix A); that model has layer-specific experts and an RMS-normalised router. The revision extends this to seven layers, with top-8 reconstruction Jaccard of at least 0.996, and adds the sign-shift, pair-operator and influence-matrix analyses. This covers one additional, released pretrained model; we do not claim generality beyond it.

Actionable conclusions. The revision states two practical rules. Each rests on a proposition and is tested on both our models and OLMoE.

  1. Do not read meaning into the signs of raw router contributions; use pairwise boundary edges.
  2. Treat the margin as a certificate: score perturbations with ‖δ‖∞ < m/2 provably cannot change that router call's selected set.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.