acceptodds
Withdrawn from ICLR 2027

A Persistent State for Auditable Mixture-of-Experts Routing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mixture-of-Experts (MoE) models repeatedly route tokens to sparse subsets of experts, but conventional routers expose no routing-specific record of how cross-layer influences accumulate. We introduce Scratchpad-Augmented Mixture-of-Experts (SA-MoE), which gives each router access to a low-dimensional persistent state that is not provided to the experts. Learned layerwise writes update this state, and their realized post-update changes exactly decompose the state-mediated contribution to any later routing margin, forming a routing ledger. Across sparsely upcycled SmolLM2- and Gemma-based models and three independent training seeds per architecture, this pathway adds less than 1% analytical forward compute and is strongly used by trained routers: local removal of its router contribution changes the selected Top-2 expert set in 87.6% and 69.9% of decisions, respectively. Relative to a matched latest-write-only control, persistent accumulation increases long-horizon future-routing accessibility by 19.4 and 12.2 percentage points, with positive effects in every seed. More than 90% of absolute ledger contribution comes from non-recent writes in both families, and full-forward suppression of ledger-selected writes changes later routing and output distributions. The ledger is an exact provenance object for the persistent-state pathway, not a complete causal explanation of routing. Sensitivity-aware scores better predict full-forward intervention effects, and post-hoc methods recover related cross-layer attribution without architectural modification. SA-MoE instead makes one routing-specific computational history explicit and directly inspectable within the model's natural forward computation.

Primary area:
interpretability and explainable AI
Keywords:
Mixture-of-Experts, Persistent State, MoE Routing, Interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_de9NOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SA-MoE, which introduces a persistent state to trace cross-layer routing contributions in MoE models. The paper evaluates its effects on routing behavior and attribution. While the idea is interesting, the practical significance and advantages over existing post-hoc attribution methods remain unclear.

Critical strengths
  1. The paper introduces a persistent state to explicitly track cross-layer contributions to MoE routing decisions, with a mathematically grounded decomposition of state-mediated routing margins.

  2. The paper conducts intervention-based experiments to investigate how persistent state updates affect routing decisions and downstream model behavior.

Critical weaknesses
  1. Unclear practical significance. The paper demonstrates that the persistent state enables explicit decomposition of cross-layer routing contributions, but it remains unclear how this capability translates into meaningful benefits for MoE analysis, auditing, or debugging. The paper would benefit from concrete evidence that the proposed mechanism enables insights or tasks that are difficult to achieve with existing methods.

  2. Insufficient justification over existing attribution methods. The paper does not adequately establish why architectural modification is necessary compared with applying post-hoc attribution methods to standard MoEs. In particular, gradient-times-write attribution achieves stronger correlation with full-forward intervention effects than the proposed ledger in some reported settings. This raises questions about the unique explanatory value of the ledger, especially given that it only decomposes the state-mediated component of routing decisions.

Questions for authors
  1. How sensitive are the results to the dimensionality and clipping range of the persistent state? Do different configurations lead to substantially different routing behavior or attribution quality?

  2. How does the proposed method scale to larger MoE models with more layers and experts? Does the routing ledger remain informative as model scale increases?

  3. Given that gradient-times-write attribution achieves stronger correlation with full-forward intervention effects in some settings, what unique advantages does the proposed routing ledger offer?

Reviewer_47mgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SA-MoE (Scratchpad-Augmented Mixture-of-Experts), which introduces a persistent state shared across MoE routers. Each layer updates this state through a learned write operation, allowing the state-mediated contribution to later routing decisions to be decomposed into contributions from earlier writes.

The method is evaluated on sparsely upcycled SmolLM2- and Gemma-based models, comparing against ordinary MoE and a latest-write-only variant. The experiments show that routers use the additional state and that persistence improves future-routing predictability. However, the proposed ledger does not capture the full routing computation, and existing attribution methods often provide stronger causal rankings.

Critical strengths
  • The proposed method is straightforward, and the routing ledger provides an exact decomposition of the persistent-state contribution to routing margins.
  • Multiple seeds, intervention studies, and architectural controls are a plus.
Critical weaknesses
  • The architectural contribution seems incremental; the routing ledger essentially relies on a telescoping decomposition that can also be applied to existing recurrent routers such as RMoE. A direct comparison with an instrumented recurrent router is missing.
  • The ledger only captures the persistent-state contribution, ignoring routing influences through the ordinary hidden representations. Moreover, exact decomposition does not necessarily imply causal explanation, making the actual interpretability benefit unclear.
  • The experiments mainly confirm that the additional state influences routing, but do not demonstrate how this helps diagnose failures or improve model behavior. I find the practical motivation for introducing this mechanism insufficiently supported.
  • The baseline comparison is confounded. The latest-write-only state is bounded within [-1, 1], whereas the persistent state can accumulate toward [-5, 5]. Thus, the observed improvements may partly reflect differences in state magnitude or saturation rather than persistence itself.
  • It seems that gradient x write substantially outperforms the ledger in predicting intervention effects, particularly on SmolLM2. Existing post-hoc methods also recover meaningful routing information without architectural changes, raising questions about the proposed method's advantage.
  • Supervised hidden-state probes outperform the persistent state in future-routing prediction, while downstream intervention advantages over latest-write-only routing are inconsistent. This weakens the argument that explicitly accumulated routing information provides substantial additional value.
  • Despite the small analytical FLOP overhead, the SmolLM2 experiments show noticeable runtime increases. Evaluation on only two relatively small model families also leaves scalability and broader applicability unclear.

Overall, I find the motivation reasonable, but the paper does not convincingly establish that the proposed auditability mechanism provides sufficient value beyond existing approaches.

Questions for authors

Please refer to the weaknesses above.

Reviewer_ACtQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SA-MoE, which introduces a persistent routing-specific state updated by explicit layerwise writes. These writes form an exact routing ledger that decomposes the state's contribution to later routing decisions and supports source-level interventions. Experiments on two sparsely upcycled model families show that routers rely heavily on this state, persistent accumulation improves both long-horizon routing accessibility and future-routing predictability, and suppressing selected writes affects downstream routing and model outputs.

Critical strengths

(1) The architectural design of introducing a dedicated state space for routing is interesting, and I find the observations on future-routing predictability insightful.

(2) The paper is mostly easy to follow.

Critical weaknesses

(1) Similar attribution and intervention analyses could largely be performed within existing archs such as RMoE by recording and intervening on GRU state updates. This raises questions about the necessity and novelty of the proposed design.

(2) The practical efficiency of SA-MoE is not sufficiently evaluated. The authors make efforts to show that the overhead is small (he small parameter and analytical FLOP overhead of SA-MoE). However, a potentially more important bottleneck lies in the memory occupation of the additional states and the associated memory I/O costs. These may reduce training throughput in realistic large-scale pretraining, but the paper does not adequately evaluate such costs in a realistic modern training framework.

(3) The paper lacks an analysis of how SA-MoE's routing patterns relate to those of conventional MoEs. The current results do not convince me that SA-MoE is a compelling new architecture. If the proposed design is mainly intended as an interpretability tool to help understand MoE routing, the authors should examine whether its routing behavior is consistent with that of conventional MoEs. Otherwise, the resulting explanations may be relevant only to this particular architecture.

(4) The experiments should go beyond sparse upcycling. It is unclear whether the reported observations would still hold for MoE pretraining. The authors should provide experiments under a more realistic MoE pretraining setting, even if only on a small-scale scaling ladder.

Questions for authors

See Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.