acceptodds
Under review as a conference paper at ICLR 2027

Reconciling Routing and Learning for Mixture-of-Experts Expansion

Abstract

Expert expansion adds task-specific capacity to mixture-of-experts (MoE) models without increasing inference-time computation. However, experts receive task gradients only from tokens assigned by pretrained routing, which may exclude critical downstream learning signals and ultimately limit task performance. In this paper, we propose ReconExp, a proactive expansion approach that addresses this misalignment in both expert initialization and training. Specifically, ReconExp uses downstream gradients to rank pretrained experts and initialize the new expert from the highest-ranked expert. ReconExp also distills pretrained expert outputs to expand learning beyond the inherited routing region. Together, these designs make the new expert responsive to downstream learning rather than solely to pretrained routing. Experiments show that ReconExp improves the accuracy on BFCL over the strongest expansion by 5.3 points and reaches 97% of the Full-Finetune score on WebShop, while updating only 1.45% of the model parameters.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Mixture-of-Experts, Large Language Models, Expert Expansion, Sparse Routing, Parameter-Efficient Adaptation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1xinOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the mismatch between pretrained routing and downstream learning needs when expanding MoE models. ReconExp uses downstream gradients to select initialization donors and off-route distillation to broaden supervision for newly added experts. With pretrained parameters frozen and the active-expert count unchanged, the method updates approximately 1.45% as many parameters as full fine-tuning and outperforms the structurally matched expansion baseline on BFCL and WebShop.

Critical strengths
  1. Clear gains over a structurally matched baseline. ReconExp improves BFCL accuracy from 43.99% to 49.25% over LEGO and also improves WebShop performance, supporting its practical value for downstream adaptation.
  2. Convincing component-interaction analysis. Ablations show that SED helps only when paired with LGI, supporting the design rationale of establishing task-relevant initialization before expanding supervision coverage.
Critical weaknesses
  1. The mechanistic evidence is not sufficiently independent. The gradient-coverage analysis compares task gradients with distillation gradients, rather than task-gradient coverage before and after adding SED. It therefore does not directly establish broader access to useful downstream learning signals.

  2. Limited evaluation scope and baseline coverage. Experiments primarily cover one backbone and two task settings, while Expert-LoRA uses substantially fewer trainable parameters. This limits conclusions about generality and superiority over strong parameter-efficient adaptation methods.

Questions for authors

See weakness.

1 comment

Thanks for the detailed response. I believe the author team has addressed my concerns. Therefore, I will maintain my positive score.

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

The paper proposes RECONEXP for adapting pretrained mixture-of-experts models by adding one expert. Learning-Guided Initialization (LGI) selects a donor expert using token-logit gradients weighted by pretrained routing responses, then copies the expert and its router row. Support-Expanding Distillation (SED) provides supervision including tokens not routed to the new expert and introduces routing pressure according to its relative agreement with pretrained expert outputs. Experiments on DeepSeek-V2-Lite-Chat show improvements over the evaluated Expert-LoRA and LEGO baselines on BFCL and WebShop.

Critical strengths
  1. Well-motivated problem and coherent method design. Under a fixed top-k budget, task gradients reach the new expert only through tokens routed to it, as made explicit in Equation 3. The routing analysis and rollback diagnostic usefully characterize this constraint and the limitations of usage-based donor selection. The two components target these issues directly, and the ablation study suggests that both contribute and that SED is most helpful when combined with LGI.

  2. Improved performance over the evaluated baselines. RECONEXP outperforms Expert-LoRA and LEGO on BFCL overall accuracy and WebShop score. In particular, the comparison with LEGO shows gains under the same expansion structure and active-expert budget.

Critical weaknesses
  1. The paper does not compare against simpler or better-matched alternatives. First, Expert-LoRA trains about 40× fewer parameters than RECONEXP (5.75M vs. 228M), so outperforming it says little about parameter-matched LoRA on expert FFNs or ESFT. Second, the paper’s motivation is that sparse routing withholds task gradients from the new expert. Always activating the new expert alongside k−1 pretrained experts, or unfreezing the pretrained router rows, would address this directly. If either performs comparably, the added complexity of LGI and SED is hard to justify.

  2. SED may conflict with the purpose of adding an expert. SED trains the new expert to imitate the frozen Top-1 expert, even on tokens not routed to it. However, the purpose of adding an expert is to learn new functions that existing experts cannot provide. Imitation may instead make the new expert functionally redundant with existing experts. Figures 4(c) and 5(c) in fact show the new expert moving closer to existing expert outputs, and the weighting in Eqs. 12–13 increases distillation when the new expert moves away from them. The paper states that the task loss drives specialization (Appendix C.2), but provides no evidence for this. If the new expert ends up behaving like existing experts, it is unclear what adding an expert offers over directly training existing experts as in ESFT[r1], beyond keeping the original experts frozen.

    [r1] Wang, Zihan, et al. "Let the expert stick to his last: Expert-specialized fine-tuning for sparse architectural large language models." Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Questions for authors
  1. Comparison with existing-capacity adaptation: How does RECONEXP compare with LoRA applied across expert FFNs and with ESFT at comparable trainable-parameter budgets?
  2. Comparison with a direct routing alternative: Always activating the new expert alongside k−1 pretrained experts would prevent routing-based exclusion from task-gradient updates while preserving the active-expert count. How does RECONEXP compare with this simpler alternative, and what additional benefits justify its selective routing and training machinery?
  3. Hyperparameter selection: Table 10 reports different hyperparameter settings across the two datasets, including rollout temperature. Please clarify how these settings were selected, whether comparable tuning protocols were used for the baselines, and how sensitive the relative performance is to these choices.
  4. Effect on expert specialization: SED encourages the new expert to imitate pretrained Top-1 expert outputs, whereas expansion is intended to provide capacity for task-specific adaptation. Could this supervision constrain the new expert’s specialization or make it functionally redundant with existing experts? What empirical evidence shows that SED preserves the ability to learn behaviors beyond those of the pretrained experts?
Reviewer_7fpxOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses an interesting problem in MoE expansion: adding new parameters does not necessarily lead to effective parameter updates or improved downstream adaptation. To address this issue, the authors propose a method that combines task-gradient-guided expert initialization with distillation from pretrained experts and additional routing pressure. The intuition is reasonable: the new expert can learn from pretrained expert outputs even on inputs that are not routed to it, while progressively gaining opportunities to participate in downstream task learning. The proposed mathematical formulation is consistent with this intuition, and the experiments show that the method expands effective gradient support from 0.180 to 0.714 without a corresponding increase in routing traffic, alongside improvements in downstream performance.

Critical strengths
  1. Well-motivated problem and clear progression. The paper identifies a relevant limitation in MoE expansion and develops its method in a logical, step-by-step manner, connecting the limitations of existing routing to the proposed initialization and training mechanisms.

  2. Reasonable mathematical formulation. The two auxiliary losses have clear intuitive purposes, and their mathematical formulations are consistent with the intended roles of distillation and routing pressure. The experiments also provide evidence that the proposed method achieves its intended effects.

Critical weaknesses

none

Questions for authors

The intuition behind the proposed method seems reasonable to me: the newly added expert is encouraged to learn useful capabilities from all the existing experts, while its own parameters are updated to adapt to the downstream task. The experiments also demonstrate that this approach improves effective training coverage and downstream performance. My question is not about the mathematical consistency of the proposed losses or whether the method works empirically, but whether the underlying intuition can be supported by a more principled mathematical analysis. In particular, can the authors provide a more rigorous account of how distillation from existing experts interacts with task-specific optimisation, and why this combination enables the newly added expert to acquire useful task-relevant capabilities beyond its initial routing region? I would be interested in an analysis that goes beyond validating the proposed design and explains more fundamentally why this learning strategy works, potentially identifying conditions under which it is expected to be effective.

1 comment

Dear Reviewer 7fpx,

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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.