acceptodds
Under review as a conference paper at ICLR 2027

Gradient Rebalancing: Mitigating Exposure-Induced Routing Drift in MoE Fine-Tuning

Abstract

Pretrained mixture-of-experts (MoE) models develop expert specialization across different input patterns. Narrow downstream distributions shift the relative frequencies of these patterns, naturally leading to imbalanced expert utilization under the pretrained router. Enforcing uniform utilization during fine-tuning may therefore disrupt useful pretrained routing preferences. However, simply removing load balancing leaves exposure asymmetry in router training unaddressed. Frequently selected experts receive direct gradient signals from more tokens. Their router updates can therefore have greater influence on routing boundaries than those of rarely selected experts. This imbalance creates a mismatch with the more balanced training conditions of pretraining, which may amplify deviations from pretrained routing behavior. To address this issue, we propose Gradient Rebalancing (GR) to mitigate exposure-induced routing drift without imposing load-balancing constraints. GR targets each expert's self-region tendency, a first-order estimate of the change in its routing-region mass induced by its own router update. We decompose this tendency into token-level contributions to region expansion and contraction. GR selectively attenuates these contributions to bring each expert's estimated tendency close to zero, regardless of its exposure. This limits exposure-driven asymmetry in boundary adjustments, helping reduce drift from pretrained routing behavior. By preserving the signs of individual gradient signals, GR retains flexibility for task-driven router adaptation. Experiments across diverse tasks and datasets show that GR achieves the best overall performance among the evaluated methods and substantially reduces routing drift with only a small computational overhead.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Mixture-of-Experts, MoE Fine-Tuning, Sparse Routing, Routing Drift, Gradient Rebalancing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wpKQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies how the router in a pretrained MoE changes during downstream fine-tuning. The authors argue that imbalanced expert usage during fine-tuning is not bad (as the fine-tuning data may just be suited to that expert). However, this imbalance can create its own problem caused by gradient updates. Experts that are selected more frequently receive more direct router-gradient updates, allowing them to have a disproportionate effect on routing boundaries and causing the pretrained routing structure to change. Standard load-balancing losses address the imbalance itself, but can interfere with useful specialization rather than addressing the effect that unequal exposure has on router updates.

The paper proposes Gradient Rebalancing (GR), which estimates how the gradient contributions received by each expert tend to expand or contract that expert’s routing region. It then dampens whichever direction dominates so that the expert’s own update has approximately zero net "tendency" to expand or contract its region. Importantly, this does not force expert utilization to become uniform or freeze the router, so routing can still adapt during fine-tuning. Across several pretrained MoEs and downstream tasks, GR generally improves downstream performance while producing substantially less routing drift than standard fine-tuning and other routing-preservation methods. The authors also show that the method adds relatively little training overhead and that simply scaling gradients according to expert exposure or update magnitude does not produce the same improvements.

Critical strengths
  1. They identify an interesting phenomenon: They show that imbalanced expert exposure during fine-tuning can itself drive routing changes, as frequently selected experts receive more direct gradient signals and therefore have a greater influence on how routing space changes.
  2. The validation shows that fixing this phenomenon actually helps. Rebalancing these gradient contributions reduces routing drift while improving downstream performance.:
Critical weaknesses
  1. GR depends on an approximate measure of routing-region change. The method uses a batch-level smooth approximation of the Top-k boundary and then reasons about its first-order change. More evidence that this quantity actually tracks the real routing changes during training would make the proposed mechanism more convincing.
Questions for authors
  1. When is GR actually necessary? The Commonsense-15K experiment suggests that its benefit depends on the degree of expert imbalance. Can the authors characterize the level of load imbalance or exposure asymmetry at which GR becomes beneficial, and could this be used to decide when to apply GR?

Not a question but the related work is missing relevant prior work on understanding router behavior during MoE fine-tuning and adaptation: Thérien, Benjamin, et al. "Continual Pre-training of MoEs: How robust is your router?." arXiv preprint arXiv:2503.05029 (2025)

2 comments

Dear Reviewer wpKQ, thank you for highlighting our analysis of exposure-induced routing drift and recognizing GR’s effectiveness. We appreciate your constructive comments and address your concerns below.


Weakness 1: Insufficient validation of the routing-region change estimate

Conclusion:

Our single-step experiment supports the estimate’s ability to predict hard Top-K routing changes, with 96.9% sign agreement under self-only updates and 89.1% order agreement under joint updates.

Explanation:

We evaluated 216 step–layer–batch combinations from a DeepSeek-V2-Lite FT w/o LB trajectory on Math14K. We estimated self-region tendency from training batches and replayed AdamW updates using fixed representations of 61,957 held-out tokens, isolating router updates from representation changes. Joint updates all router rows together; Self-only updates each row separately.

Update scopeOrder agreement (%)Sign agreement (%)
Joint89.186.8
Self-only94.496.9

We have added and highlighted this experiment in the revised main text, alongside a comparison showing that GR reduces single-step load TV by 78.7% relative to norm-matched FT.


Question 1: When to apply GR based on the degree of exposure imbalance

Conclusion:

Our results suggest that GR is particularly beneficial under pronounced expert load imbalance, and future work could establish a quantitative criterion for when to apply it across downstream tasks.

Explanation:

We have not yet established a precise threshold for applying GR, given the diversity of downstream tasks. Nevertheless, our Commonsense-15K experiment shows that most fine-tuning methods perform Comparably in a relatively balanced setting. Given GR's low computational overhead, we encourage evaluating it on downstream tasks, especially those exhibiting exposure imbalance. We could examine how the benefits of GR vary with the degree of imbalance across tasks and datasets, as well as what other factors influence its effectiveness in our future work. We have added a discussion of our contributions and limitations to the main text.


Regarding the work you mentioned [1] , we have read the paper and found its analysis of layer-wise routing changes and their association with forgetting to be informative and relevant to our study. We have cited, discussed and highlighted this paper in the Related Work section of our revised manuscript.

We would appreciate your feedback on whether our responses have addressed your concerns. And we would be happy to clarify any remaining points.

[1] Continual Pre-training of MoEs: How robust is your router? TMLR, 2025.

Thanks for clarifying on the my points. I will maintain my original score for this paper.

Reviewer_1vb6OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+2)
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The authors propose a new anchoring mechanism for pretrained Mixture-of-Experts (MoE) routers. The objective is to preserve useful pretrained routing behavior when fine-tuning on downstream tasks.

The authors introduce a geometric perspective on routing preservation by defining a quantity called routing-region mass, which measures the probability that an expert is selected for an input. They argue that during fine-tuning, frequently activated experts receive more gradient contributions, potentially causing undesirable changes in routing behavior.

To address this, they propose Gradient Rebalancing (GR), which modifies router gradients to approximately neutralize the change in each expert's routing-region mass induced by its own parameter update. This is intended to preserve pretrained routing behavior while still allowing task-driven adaptation.

Critical strengths
  1. Well-written paper. The paper is clearly written and well organized. The motivation, proposed method, and experimental evaluation are easy to follow.

  2. Interesting geometric perspective on routing preservation. The idea of controlling changes in routing-region mass rather than directly anchoring router parameters is an interesting perspective. However, the advantages of this particular geometric constraint over existing anchoring mechanisms remain unclear.

Critical weaknesses
  1. Limited novelty relative to existing routing-preservation methods. The idea of preserving pretrained routing behavior during fine-tuning has already been explored through various forms of anchoring, regularization, and gradient constraints. The proposed method introduces a different quantity to preserve, namely the self-region tendency, but it is unclear what fundamental advantage this provides over existing approaches. Furthermore, the idea of constraining gradient updates to preserve a particular quantity is well established in optimization and continual learning. The authors should better distinguish their contribution from existing work.

  2. insufficient comparisons with existing routing-anchoring methods. The proposed method is closely related to Router Prior Bias: Preserving Base Routing Structure in MoE Post-Training (Lee et al., 2026, arXiv:2609.08115), which studies the same underlying problem of preserving pretrained routing behavior without enforcing uniform expert utilization. That work evaluates several forms of soft anchoring, including parameter-space, logit-space, and probability-space constraints. However, the authors only compare against a simple L2 Router Anchor baseline. Comparisons against RPB itself and its KL-to-base and logit-matching alternatives are important for establishing the novelty and effectiveness of the proposed approach. Additionally, SafeMoE (Kim et al., 2025, arXiv:2509.22745) provides a relevant precedent for directly constraining routing drift using KL-based routing alignment, albeit in a safety-specific setting.

  3. Insufficient justification for the proposed optimization objective. The authors propose enforcing approximately zero self-region tendency, Si=∇θiVi⊤Δθi≈0S_i=\nabla_{\theta_i}V_i^\top\Delta\theta_i\approx0, for each expert. However, it is unclear why zero is the appropriate target. Expanding or contracting an expert's routing region may be beneficial for downstream adaptation, and suppressing such changes could interfere with useful task gradients. The paper would benefit from a stronger theoretical or empirical justification for this particular constraint. Again, the idea of anchoring the router during finetuning might not be beneficial and would hurt you downstream performance metric that one might care about.

  4. Controlling self-region tendency does not necessarily control overall routing drift. The proposed method only constrains the contribution of an expert's own router update to its routing-region mass. However, routing boundaries are also affected by competing experts' updates and changes in upstream representations. These effects are not directly controlled by the proposed objective. While the experiments demonstrate reduced routing drift, the paper does not sufficiently explain why controlling only the self-induced component should consistently reduce overall drift.

  5. Missing in-domain evaluation and pretrained-model baselines. Although the authors evaluate their models on multiple held-out benchmarks, the evaluation does not consistently report performance on held-out examples from the fine-tuning distribution. For instance, the mathematical reasoning experiments fine-tune on Math14K but report accuracy on other benchmarks, without reporting performance on the held-out Math14K validation set. This makes it difficult to distinguish improvements in downstream adaptation from better preservation of pretrained capabilities. Additionally, the paper does not report the original pretrained models' performance before fine-tuning. The Frozen Router baseline is not equivalent, since the remaining trainable parameters are still updated. Reporting both pretrained performance and in-domain held-out performance would provide a clearer picture of the actual benefits and trade-offs of fine-tuning with GR.

  6. Experimental reliability and missing optimization controls. The paper does not report performance across multiple random seeds or provide confidence intervals. Moreover, the experiments do not include comparisons with independently tuned router learning rates, which could provide a simpler way to limit routing drift. The norm-matched ablation is useful but does not fully address this concern.

  7. Limited understanding of when the proposed method is beneficial. The commonsense-reasoning experiments show that standard load balancing outperforms GR, suggesting that the benefits of GR depend on the downstream task and the degree of expert-utilization imbalance. Although the authors investigate selective application of GR, the paper does not establish a clear criterion for when GR should be preferred over standard fine-tuning or load balancing.

Questions for authors
  1. Relationship to existing anchoring methods: The proposed method appears closely related to existing routing-preservation approaches, particularly Router Prior Bias (Lee et al., 2026). What is the fundamental advantage of constraining the self-region tendency over existing parameter-space, logit-space, or probability-space anchoring methods? Could the authors provide comparisons against the actual Router Prior Bias method and KL-based routing regularization?

  2. Choice of optimization objective: Why is approximately zero self-region tendency the appropriate target? Wouldn't some experts naturally need to expand or contract their routing regions during downstream adaptation? Could enforcing this constraint suppress useful task gradients?

  3. Missing evaluation baselines: Could the authors report the performance of the original pretrained models before fine-tuning? Additionally, for the mathematical reasoning experiments, why is accuracy on the held-out Math14K validation set not reported? This would help distinguish improvements in adapting to the fine-tuning distribution from better preservation of pretrained capabilities on other benchmarks.

  4. When should GR be used? The commonsense-reasoning experiments show that standard load balancing outperforms GR. Can the authors better characterize the conditions under which GR is expected to help, and whether expert-load imbalance can reliably predict its effectiveness?

3 comments

Dear Reviewer 1vb6, thank you for recognizing the clarity of our writing, highlighting our geometric perspective on routing preservation, and offering constructive feedback. Since Questions 1–4 correspond to Weaknesses 2, 3, 5, and 7, respectively, we organize our responses by weakness and address each of them within the corresponding response below.


Weakness 1: Limited novelty relative to existing routing-preservation methods

Conclusion:

Our novelty lies in analyzing routing drift caused by exposure imbalance and deriving an expert-specific geometric correction.

Explanation:

Existing router anchoring methods encourage proximity to pretrained parameters or outputs. We identify expert exposure imbalance as a more fundamental concern in downstream MoE fine-tuning without load balancing. We analyze how each expert’s own router update changes its routing region and derive a geometric correction to mitigate exposure-induced drift, with pretrained routing preservation as a secondary goal.


Weakness 2: insufficient comparisons with existing routing-anchoring methods

Conclusion:

RPB [1] reported comparable performance across the four anchoring formulations it evaluated, motivating our choice of the parameter-space Router Anchor as a representative baseline.

Explanation:

RPB [1] compares four soft router-anchoring formulations and reports: “Anchors defined on the router weights, on its logits, or on its output distribution perform comparably with no consistent ordering.” Its Table 3 shows:

Methodmath20kcoding20k
RPB (span-level logit bias)28.27 ± 0.1639.59 ± 0.14
Router-L2 (parameter space)27.69 ± 0.2839.35 ± 0.35
KL-to-base (probability space)27.50 ± 0.6239.27 ± 0.33
Logit-L2 (centered logit space)27.24 ± 0.7739.81 ± 0.67

These findings and computational constraints motivated our choice of L2 Router Anchor as a representative baseline. SafeMoE [2] follows the same general KL-based anchoring formulation as the KL-to-base baseline in RPB, with safety-specific reference distributions and constrained tokens, therefore we did not include it separately.


Weakness 3: Insufficient justification for the proposed optimization objective

Conclusion:

We set a near-zero target for self-region tendency primarily to mitigate exposure-induced routing drift. Preserving pretrained routing behavior is a secondary goal.

Explanation:

Frequently selected experts receive more direct gradient contributions, giving their updates additional influence on routing boundaries due to higher exposure alone, rather than downstream adaptation needs. GR targets approximately zero self-region tendency to suppress this asymmetry, with pretrained routing preservation as a secondary goal.

This constraint may also attenuate useful adaptation, since exposure effects and task-driven updates are difficult to disentangle. Nevertheless, GR improves average performance by 2.57 percentage points over FT w/o LB on exposure-imbalanced tasks, suggesting that the benefits of mitigating exposure-induced routing drift outweigh the loss of downstream adaptability caused by GR’s constraint. Its comparable performance on the more balanced Commonsense-15K suggests limited adaptation costs in that setting.

GR constrains only each expert’s self-induced net region change, while updates to other router rows can still affect its region through a cross-induced adaptation channel. We replace “competition-induced” with “cross-induced” in the revised manuscript because self-induced update also involve Top-kk competition.


Weakness 4: Controlling self-region tendency does not necessarily control overall routing drift

Conclusion:

GR intentionally controls estimated self-induced region changes while retaining other directions for task adaptation.

Explanation:

We agree that controlling self-region tendency does not guarantee reduced overall routing drift. GR intentionally neutralizes only self-induced net region changes to mitigate exposure-induced drift, while retaining cross-induced updates for adaptation. Constraining each expert’s total region mass to its initial value would impose a stronger restriction that could hinder downstream adaptation.

Weakness 5: Missing in-domain evaluation and pretrained-model baselines

Conclusion: The official GSM8K and AQuA test sets can be regarded as in-domain evaluation sets for Math14K. The original performance of base models are very low so the result in our experiment clearly reflect the performance of fine-tuning.

Explanation: According to the original paper [3] and its GitHub repository [4], Math14K derives from GSM8K and AQuA training data and has no official validation split. The official GSM8K and AQuA test sets therefore provide in-domain evaluation, while MedMCQA and WikiSQL provide explicit in-domain evaluation for our other tasks.

We use base checkpoints rather than instruction-tuned models. Under our zero-shot prompting and answer-extraction protocol, pretrained DeepSeek-V2-Lite scored below 4% on every mathematical benchmark. Most generated responses were excessively long and incoherent. These results provide a reference for fine-tuning gains:

MetricSingleEqMultiArithAddSubGSM8KSVAMPAQuAAverage
Accuracy0.7890%0.8621%3.5443%1.1372%2.1000%1.5748%1.6679%

Weakness 6: Experimental reliability and missing optimization controls

Conclusion:

We provide new multi-seed results with standard deviation in ablation studies. Our new experiments suggest that an independently tuned router learning rate is not effective in addressing load imbalanced fine-tuning.

Explanation: All submitted experiments used seed 42. Given the cost of repeating the main experiments across three models and three tasks, particularly for Qwen3, we prioritized multi-seed ablations for mechanism validation. We retain the original single-seed results in the manuscript for consistency and report three-seed means and standard deviations below.

MethodSingleEqMultiArithAddSubGSM8KSVAMPAQuAAvg.
FT w/o LB (No-Aux)89.15 ± 0.7188.16 ± 0.7873.25 ± 1.1454.06 ± 0.4662.37 ± 0.9239.24 ± 1.8267.70
GR w/o ψ90.01 ± 0.6387.76 ± 1.2474.60 ± 1.0554.33 ± 0.6264.67 ± 1.6341.08 ± 2.1768.74
Layer-global GR89.94 ± 0.3487.59 ± 0.6973.50 ± 1.7853.37 ± 0.6063.87 ± 0.1539.24 ± 2.5667.92
Projection GR90.40 ± 0.6387.36 ± 0.4374.77 ± 0.7355.55 ± 1.3764.20 ± 0.7940.29 ± 1.2068.76
Norm-matched FT87.84 ± 0.3087.99 ± 0.8874.01 ± 1.2553.75 ± 0.5064.27 ± 1.0439.37 ± 3.5067.87
Exposure scaling89.48 ± 0.8087.47 ± 2.1873.00 ± 2.0554.13 ± 0.7263.37 ± 0.3139.50 ± 0.4567.82
GR (Ours)90.20 ± 0.6387.76 ± 1.2176.03 ± 1.0257.11 ± 1.1066.80 ± 1.3545.67 ± 0.7970.60

Separately tuning the router learning rate yields limited gains in this setting: the best tested rate improves average accuracy by only 0.32 percentage points over the default. We have added and highlighted these results in the revised appendix.

Router LRSingleEqMultiArithAddSubGSM8KSVAMPAQuAAvg.
1e-588.9587.5972.4155.1262.3038.1967.43
3e-587.5788.2872.4154.9764.0040.1667.90
1e-4 (default)88.5688.1073.1654.5962.9038.1967.58
3e-488.5688.4574.6853.9861.7035.0467.07
1e-386.5983.2868.6149.4357.4039.7664.18

Weakness 7: Limited understanding of when the proposed method is beneficial

Conclusion:

Our results suggest that GR is particularly beneficial under pronounced expert load imbalance, and future work could establish a quantitative criterion for when to apply it across downstream tasks.

Explanation:

Our results suggest that GR is particularly useful under pronounced expert-load imbalance, although we have not yet established a clear criterion for when to prefer it. Given its low overhead and comparable performance on Commonsense-15K, we encourage evaluating GR, especially in imbalanced settings. Future work could explore how imbalance and other task factors affect its benefits. We have added a discussion of our contributions and limitations to the main text.

We would appreciate your feedback on whether our responses have addressed your concerns. If so, would you consider raising your rating? We would be happy to clarify any remaining points.


[1] Router Prior Bias: Preserving Base Routing Structure in MoE Post-Training. Findings of EMNLP, 2026.

[2] SafeMoE: Safe Fine-Tuning for MoE LLMs by Aligning Harmful Input Routing. ICLR, 2026.

[3] LLM-Adapters: An Adapter Family for Parameter-Efficient Fine-Tuning of Large Language Models. EMNLP, 2023.

[4] LLM-Adapters: Fine-tuning datasets. GitHub repository. https://github.com/AGI-Edgerunners/LLM-Adapters/tree/main/ft-training_set

Dear Reviewer 1vb6, thank you again for your constructive feedback. As the discussion deadline approaches, we would appreciate any feedback on whether our responses, additional experiments, and manuscript revisions have addressed your concerns. If your main concerns have been resolved, we would be grateful if you would consider revising your rating upward. Please let us know if any points remain unclear. We would be happy to provide further clarification.

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

This paper proposes Gradient Rebalancing (GR), a method for fine-tuning pretrained MoE models while preserving useful pretrained routing behavior. GR addresses exposure-induced routing drift, where frequently selected experts receive more direct router-gradient contributions and may disproportionately alter routing boundaries. The method estimates each expert’s first-order self-region tendency and selectively attenuates token-level gradient contributions on the dominant expansion or contraction side while preserving their signs. An AdamW-compatible formulation is also provided. Experiments on DeepSeek-V2-Lite, OLMoE, and Qwen3 across mathematical reasoning, medical QA, text-to-SQL, commonsense reasoning, and continual learning show that GR generally improves downstream performance and reduces routing drift with small computational overhead.

Critical strengths

The paper identifies a meaningful problem in MoE fine-tuning by distinguishing useful downstream load imbalance from exposure-induced routing drift. The proposed expert-specific, sign-preserving gradient correction is technically interesting and supported by experiments across multiple MoE architectures and task families. The ablation studies provide useful evidence for the role of the region-influence estimate.

Critical weaknesses

The experimental results appear to rely on single runs without multi-seed results or uncertainty estimates. This makes some relatively modest improvements difficult to assess, especially because GR is not uniformly better than the baselines. The method relies on a batch-level smooth approximation of routing-region mass, but the paper provides limited sensitivity analysis and direct validation of whether the estimated self-region tendency predicts actual hard Top-k routing changes.

Questions for authors

How many random seeds were used for each configuration? If the results are based on single runs, can the authors provide multi-seed results with standard deviations or confidence intervals? How sensitive is GR to the surrogate temperature, batch size, EMA decay, neutrality tolerance, and bisection settings, particularly around the reported temperature T=0.05? Can the authors provide evidence that the estimated self-region tendency predicts subsequent changes in hard Top-k routing regions, while distinguishing router updates from changes in router input representations? How does GR compare with carefully tuned router-preservation baselines, such as stronger router-anchor penalties, routing-logit distillation, or explicit KL regularization?

2 comments

Dear Reviewer kKMa, thank you for highlighting the significance of the problem we address, the technical merits of our method, and the experimental evidence supporting it. We appreciate your constructive feedback and address your concerns below.

We summarize your concerns as follows:

  1. Statistical reliability. Clarify the number of random seeds and provide multi-seed results with standard deviations or confidence intervals.

  2. Validation of the self-region tendency estimate. Verify whether the estimated self-region tendency predicts actual hard Top-kk routing-region changes.

  3. Hyperparameter sensitivity. Evaluate sensitivity to the surrogate temperature, batch size, EMA decay, neutrality tolerance, and bisection settings.

  4. Comparison with stronger routing-preservation baselines. Compare GR with carefully tuned router-anchor penalties, routing-logit distillation, and explicit routing KL regularization.

We have carefully addressed each of these points in the following responses.


1. Regarding statistical reliability

Conclusion: We only use one seed in our experiment due to time and computational resource constraints. We have provided new multi-seed results with standard deviations for the ablation experiment.

Explanation: All submitted experiments used seed 42. Given the cost of repeating the main experiments across three models and three tasks, particularly for Qwen3, we prioritized multi-seed ablations to strengthen mechanism validation. We retain the original seed-42 results in the manuscript for consistency and report means and standard deviations across seeds 42, 43, and 44 below.

MethodSingleEqMultiArithAddSubGSM8KSVAMPAQuAAvg.
FT w/o LB (No-Aux)89.15 ± 0.7188.16 ± 0.7873.25 ± 1.1454.06 ± 0.4662.37 ± 0.9239.24 ± 1.8267.70
GR w/o ψ90.01 ± 0.6387.76 ± 1.2474.60 ± 1.0554.33 ± 0.6264.67 ± 1.6341.08 ± 2.1768.74
Layer-global GR89.94 ± 0.3487.59 ± 0.6973.50 ± 1.7853.37 ± 0.6063.87 ± 0.1539.24 ± 2.5667.92
Projection GR90.40 ± 0.6387.36 ± 0.4374.77 ± 0.7355.55 ± 1.3764.20 ± 0.7940.29 ± 1.2068.76
Norm-matched FT87.84 ± 0.3087.99 ± 0.8874.01 ± 1.2553.75 ± 0.5064.27 ± 1.0439.37 ± 3.5067.87
Exposure scaling89.48 ± 0.8087.47 ± 2.1873.00 ± 2.0554.13 ± 0.7263.37 ± 0.3139.50 ± 0.4567.82
GR (Ours)90.20 ± 0.6387.76 ± 1.2176.03 ± 1.0257.11 ± 1.1066.80 ± 1.3545.67 ± 0.7970.60

2. Regarding validation of the self-region tendency estimate

Conclusion: We designed and conducted a new experiment to test whether self-region tendency accurately predicts the direction of routing-region changes.

Explanation: We tested whether batch-estimated self-region tendency predicts hard Top-K routing changes across 216 step–layer–batch combinations from a DeepSeek-V2-Lite FT w/o LB trajectory on Math14K. We replayed AdamW router updates using fixed representations of 61,957 held-out tokens, isolating router updates from representation changes.

Joint updates all router rows together; Self-only updates each row separately. Order and sign agreement measure agreement in expert-pair rankings and expansion/contraction directions, respectively.

Update scopeOrder agreement (%)Sign agreement (%)
Joint89.186.8
Self-only94.496.9

These results support the estimate’s predictive value under both update settings. We have added and highlighted this experiment in the revised main text, alongside a comparison showing that GR reduces single-step load TV by 78.7% relative to norm-matched FT.


3. Regarding hyperparameter sensitivity

Conclusion: GR is relatively insensitive to the precise values of these hyperparameters within reasonable operating ranges.

Explanation: GR targets approximately zero self-region tendency; its hyperparameters control estimation and solver accuracy. We tested sensitivity on DeepSeek-V2-Lite with Math14K, varying one setting at a time.

Temperature, batch size, and EMA decay. GR retained higher average accuracy than FT w/o LB across all tested settings, with matched-batch gains of 2.73–3.47 percentage points.

SettingSingleEqMultiArithAddSubGSM8KSVAMPAQuAAvg.
GR (default)89.7488.6275.4458.3865.3045.6770.53
FT w/o LB (default)88.5688.1073.1654.5962.9038.1967.58
GR (T = 0.025)90.5387.7675.1958.3865.9039.7669.59
GR (T = 0.04)88.3687.4175.7058.0065.7044.4969.94
GR (T = 0.06)89.5587.9373.4257.3266.0043.3169.59
GR (T = 0.10)89.1588.4575.1958.9165.5042.5269.95
GR (batch size = 16)88.7688.7975.4456.6367.8047.6470.84
FT w/o LB (batch size = 16)86.0088.7972.4154.0663.7043.7068.11
GR (batch size = 64)88.7691.2177.2260.0564.9048.8271.82
FT w/o LB (batch size = 64)87.9790.8673.1656.6362.1039.3768.35
GR (EMA decay = 0)89.9487.4176.4657.7765.1042.5269.87
GR (EMA decay = 0.99)89.9488.2876.2057.8566.6042.5270.23

Neutrality tolerance and bisection iterations. Using the same 216 single-step replay cases, all tested settings substantially reduced load TV relative to FT w/o LB.

SettingNeutrality ToleranceBisection IterationsLoad TV (%)
FT w/o LB baseline——0.1693
GR (default)0.0200120.0307
Stricter tolerance0.0100120.0299
Looser tolerance0.0400120.0374
Fewer bisection iterations0.020060.0311
More bisection iterations0.0200240.0307

These results support GR’s robustness within the tested ranges. We have added and highlighted the full sensitivity analysis in the revised appendix.


4. Regarding comparison with stronger routing-preservation baselines

Conclusion: RPB [1] compared four router anchoring methods and found that they performed comparably in its experiments. Based on this finding, we included one of the four methods, parameter-space L2 anchoring, as a representative baseline.

Explanation: RPB [1] compares four router-anchoring formulations, including the KL and logit-matching alternatives suggested in your review. Its abstract states, “Anchors defined on the router weights, on its logits, or on its output distribution perform comparably with no consistent ordering.” The experimental results below are reproduced directly from Table 3 of the original RPB paper:

Methodmath20kcoding20k
RPB (span-level logit bias)28.27 ± 0.1639.59 ± 0.14
Router-L2 (parameter space)27.69 ± 0.2839.35 ± 0.35
KL-to-base (probability space)27.50 ± 0.6239.27 ± 0.33
Logit-L2 (centered logit space)27.24 ± 0.7739.81 ± 0.67

These findings and computational constraints motivated our choice of L2 Router Anchor as a representative baseline, which GR outperformed.


We would appreciate your feedback on whether our responses have addressed your concerns. And we would be happy to clarify any remaining points.

[1] Router Prior Bias: Preserving Base Routing Structure in MoE Post-Training. Findings of EMNLP, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.