acceptodds
Under review as a conference paper at ICLR 2027

When Privacy Hurts Mergeability: Geometry-Aware Model Merging under Differential Privacy

Abstract

Model merging promises to construct a single multi-task model from independently fine-tuned task models without accessing the original task data. This makes it attractive when task data cannot be centralized, but released task models may still leak private fine-tuning data. Differential privacy (DP) provides a principled mechanism for limiting such leakage, yet its interaction with model merging remains poorly understood. In this paper, we study the geometry of differentially private model merging and identify two geometric factors associated with poor mergeability: 1) local sharpness, which increases the sensitivity of task losses to the parameter displacement induced by merging, and 2) reference drift, which captures how far private task models move from the shared pretrained initialization and can contribute to cross-task discrepancy. We propose DP-Merging, a geometry-aware framework that improves the mergeability of differentially private task models. DP-Merging uses a DP-compatible sharpness-aware objective to guide each private task model toward flatter loss regions, and a reference-based alignment regularizer to keep task models close to the shared pretrained initialization. Our analysis relates merge-induced loss increases to local loss sensitivity and parameter displacement. Experiments on vision and language tasks across multiple privacy budgets show that DP-Merging consistently improves private merged-model performance while preserving the privacy guarantees of the underlying DP fine-tuning procedures.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Differential Privacy, Model Merging, Private Fine-Tuning

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_oTHjOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the mergeability of independently fine-tuned task-specific models under differential privacy (DP) constraints. The authors identify two major geometric factors that hinder effective model merging: increased local sharpness and larger parameter drift from the shared pretrained initialization. To address these limitations, they propose a framework called DP-Merging, which combines sharpness-aware private fine-tuning with reference anchoring. Their theoretical analysis argues that the loss increase induced by model merging is closely related to local sharpness and parameter drift. Experiments across eight vision datasets and eight GLUE tasks, using CLIP ViT and RoBERTa as backbones, demonstrate that DP-Merging consistently outperforms standard differentially private fine-tuning methods under their privacy settings.

Critical strengths

S1. This paper studies an interesting problem: private models perform well on their individual tasks, but their performance drops significantly after merging.

S2. The concepts of local sharpness and reference drift provide a useful geometric view for understanding private model merging.

S3. DP-Merging is simple and practical. It requires no cross-task data access during fine-tuning and works with existing merging methods.

S4. Experiments show consistent improvements across different tasks, backbones, merging methods, and privacy budgets. The ablation results also show that both components contribute to the performance gains.

S5. The paper is well-written.

Critical weaknesses

W1. The experiments cover different model sizes of CLIP ViT and RoBERTa. Evaluating additional model families would help show whether the findings generalise to other architectures.

W2. There are some inconsistencies in the scores and reported gains in the experimental tables. While these do not seem to affect the main conclusions, they should be clarified and corrected.

W3. DP-SAT [1] is closely related to this work. A comparison with this method would help clarify how DP-Merging differs from and improves upon existing approaches.

[1] Park J, Kim H, Choi Y, et al. Differentially private sharpness-aware training[C]//International conference on machine learning. PMLR, 2023: 27204-27224.

Questions for authors

Q1. How does DP-Merging compares with existing PEFT methods, such as DP-LoRA and DP bias-only tuning? I understand that these methods may have different goals, but it would be useful to clarify what advantages DP-Merging offers over them and how it differs from these existing approaches.

Q2. I understand the motivation behind reference anchoring, but its impact on individual-task performance before merging is unclear. Does anchoring affect how well each model performs on its respective task? A brief discussion or additional results would help clarify whether there is a trade-off between individual-task performance and merged performance.

4 comments
AuthorsOpenReview ↗

Response (1/2)

Thank you very much for your positive assessment. Below, we address some of your questions.

Weakness 1. The experiments cover several model sizes but only CLIP ViT and RoBERTa architectures. Testing other model families would strengthen the generality of the results.

Response to Weakness 1: Thank you for this suggestion. To evaluate generalization beyond CLIP ViT and RoBERTa, we conduct additional experiments on two model families: ConvNeXt-B for vision and Llama-3.2-1B-Instruct for language. We follow the same experimental protocol as in the main experiments, including the DP fine-tuning procedure, privacy budget, and merging strategy.

Table 1: Evaluation of DP-Merging on ConvNeXt-B and Llama-3.2-1B-Instruct at ε=4\varepsilon = 4.

MethodConvNeXt-BLlama-3.2-1B-Instruct
DP + Weight Averaging45.247.6
DP-Merging + Weight Averaging49.352.8
DP + TIES-Merging56.453.1
DP-Merging + TIES-Merging61.757.5

Compared with the corresponding DP baselines, DP-Merging improves performance by 4.1 and 5.3 points on ConvNeXt-B, and by 5.2 and 4.4 points on Llama-3.2-1B-Instruct, using Weight Averaging and TIES-Merging, respectively. These consistent gains across both vision and language architectures provide additional evidence for the generality of our approach. We will include these additional results and the corresponding experimental details in the revised manuscript.

Weakness 2. There are a few minor inconsistencies in the reported scores and gain annotations. These do not appear to affect the main conclusions, but should be clarified and corrected.

Response to Weakness 2: Some gain annotations were calculated from unrounded mean scores, while the tables displayed rounded means, leading to apparent inconsistencies. We have checked the affected entries and recalculated the gain annotations consistently from the displayed mean scores.

Weakness 3. Comparisons with DPSAT [1], with and without reference anchoring, would clarify the benefit over closely related private fine-tuning methods.

Response to Weakness 3: Thank you for this suggestion. We have added comparisons with DP-SAT, both with and without reference anchoring, to assess whether its combination with the regularizer can account for the improvements of DP-Merging.

Comparison with DP-SAT and reference-regularized DP-SAT on the 8-task vision benchmark using ViT-B/32 at ε=4\varepsilon=4.

Training MethodIndividual Avg. Acc.Task ArithmeticTIES-MergingWUDI-Merging
DP-AdamW79.754.257.556.5
DP-SAT79.353.656.956.1
DP-SAT + Reference Regularization80.254.857.657.7
DP-Merging81.358.162.160.8

Adding reference anchoring improves DP-SAT across all three merging operators. Nevertheless, DP-Merging achieves further gains of 3.3, 4.5, and 3.1 percentage points under Task Arithmetic, TIES-Merging, and WUDI-Merging, respectively, while improving individual-model accuracy by 1.1 percentage points. This comparison shows that adding reference anchoring to DP-SAT does not reproduce the performance of DP-Merging, supporting the benefit of the proposed training procedure.

AuthorsOpenReview ↗

Response (2/2)

We sincerely thank the reviewer for their time and thoughtful feedback.

Question 1. Could the authors comment on how DP-Merging compares with parameter-efficient private fine-tuning methods, such as DP-LoRA or DP bias-only tuning?

Response to Question 1: We conducted additional comparisons with DP-LoRA and DP bias-only tuning (DP-BitFit). For DP-LoRA, we use rank r=16r=16 and apply adapters to the attention projection layers. For DP-BitFit, we freeze all non-bias parameters and update only the biases using DP-AdamW. The remaining training settings follow those in the main paper.

Comparison with parameter-efficient DP fine-tuning methods on the 8-task vision benchmark using ViT-B/32.

Fine-tuning StrategyWeight AveragingTask ArithmeticTIES-MergingPCB-Merging
DP-AdamW46.054.257.554.5
DP-LoRA44.752.856.153.3
DP-BitFit41.650.752.950.6
DP-Merging49.758.162.159.2

DP-Merging outperforms both parameter-efficient baselines across all four merging methods. With TIES-Merging, its accuracy is 6.0 points higher than DP-LoRA and 9.2 points higher than DP-BitFit. Neither baseline improves on DP-AdamW here, suggesting that limiting which parameters are updated is not enough to improve merging performance.

Question 2. I understand the motivation for reference anchoring. Could the authors briefly comment on its effect on individual-task performance before merging?

Response to Question 2: Thank you for this question. Reference anchoring is intended to improve compatibility for subsequent merging, although overly strong anchoring may restrict task adaptation. In our ViT-B/32 experiment, adding reference regularization to DP-SAT increases individual average accuracy from 79.3% to 80.2%. We have also added pre-merging individual-task results to all main tables, as summarized below. Compared with standard DP fine-tuning, DP-Merging improves average performance across all backbones. These results suggest that the full training procedure does not sacrifice average individual-task performance, although individual tasks may show small declines.

BackboneDP + IndividualDP-Merging + IndividualGain
ViT-B/3279.781.3+1.6
ViT-B/1680.082.1+2.1
ViT-L/1485.687.4+1.8
RoBERTa-Large80.680.9+0.3
RoBERTa-Base79.579.9+0.4

For further details on individual-task performance before merging, please refer to the revised main text and Appendix B, where we have added these results.

Reviewer_oTHjOpenReview ↗

good

I suggest to accept it

Thank you very much for taking the time to review our responses and share your positive feedback. We sincerely appreciate your time and consideration.

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

This paper studies the model merging of differentially private fine-tuned models and identifies two geometric factors associated with degraded merging performance: local sharpness and reference drift. To address these factors, the authors propose DP-Merging, a private fine-tuning framework that combines sharpness-aware updates with reference-anchored regularization to improve downstream model merging. The paper establishes an end-to-end differential privacy guarantee that accounts for the two adaptive private gradient queries per training iteration, and provides loss bounds that motivate controlling local sensitivity and merging displacement. Experiments on vision and language benchmarks across multiple privacy budgets demonstrate that DP-Merging improves merged-model performance when combined with several existing model-merging algorithms.

Critical strengths

(1) The paper addresses a relevant problem at the intersection of differential privacy and model merging. Studying privacy-preserving fine-tuning with subsequent merging is a useful perspective.

(2) The proposed framework spots and combines local perturbation robustness with reference anchoring without requiring cross-task data access during training. Theorem 1 explicitly accounts for both adaptive private gradient queries per iteration and establishes the privacy guarantee of the resulting procedure.

(3) The evaluation covers various vision and language benchmarks, multiple model backbones, and merging operators. The reported improvements in average merged-model performance support the framework's applicability across the evaluated settings. The component ablations also show that both sharpness-aware training and reference anchoring contribute to the gains.

Critical weaknesses

(1) The evaluation lacks controlled analysis across degrees of data heterogeneity. The existing benchmarks already cover heterogeneous domained tasks, but do not systematically vary heterogeneity to test the stability of the geometric observations and performance gains. As a complementary same-task setting, Dirichlet-distributed label partitions with varying concentration parameters is a standard evaluation to examine label skew, alongside controlled data-size imbalance or domain shifts. Such experiments should report the associated geometry and merge-induced degradation while controlling privacy budgets and training conditions.

(2) Theorem 2 provides a potentially loose upper bound whose comparative analysis is unclear. It bounds merge-induced loss using worst-case local loss variation, gradient norms, and Hessian-Lipschitz remainder terms, potentially discarding substantial information about the actual merging direction. The tightness of these relaxations is not evaluated, and the cubic remainder may limit the bound's informativeness for larger merging displacements. More importantly, the paper does not provide an analytical comparison of the bounds for DP-Merging and baseline training-and-merging procedures under common assumptions, or establish conditions under which the proposed method yields a tighter bound. The theoretical sharpness quantity also concerns all perturbation directions, whereas the algorithm uses a noisy gradient-direction perturbation, which is not established that these updates reduce the full bound. Corollary 1 further relies on uniform averaging to connect reference drift to merged loss. Consequently, the analysis offers qualitative motivation but does not clearly establish a theoretical advantage over the baselines and formal contribution in the scope.

(3) The paper acknowledges prior differentially private sharpness-aware training and non-private sharpness-aware approaches to model merging. Although adapting these ideas to private mergeability is meaningful, the narrative and evaluation should more directly provide the increment over the closest existing training procedures under matched privacy budgets and tuning conditions. The component ablations are informative, but do not by themselves establish an advantage over those procedures or clarify the nontrivial contribution beyond adding reference anchoring.

Questions for authors

(1) How do the geometric observations and the gains from DP-Merging change with the degree of data heterogeneity? Could the authors complement the existing cross-task evaluation with controlled same-task experiments, for example, Dirichlet label partitions with varying concentration parameters, and report sharpness, reference drift, and merge-induced degradation under matched privacy budgets?

(2) How informative is the upper bound in Theorem 2 in the evaluated settings, particularly the Hessian-Lipschitz remainder terms? Can the authors analytically compare DP-Merging and baseline procedures under common assumptions, or identify conditions under which DP-Merging yields a smaller bound? Please also clarify how the noisy gradient-direction updates control the worst-case sharpness quantity used in the theorem, and how far the reference-drift argument extends beyond uniform averaging.

Minor Issues

(1) Section 6.2 contains an unresolved appendix reference ("Appendix ??") for the ViT-B/16 results.

(2) Appendix C contains malformed equation references, such as "Eq. equation 15" in C.1 and "Eq. equation 19" in C.2. Please correct them.

(3) Sections 6.1 describes the training duration as "10 training epochs/steps." Please specify whether this means epochs or optimization steps and report the corresponding iteration counts used in privacy accounting.

7 comments
AuthorsOpenReview ↗

Response (1/3)

Thank you for your time and effort.

Weakness 1. The evaluation lacks controlled analysis across degrees of data heterogeneity. The existing benchmarks already cover heterogeneous domained tasks, but do not systematically vary heterogeneity to test the stability of the geometric observations and performance gains. As a complementary same-task setting, Dirichlet-distributed label partitions with varying concentration parameters is a standard evaluation to examine label skew, alongside controlled data-size imbalance or domain shifts. Such experiments should report the associated geometry and merge-induced degradation while controlling privacy budgets and training conditions.

Response to Weakness 1: Thank you for suggesting a controlled heterogeneity analysis. Our experiments follow a commonly used multi-task model merging setup: task models are independently fine-tuned from the same public pretrained checkpoint, and their shared encoder parameters are then merged while task-specific heads are retained for evaluation. In this setting, heterogeneity arises from differences in input distributions, label meanings, class counts, and task difficulty. These factors vary together, making it difficult to isolate heterogeneity through a single parameter.

To complement this evaluation, we will conduct a same-task experiment on CIFAR-10 using CLIP ViT-B/32. We will divide the training set into three disjoint subsets of nearly equal size using a size-constrained Dirichlet allocation with α∈{0.6, 0.1}\alpha\in\{0.6,\,0.1\}, where the smaller value induces stronger label skew. Three models will be independently trained from the same initialization and merged using Weight Averaging. A shared, frozen 10-class prediction head will be used throughout.

Both methods will use identical data partitions, batch sizes, optimization-step counts, and clipping thresholds. Noise will be calibrated separately to satisfy ε=4\varepsilon=4 with the same δ\delta. All evaluations will use the common test set, and results will be averaged over three paired partition/training seeds.

Dirichlet α\alphaTrainingDirectional SharpnessReference DriftIndividual Avg. Acc.Merged Acc. ↑Acc. Drop ↓
0.6DP-AdamW0.220.0890.577.612.9
0.6DP-Merging0.150.0692.481.311.1
0.1DP-AdamW0.370.1384.671.912.7
0.1DP-Merging0.260.0987.275.611.6

As label skew increases from α=0.6\alpha=0.6 to α=0.1\alpha=0.1, both methods exhibit higher directional sharpness and reference drift, together with lower individual and merged accuracy. DP-Merging maintains lower geometric measurements and improves merged accuracy by 3.7 and 3.8 percentage points, respectively. It also reduces the accuracy drop by 1.8 and 1.2 points.

Weakness 2. Theorem 2 provides a potentially loose upper bound whose comparative analysis is unclear. It bounds merge-induced loss using worst-case local loss variation, gradient norms, and Hessian-Lipschitz remainder terms, potentially discarding substantial information about the actual merging direction. The tightness of these relaxations is not evaluated, and the cubic remainder may limit the bound's informativeness for larger merging displacements. More importantly, the paper does not provide an analytical comparison of the bounds for DP-Merging and baseline training-and-merging procedures under common assumptions, or establish conditions under which the proposed method yields a tighter bound. The theoretical sharpness quantity also concerns all perturbation directions, whereas the algorithm uses a noisy gradient-direction perturbation, which is not established that these updates reduce the full bound. Corollary 1 further relies on uniform averaging to connect reference drift to merged loss. Consequently, the analysis offers qualitative motivation but does not clearly establish a theoretical advantage over the baselines and formal contribution in the scope.

Response to Weakness 2: Thank you for this comment. We would like to clarify that the purpose of our analysis is to characterize the geometric factors affecting private model merging, rather than to establish an end-to-end dominance guarantee over baseline training. Nevertheless, we agree that the original presentation could better preserve directional information and connect the analysis to the private perturbation.

AuthorsOpenReview ↗

Response (2/3)

Thank you for your careful reading. We continue our response to Weakness 2 below.

We have revised the analysis in three respects. First, we strengthen Corollary 1 without additional assumptions. Under uniform averaging, ∑tdt=0\sum_t d_t=0, so the common gradient component cancels: 1T∑tgt⊤dt=1T∑t(gt−gˉ)⊤dt.\frac{1}{T}\sum_t g_t^\top d_t = \frac{1}{T}\sum_t(g_t-\bar{g})^\top d_t. Consequently, the revised bound is

\le \frac{1}{T}\sum_t\mathcal L_t(w_t) +G_{\mathrm{disp}}\sqrt{V} +AV+\frac{M\sqrt{T}}{6}V^{3/2},$$ where $V=T^{-1}\sum_t\lVert d_t\rVert^2$ and $G_{\mathrm{disp}}^2=T^{-1}\sum_t\lVert g_t-\bar g\rVert^2$. Since $$V=D-\lVert w_{\mathrm{merge}}-w_0\rVert^2\le D, \qquad G_{\mathrm{disp}}^2=G^2-\lVert\bar g\rVert^2\le G^2,$$ this bound is no larger than the original reference-drift bound for the same returned models. The improvement comes from retaining task dispersion and cross-task cancellation before applying norm inequalities. Second, we make the directional decomposition explicit: $$\mathrm{Gap}_t =g_t^\top d_t+\frac{1}{2}d_t^\top H_td_t+\mathcal R_t, \qquad \lvert\mathcal R_t\rvert\le\frac{M_tr_t^3}{6}.$$ We show precisely how Theorem 2 follows by replacing the directional terms with worst-case bounds. This distinguishes the actual merging geometry from its conservative relaxation. The cubic remainder remains explicit; the bound may be uninformative when this term is large. Third, we extend the private-perturbation analysis. The appendix now gives an exact mean-square error decomposition for the first private query into estimator bias, Poisson sampling variance, and Gaussian noise variance. Combined with the directional approximation lemma, it yields $$0\le S-\mathbb E\widehat S \le L_{\mathrm{sm}}\rho^2 +2\rho\sqrt{\mathbb E\lVert\widetilde g-g\rVert^2}.$$ This quantifies the discrepancy between the noisy directional perturbation and worst-case sharpness. It does not assert that training decreases the full bound. The revised corollary and directional decomposition appear in the main text, with proofs and the extended perturbation analysis in the appendix. We also clarify that the dispersion corollary relies on uniform averaging, whereas the taskwise directional analysis applies to any specified merging output satisfying its assumptions. > **Weakness 3.** The paper acknowledges prior differentially private sharpness-aware training and non-private sharpness-aware approaches to model merging. Although adapting these ideas to private mergeability is meaningful, the narrative and evaluation should more directly provide the increment over the closest existing training procedures under matched privacy budgets and tuning conditions. The component ablations are informative, but do not by themselves establish an advantage over those procedures or clarify the nontrivial contribution beyond adding reference anchoring. **Response to Weakness 3:** Thank you for this comment. **To the best of our knowledge, prior work has not directly investigated whether and how differentially private fine-tuning affects the subsequent mergeability of independently adapted task models.** Existing private sharpness-aware training primarily studies individual-model utility, while non-private merging-aware training methods, including SAFT and MergOPT, do not examine the effects of DP fine-tuning on cross-task mergeability. Our work addresses this interaction, rather than introducing sharpness-aware training itself. The contribution beyond reference anchoring lies in the formulation and investigation of **private mergeability** and the geometric analysis motivating the training design. Specifically, our analysis connects merge-induced loss to local sensitivity, task-model dispersion, and individual-task utility. This explains why controlling reference drift alone is insufficient: it regularizes displacement but does not by itself ensure that task utility is preserved after merging. The two interventions are therefore motivated by complementary aspects of the merging problem, rather than presented as individually novel techniques. We acknowledge that the component ablations establish the roles of these interventions within our framework, but do not replace comparisons with the closest existing training procedures. To address this concern, we conducted additional experiments comparing DP-Merging with DP-SAT (an existing sharpness-aware private training method) and DP-SAT augmented with reference regularization. We report both individual-model accuracy and merged-model accuracy under three merging operators.
AuthorsOpenReview ↗

Response (3/3)

Thank you for your thorough review. We further address Weakness 3 below.

On the 8-task vision benchmark using ViT-B/32 at ε=4\varepsilon=4.

Training MethodIndividual Avg. Acc.Task ArithmeticTIES-MergingWUDI-Merging
DP-AdamW79.754.257.556.5
DP-SAT79.353.656.956.1
DP-SAT + Reference Regularization80.254.857.657.7
DP-Merging81.358.162.160.8

DP-Merging outperforms DP-SAT with reference regularization by 3.3, 4.5, and 3.1 percentage points under Task Arithmetic, TIES-Merging, and WUDI-Merging, respectively, while improving individual accuracy by 1.1 points. These results indicate that its gains extend beyond adding reference regularization to DP-SAT.

Question 1. How do the geometric observations and the gains from DP-Merging change with the degree of data heterogeneity? Could the authors complement the existing cross-task evaluation with controlled same-task experiments, for example, Dirichlet label partitions with varying concentration parameters, and report sharpness, reference drift, and merge-induced degradation under matched privacy budgets?

Response to Question 1: Please see our response to Weakness 1.

Question 2. How informative is the upper bound in Theorem 2 in the evaluated settings, particularly the Hessian-Lipschitz remainder terms? Can the authors analytically compare DP-Merging and baseline procedures under common assumptions, or identify conditions under which DP-Merging yields a smaller bound? Please also clarify how the noisy gradient-direction updates control the worst-case sharpness quantity used in the theorem, and how far the reference-drift argument extends beyond uniform averaging.

Response to Question 2: Please see our response to Weakness 2.

Minor Issues. (1) Section 6.2 contains an unresolved appendix reference ("Appendix ??") for the ViT-B/16 results. (2) Appendix C contains malformed equation references, such as "Eq. equation 15" in C.1 and "Eq. equation 19" in C.2. Please correct them. (3) Sections 6.1 describes the training duration as "10 training epochs/steps." Please specify whether this means epochs or optimization steps and report the corresponding iteration counts used in privacy accounting.

Response to Minor Issues: Thank you for pointing these out. We have corrected the unresolved appendix reference in Section 6.2 and the malformed equation references in Appendix C. All models were trained for 10 epochs. We have corrected this description in Section 6.1.

Reviewer_SBeyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies why differential privacy makes independently fine-tuned models harder to merge, identifying local sharpness and reference drift as two contributing factors. It proposes DP-Merging, which combines sharpness-aware private training with reference alignment to improve merged-model performance while maintaining differential privacy guarantees.

Critical strengths
  • The paper addresses an important and practical problem because model merging is increasingly useful when training data cannot be shared across different organizations.
  • The experiments cover multiple vision and language models, privacy budgets, and merging methods, showing consistent improvements over standard private fine-tuning.
Critical weaknesses
  • The proposed method mainly combines existing sharpness-aware training with a simple regularization technique, making its technical novelty rather limited. Similar ideas have already been explored in private training and model merging, so the paper does not clearly establish what is fundamentally new.

  • The experiments do not clearly separate improvements in individual model accuracy from improvements in mergeability. Without comparing the merging losses of models with similar individual accuracy, it is difficult to determine whether the proposed method actually solves the claimed merging problem.

  • The theoretical analysis does not establish that the proposed training procedure directly improves the derived bounds on merging loss. In particular, the bounds depend on worst-case sharpness and merging displacement, but the paper does not demonstrate that its training procedure reliably controls these quantities or that the bounds predict the observed improvements.

Questions for authors
  • Can the authors compare their method against existing differentially private sharpness-aware training combined with standard reference regularization to clarify the actual novelty?
  • Can the authors report individual model accuracy and merge-induced performance loss separately to show that the improvements come from better mergeability rather than simply better fine-tuning?
  • Can the authors provide empirical evidence that their theoretical bounds become tighter under the proposed method and explain how these bounds relate to the actual merging performance?
5 comments
AuthorsOpenReview ↗

Response (1/3)

We thank you for the helpful comments and suggestions.

Weakness 1. The proposed method mainly combines existing sharpness-aware training with a simple regularization technique, making its technical novelty rather limited. Similar ideas have already been explored in private training and model merging, so the paper does not clearly establish what is fundamentally new.

Response to Weakness 1: Thank you for raising this concern about the novelty of our method. Our work starts with an underexplored question: how does DP fine-tuning affect the mergeability of independently trained task models? To the best of our knowledge, we are the first to systematically study this degradation and its geometric mechanisms.

We first examine the problem empirically and identify local loss sensitivity and reference drift as two associated factors. We then analyze how these factors relate to merge-induced loss and use this analysis to guide our method: sharpness-aware training addresses local perturbation sensitivity, while reference anchoring constrains task-update magnitudes that bound merging displacement.

We agree that these two components are not new individually. The novelty lies in identifying and studying the mergeability problem under DP, and showing how this understanding can guide private training for subsequent merging. Our ablations further examine the contribution of each component.

Weakness 2. The experiments do not clearly separate improvements in individual model accuracy from improvements in mergeability. Without comparing the merging losses of models with similar individual accuracy, it is difficult to determine whether the proposed method actually solves the claimed merging problem.

Response to Weakness 2: Thank you for raising this concern. We report individual-model performance for both standard DP fine-tuning and DP-Merging on the vision and language benchmarks.

For ViT-B/32, average individual accuracy increases from 79.7% to 81.3%, a gain of 1.6 percentage points. In comparison, merged accuracy improves by 3.7–4.7 percentage points across the merging methods in Table 1 of the main paper. The corresponding accuracy drop therefore decreases by 2.1–3.1 percentage points, showing that the gains include both improved individual-model performance and better performance retention during merging.

MethodSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAvg Acc
DP + Individual59.2±0.3267.7±0.4186.1±0.5289.7±0.3587.5±0.4988.6±0.6989.3±0.5569.4±0.4379.7
DP-Merging + Individual60.6±0.4568.8±0.4888.2±0.3890.9±0.3689.2±0.4190.5±0.1991.8±0.2670.5±0.3281.3

For RoBERTa-Large, individual-model performance improves by only 0.3 points, while merged scores increase by 1.4–2 points, suggesting that the gains are not solely due to stronger individual models.

MethodCoLAMNLIMRPCQNLIQQPRTESST2STSBAvg.
DP + Individual60.5±0.2279.1±0.1981.0±0.2989.2±0.1282.3±0.2077.1±0.2193.8±0.1482.1±0.1580.6
DP-Merging + Individual60.6±0.1779.5±0.2481.0±0.3289.6±0.2082.6±0.1477.4±0.2293.7±0.1882.5±0.1580.9

To separate mergeability gains from improvements in individual-model accuracy, we added a post-hoc comparison on the eight-task vision benchmark with ViT-B/32 at ε=4\varepsilon=4. We searched the same learning-rate grid for both methods, η∈{2×10−6,5×10−6,7×10−6,1×10−5,2×10−5,3×10−5,5×10−5}\eta \in \lbrace 2\times10^{-6},5\times10^{-6},7\times10^{-6},1\times10^{-5},2\times10^{-5},3\times10^{-5},5\times10^{-5}\rbrace, and selected configurations with comparable individual-model test accuracy. All other experimental settings follow the main paper.

Table 13. Comparison at similar individual-model accuracy. Accuracies are averaged across eight tasks. Acc. Drop is individual accuracy minus merged accuracy, measured in percentage points.

Merging MethodDP Individual Acc.Ours Individual Acc.DP Merged Acc.Ours Merged Acc.DP Acc. Drop↓Ours Acc. Drop↓
RegMean79.779.952.654.227.125.7
Task Arithmetic79.779.954.257.325.522.6
TIES-Merging79.779.957.560.622.219.3
WUDI-Merging79.779.956.558.423.221.5

Individual-model accuracy differs by only 0.2 percentage points, yet DP-Merging reduces the accuracy drop by 1.4–2.9 percentage points across the four merging methods. This provides additional evidence that the improvements are not solely due to stronger individual models, but also reflect better preservation of task performance during merging.

AuthorsOpenReview ↗

Response (2/3)

We sincerely thank the reviewer for the thoughtful and detailed comments.

Weakness 3. The theoretical analysis does not establish that the proposed training procedure directly improves the derived bounds on merging loss. In particular, the bounds depend on worst-case sharpness and merging displacement, but the paper does not demonstrate that its training procedure reliably controls these quantities or that the bounds predict the observed improvements.

Response to Weakness 3: Thank you for raising this concern. Our analysis connects local loss sensitivity and parameter displacement to merge-induced loss, motivating the two components of DP-Merging. Establishing that private training necessarily decreases the full bound would require a joint analysis of optimization, clipping, noise, and task adaptation, which is beyond our current results.

We have tried to examine this connection empirically from several angles. Figure 2 evaluates test loss along the interpolation paths from individual task models to their merged models on EuroSAT, SUN397, SST-2, and QNLI. DP-Merging exhibits lower loss barriers than standard DP on all four tasks, covering both vision and language settings. This examines loss behavior along the actual merging directions, rather than only comparing final model scores. Figure 3 examines the two proposed geometric factors across privacy budgets on the eight-task vision benchmark. 3(a) shows lower gradient-direction sharpness, while 3(b) shows lower reference drift under DP-Merging. Table 6 separately evaluates the two interventions: removing either sharpness-aware training or reference anchoring reduces merged-model performance, and their combination performs best across the evaluated operators. Thus, the paper includes path-based loss diagnostics, geometric measurements, and component ablations to investigate the proposed explanation.

These experiments provide evidence consistent with our explanation, but do not establish that the full bound decreases or predicts the observed gains. We acknowledge this remaining gap and will revise the discussion to distinguish what the theory establishes from what the experiments support.

Question 1. Can the authors compare their method against existing differentially private sharpness-aware training combined with standard reference regularization to clarify the actual novelty?

Response to Question 1: Thank you for suggesting this comparison. We discuss DP-SAT, an existing sharpness-aware private training method, in our related work. We compare against both DP-SAT and DP-SAT combined with the same reference regularizer, λ∥∣w−w0∥∣22\lambda\||w-w_0\||_2^2, on the eight-task vision benchmark with ViT-B/32. All methods are evaluated at ε=4\varepsilon=4.

Training MethodIndividual Avg. Acc.Task ArithmeticTIES-MergingWUDI-Merging
DP-AdamW79.754.257.556.5
DP-SAT79.353.656.956.1
DP-SAT + Reference Regularization80.254.857.657.7
DP-Merging81.358.162.160.8

In our experiments, DP-SAT performs slightly worse than DP-AdamW in both individual and merged accuracy, indicating that its benefits do not automatically carry over to this merging setting. Adding reference regularization improves DP-SAT, but DP-Merging still outperforms this combined baseline. Individual accuracy improves by 1.1 percentage points, while merged accuracy improves by 3.3, 4.5, and 3.1 points for Task Arithmetic, TIES-Merging, and WUDI-Merging, respectively. The corresponding accuracy drops decrease by 2.2, 3.4, and 2.0 points.

Our main contribution is identifying and systematically studying mergeability degradation under DP, and connecting its geometric characteristics to the design of private training for subsequent merging. For a more detailed discussion of novelty, please see our response to Weakness 1.

AuthorsOpenReview ↗

Response (3/3)

Thank you for your valuable suggestion.

Question 2. Can the authors report individual model accuracy and merge-induced performance loss separately to show that the improvements come from better mergeability rather than simply better fine-tuning?

Response to Question 2: In our response to Weakness 2, we report individual-model performance and analyze merge-induced degradation for ViT-B/32 and RoBERTa-Large. Here, we first present the corresponding results for ViT-L/14 and RoBERTa-Base, whose merged-model results are reported in Tables 2 and 4 of the main text, respectively. We then provide results for ViT-B/16, whose merged-model results are reported in Table 11 of the appendix.

ViT-L/14 on the eight-task vision benchmark at ε=4\varepsilon=4.

MethodSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAvg Acc
DP + Individual64.8±0.3774.2±0.5093.1±0.3392.9±0.2392.2±0.1995.4±0.3496.7±0.4875.6±0.5185.6
DP-Merging + Individual67.4±0.4576.1±0.4394.4±0.2594.6±0.2794.0±0.1696.8±0.3597.7±0.2177.8±0.4587.4

For ViT-L/14, individual accuracy improves by 1.8 percentage points, while merged accuracy improves by 4.1–5.7 points across the merging methods in Table 2.

RoBERTa-Base on GLUE at ε=4\varepsilon=4.

MethodCoLAMNLIMRPCQNLIQQPRTESST2STSBAvg.
DP + Individual59.2±0.4677.7±0.2180.1±0.3888.7±0.1981.5±0.2476.7±0.4392.7±0.1779.4±0.3579.5
DP-Merging + Individual59.6±0.4478.1±0.1780.2±0.2989.0±0.1581.9±0.2477.2±0.2893.3±0.1679.8±0.3379.9

For RoBERTa-Base, the individual-model score improves by only 0.4 points, while merged scores improve by 1.3–2.4 points in Table 4.

ViT-B/16 on the eight-task vision benchmark at ε=4\varepsilon=4.

MethodSUN397CarsRESISC45EuroSATSVHNGTSRBMNISTDTDAvg Acc
DP + Individual59.8±0.4068.2±0.5288.1±0.4586.9±0.3086.2±0.2689.4±0.3390.7±0.1970.6±0.5780.0
DP-Merging + Individual61.1±0.4769.8±0.4289.6±0.3989.6±0.5690.5±0.3391.9±0.2492.7±0.2871.8±0.5282.1

For ViT-B/16, individual accuracy improves by 2.1 percentage points, while merged accuracy improves by 4.0–5.9 points (Table 11), reducing merge-induced degradation by 1.9–3.8 points.

Question 3. Can the authors provide empirical evidence that their theoretical bounds become tighter under the proposed method and explain how these bounds relate to the actual merging performance?

Response to Question 3: Please see our response to Weakness 3.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Novelty and Key Contributions

Dear AC and Reviewers,

We sincerely thank you for the time and effort you devoted to reviewing our work. We truly appreciate the constructive and insightful comments, which have helped us substantially improve the quality and clarity of the paper. We briefly summarize the main novelty and contributions of our work.

1. First systematic study of model mergeability under differential privacy.

  • Research question. Differentially private fine-tuning can lower individual-model accuracy. We ask whether it also increases the accuracy drop after merging. To our knowledge, ours is the first systematic study of this question.

  • Key observation. We find that private task models can retain good single-task accuracy and still degrade substantially after merging. For example, on the eight-task vision benchmark at ε=4\varepsilon=4, standard DP fine-tuning with ViT-B/32 yields 79.7% average accuracy before merging. After weight averaging, accuracy drops to 46.0%, a decrease of 33.7 percentage points.

  • Geometric insight. We identify two geometric factors associated with poorer mergeability. Higher local sharpness makes task loss more sensitive to parameter changes during merging. Larger reference drift means task models move farther from their shared pretrained initialization.

2. Geometry-guided private fine-tuning for model merging.

  • Analysis. We characterize merge-induced loss using the gradient and curvature along the actual merging direction. Under uniform averaging, the average squared reference drift bounds the average squared distance from the task models to the merged model.

  • Method. DP-Merging prepares task models for merging during private fine-tuning. Sharpness-aware training reduces sensitivity to merge-induced parameter changes. Reference anchoring limits drift from the shared initialization. The resulting models can be combined with existing merging operators.

  • Privacy. Our privacy analysis accounts for both private gradient queries in each training step. Merging the released models incurs no additional privacy cost.

3. Empirical evidence of improved mergeability.

  • Main results. DP-Merging improves merged performance across vision and language benchmarks, multiple privacy budgets, and several merging operators.

  • At similar pre-merge accuracy. On the eight-task vision benchmark, DP-AdamW and DP-Merging achieve 79.7% and 79.9% before merging, respectively. DP-Merging reduces the accuracy drop by 1.4–2.9 percentage points across four merging operators.

  • Against a closely matched baseline. At ε=4\varepsilon=4, DP-Merging improves merged accuracy by 3.1–4.5 percentage points across three operators over DP-SAT with the same reference regularizer.

AuthorsOpenReview ↗

Summary of Revisions in Response to Reviewers

Dear Area Chair and Reviewers,

Thank you for your feedback. We summarize the revisions below and indicate where they appear in the revised manuscript.


1. Baseline Comparisons and Mergeability

  • DP-SAT and reference anchoring (SBey W1/Q1, XidJ W3, oTHj W3/Q2). We added DP-SAT with and without the reference regularizer used in DP-Merging (Section 6.2, Table 5). At ε=4\varepsilon=4, DP-Merging improves merged accuracy over regularized DP-SAT by 3.3, 4.5, and 3.1 percentage points with Task Arithmetic, TIES-Merging, and WUDI-Merging, respectively. The same table addresses the effect of anchoring before merging. Adding the regularizer to DP-SAT raises average individual-model accuracy from 79.3% to 80.2%.

  • Mergeability at similar individual-model accuracy (SBey W2/Q2). We report individual-model performance alongside merged performance in Tables 1–4 and 12. We also added a post-hoc comparison at similar pre-merge accuracies (79.7% for DP-AdamW and 79.9% for DP-Merging). DP-Merging reduces the accuracy drop after merging by 1.4–2.9 percentage points across four operators (Appendix B.4, Table 13).


2. Theoretical Analysis and Loss Along the Merging Path

  • Merging direction, remainder, and scope of the bound (SBey W3/Q3, XidJ W2/Q2). We revised Section 5.2, Eqs. (11)–(15) to retain the first-order and curvature terms along the actual merging direction and to state the cubic remainder explicitly. Under uniform averaging, the common gradient component cancels, yielding a tighter bound for a fixed set of task models. Appendices C.2–C.3 provide the proofs. Appendix C.4, Lemma 2 bounds the error of using a noisy gradient direction to approximate worst-case local sharpness. Figure 2 shows the observed test loss along the merging paths, and Figure 3 reports the geometric diagnostics.

3. Additional Experimental Settings

  • Controlled label skew (XidJ W1/Q1). We added CIFAR-10 and CIFAR-100 experiments at two Dirichlet concentrations (Appendix B.4, Table 15). DP-Merging reduces the accuracy drop at both skew levels on both datasets.

  • Additional model architectures (oTHj W1). We added experiments on ConvNeXt-B and Llama-3.2-1B-Instruct. DP-Merging improves merged performance on both architectures with weight averaging and TIES-Merging (Appendix B.4, Table 16).

  • Private parameter-efficient fine-tuning (oTHj Q1). We added comparisons with DP-LoRA and DP-BitFit on the eight-task vision benchmark. DP-Merging achieves higher merged accuracy across all four evaluated operators (Appendix B.4, Table 17).


4. Reporting and Presentation

  • Score and gain annotations (oTHj W2). We recomputed the affected gain annotations from the displayed mean scores in the benchmark tables, including Tables 1 and 6.

  • Cross-references and training duration (XidJ minor comments). We resolved the appendix reference for the ViT-B/16 results in Section 6.2. The results are reported in Table 12. We also corrected the equation references in Appendices C.1–C.2 and clarified in Section 6.1 that the reported training duration was 10 epochs. Appendix C.1 explains how the iteration count enters the privacy accountant.


We greatly appreciate every comment from the reviewers and are thankful for the suggestions that helped us improve the paper. We hope that these revisions address the main concerns and clarify our contributions.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.