acceptodds
Under review as a conference paper at ICLR 2027

BeliefMerge: Bayesian Game Theoretic Model Merging with Joint Beliefs

Abstract

Model merging combines multiple task-specific models fine-tuned from a shared base model into a single model with consolidated capabilities. Yet, model merging is inherently uncertain because each task-specific model cannot directly observe the private characteristics of others, making contribution coordination difficult. Existing methods typically rely on public model information or locally available task information without explicitly accounting for such uncertainty. To address this issue, we introduce BeliefMerge, a Bayesian game-theoretic framework for uncertainty-aware model merging. BeliefMerge constructs correlated joint beliefs over unobserved model characteristics and coordinates interdependent contributions via a Bayesian Nash Equilibrium. Theoretically, sufficient conditions are established for equilibrium uniqueness and convergence of exact best-response updates. Experiments across CLIP and GLUE show that BeliefMerge achieves the best performance on all nine CLIP configurations and five of eight GLUE tasks, demonstrating its empirical superiority.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Model Merging, Bayesian Nash Equilibrium, Correlated Joint Beliefs, Contribution Coordination, Incomplete Information

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_iMM8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces BeliefMerge, a novel Bayesian game-theoretic framework for model merging that coordinates task-specific models by framing them as players with private information. The method constructs correlated joint beliefs via a shared latent variable, allowing models to reason about each other's characteristics, and determines contribution weights through a Bayesian Nash Equilibrium (BNE). Supported by theoretical guarantees for equilibrium uniqueness and convergence, the approach achieves state-of-the-art empirical performance across numerous CLIP and GLUE benchmarks, consistently outperforming strong baselines.

Critical strengths
  1. The paper innovatively reframes model merging as a Bayesian game with asymmetric information rather than a standard centralized optimization problem. By modeling strategic uncertainty via correlated joint beliefs and coordinating through BNE, it provides a principled, theoretically grounded framework (backed by uniqueness and convergence proofs) for handling incomplete information during merging.

  2. BeliefMerge achieves state-of-the-art results across diverse settings, with performance gains scaling favorably alongside task diversity. Furthermore, comprehensive ablation studies effectively validate key design choices—proving that elements like task-vector directions, posterior updating, and Nash coordination all provide genuine, independent benefits beyond simple Bayesian weighting.

Critical weaknesses
  1. Insufficient justification for game-theoretic modeling. The paper does not adequately explain why model merging should be treated as a strategic game with artificial information asymmetry rather than a standard centralized Bayesian optimization problem. To prove this elegant formalization is practically necessary rather than just decorative, the authors must compare the BNE approach against a centralized optimizer with the same information structure, or demonstrate specific scenarios (e.g., conflicting objectives) where strategic coordination fundamentally outperforms centralized optimization.

  2. Limited analysis of when and why the approach succeeds. The paper lacks a deep investigation into the underlying conditions that make BeliefMerge advantageous over simpler alternatives. Specifically, it is missing critical analyses regarding belief calibration quality (the accuracy of inferred beliefs vs. ground truth and the impact of miscalibration), sensitivity to the correlation structure (when and how much correlation is needed compared to independent beliefs), and the characterization of task compatibility that dictates when strategic coordination is truly beneficial.

Questions for authors

1.Could you provide an ablation comparing your BNE-based approach against a centralized Bayesian optimizer that directly optimizes joint utility with full access to all private information, to justify the practical necessity of the game-theoretic view?

2.How sensitive is the method to belief miscalibration; specifically, what is the actual accuracy of inferred beliefs versus ground truth, how does performance degrade with deliberately perturbed priors, and in what exact scenarios do correlated beliefs outperform independent ones?

3.How many strategy optimization iterations are typically needed for convergence, what is the minimum calibration set size required, and when should practitioners prefer BeliefMerge over computationally cheaper baselines?

7 comments
AuthorsOpenReview ↗

Response 1 to Reviewer iMM8

We appreciate your thoughtful review and constructive suggestions. We have carefully considered your feedback and provide the following clarifications to address your concerns.

Weaknesses 1/Question 1:Insufficient justification for game-theoretic modeling...

We thank the reviewer for highlighting the need to clarify the motivation and practical benefits of our game-theoretic modeling. BMM (Li et al., 2026a) uses a merged model obtained by methods such as TA (Ilharco et al., 2023) or TSV (Gargiulo et al., 2025) as a prior and estimates merged parameters through Bayesian regression using task-specific activations and residual outputs. It then uses Bayesian optimization to jointly tune module-specific hyperparameters based on average validation performance. Since task-specific models share the same pretrained base, their private information may be statistically dependent. Differences in task data and training further introduce uncertainty about other models private information and contributions. BMM does not explicitly model how other models unobserved private information affects their contributions and, in turn, a model own contribution decisions. BeliefMerge addresses this gap by forming joint beliefs about other models private information, propagating this uncertainty through their contribution strategies to distributions over actions, and coordinating interdependent contributions through conditional expected utilities.

Fig. 6 includes w/o NO, a centralized expected-utility optimizer with the same information structure as BeliefMerge. Each contribution strategy remains bm=σm(sm,o)b_m=\sigma_m(s_m,o), as in Eq. (11). This variant retains the joint beliefs, utilities, and merging rule, but replaces the individual best responses in Eq. (16) with centralized maximization of the average expected utility, max⁡σEs∼π(⋅∣o)\max_{\sigma}\mathbb{E}_{s\sim\pi(\cdot\mid o)}. On the 20-task CLIP benchmark, accuracy decreases from 86.4% and 86.9% to 84.2% and 85.0% on ViT-B/32 and ViT-B/16, respectively, corresponding to drops of 2.2 and 1.9 percentage points. These results support the practical benefit of Nash coordination over centralized average expected utility maximization under the same information structure.

To further address the reviewer suggestion, we add Centralized BO as a centralized Bayesian optimization baseline that directly maximizes joint utility with full access to all models private information. Specifically, we compute the true private statistics sm⋆s_m^\star from each model local data using Eqs. (5)–(7) and provide them to the optimizer. We replace Eq. (16) with centralized joint utility maximization, b⋆=arg⁡max⁡b1M∑m=1Mum(b;sm⋆,o)\mathbf{b}^{\star}=\underset{\mathbf{b}}{\arg\max}\frac{1}{M}\sum_{m=1}^{M}u_m(\mathbf{b};s_m^{\star},o).

Table A. Comparison of centralized joint utility optimization with full information access and BeliefMerge.

MethodViT-B/32-20ViT-B/16-20
Centralized BO85.185.8
BeliefMerge86.486.9

Table A shows that, compared with Centralized BO, BeliefMerge further improves accuracy by 1.3 and 1.1 percentage points on ViT-B/32 and ViT-B/16, respectively. These results show that coordinating contributions through individual best responses yields higher merging accuracy than centralized average utility maximization, providing empirical support for the game-theoretic formulation. We have added these comparisons and the corresponding experimental details to Section 4.3.

AuthorsOpenReview ↗

Response 2 to Reviewer iMM8

Weaknesses 2/Question 2:Limited analysis of when and why the approach succeeds...

We thank the reviewer for these valuable suggestions. Regarding belief calibration quality, we compare inferred private information with ground-truth statistics computed from each model’s local data. In Fig. 7, RMSE decreases from 0.57 at 2 steps to 0.06 at 2,048 steps. Meanwhile, merging accuracy reaches 89.2% at 512 steps and remains stable through 2,048 steps. These trends are consistent with more accurate private information estimates supporting contribution coordination and sustained merging performance.

Regarding sensitivity to the correlation structure, the Independent Beliefs (IB) variant in Fig. 5 replaces the joint belief with the product of its conditional marginals, removing cross-model dependence while preserving the conditional marginal distributions. This reduces accuracy by 1.2 and 1.4 percentage points on ViT-B/32 and ViT-B/16, respectively, supporting the benefit of modeling dependence beyond marginal information in the evaluated settings. And regarding strategic coordination across tasks, w/o NO in Fig. 6 replaces individual best responses with centralized maximization of average expected utility while retaining the joint beliefs, utility functions, and merging rule. The accuracy decreases by 2.2 and 1.9 percentage points on ViT-B/32 and ViT-B/16, respectively, supporting the practical benefit of Nash coordination.

To further examine when correlated beliefs and Nash coordination are beneficial, we evaluate three eight-task subsets from the 20-task CLIP benchmark on ViT-B/32. Using the average directional similarity in Eq. (27), we select subsets closest to the 10th, 50th, and 90th percentiles, representing low, medium, and high similarity, respectively. We compare full BeliefMerge with two existing variants: IB (Fig. 5), which replaces the joint belief in Eq. (10) with the product of its conditional marginals, and w/o NO (Fig. 6), which replaces individual best responses in Eq. (16) with centralized maximization of average expected utility. These comparisons assess how the benefits of correlated beliefs and Nash coordination vary across task combinations. Table B shows that gains from correlated beliefs and Nash coordination decrease from 1.5 to 0.9 and from 2.8 to 1.3 percentage points as task-vector similarity increases, suggesting greater benefits for less directionally similar task combinations in the evaluated subsets. We have added these comparisons and the corresponding experimental details to Section 4.3.

Table B. Benefits of correlated beliefs and Nash coordination across task combinations. Low, Medium, and High denote the three eight-task subsets.

SubsetIBw/oNOBeliefMerge
Low87.686.389.1
Medium87.787.288.9
High87.987.588.8

Question 3: How many strategy optimization iterations are typically needed for convergence...

We thank the reviewer for these valuable suggestions. We do not require an additional independent calibration dataset: the belief model learns from existing public information and local statistics during training Appendix D.3. Our training directly optimizes the contribution strategies. Fig. 7 shows that merging accuracy reaches 89.2% at the 512-step task-model fine-tuning checkpoint and remains stable through 2,048 steps.

Regarding practical use, Table 7 reports a merging time of 50.14 seconds on the eight-task CLIP ViT-B/32 benchmark, compared with 40.64 seconds for TSV-M (Gargiulo et al., 2025) and 62.92 seconds for Iso-CTS (Marczak et al., 2025), which represent subspace-based merging. BeliefMerge requires only 1.81 seconds more than the strongest overall baseline (Crisostomi et al., 2026), which represents input-dependent merging, while using 29.0% less peak memory. BeliefMerge uses 4.45 GB of peak memory, compared with 3.59–3.92 GB for Task Arithmetic, TSV-M, and Iso-CTS. Specifically, Task Arithmetic (Ilharco et al., 2023), representing simple parameter arithmetic, takes only 0.08 seconds, but its accuracy is 68.8% compared with BeliefMerge 89.2% under the same benchmark setting—a gap of 20.4 percentage points (Table 1). We therefore recommend BeliefMerge when merging quality is a priority and local task statistics are available; Task Arithmetic may be preferable when merging time is more important than accuracy.

AuthorsOpenReview ↗

Thank you & Looking forward to further discussion!

Dear Reviewer iMM8,

Thank you again for your constructive review. We briefly summarize our key clarifications and additional evidence:

  1. Practical Benefits of Game-Theoretic Modeling (Weakness 1; Question 1): BeliefMerge forms correlated joint beliefs about other models’ unobserved private information and coordinates interdependent contributions through conditional expected utilities. In Fig. 6, w/o NO replaces individual best responses with centralized average expected-utility maximization while retaining the information structure, beliefs, utilities, and merging rule. Furthermore, Table A compares BeliefMerge with Centralized BO using all models’ true private statistics. Both comparisons support the practical benefit of Nash coordination.

  2. Belief Estimation and Task Compatibility (Weakness 2; Question 2): Fig. 7 shows decreasing private-statistic estimation error alongside sustained merging accuracy. Independent Beliefs in Fig. 5 removes cross-model dependence while preserving conditional marginals, and w/o NO isolates Nash coordination. Furthermore, Table B compares these variants across task subsets with different directional similarities, showing greater benefits from correlated beliefs and Nash coordination for less similar combinations within the evaluated subsets.

  3. Efficiency and Practical Use (Question 3): Table 10 in Appendix F shows modest additional merging time and lower peak memory than the strongest overall baseline. Task Arithmetic is faster but achieves lower merging accuracy (Table 1). These comparisons support BeliefMerge when merging quality is the priority and local task statistics are available, while Task Arithmetic may be preferable when merging time matters more.

We appreciate your valuable suggestions and look forward to further discussion.

Warm regards,

Authors

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

This paper proposes BeliefMerge, which formulates the fusion of multiple models as a Bayesian game: each task model leverages publicly available model information and local private information to construct correlated joint beliefs about the private features of other models, and then employs a Bayesian Nash equilibrium to coordinate the fusion weights among these models. The paper establishes sufficient conditions for the existence and uniqueness of such equilibria, as well as for the convergence of exact best-response strategies, and conducts extensive evaluations on multiple CLIP tasks and the GLUE benchmark. Across all nine configurations of CLIP, our method outperforms all compared approaches; on the GLUE benchmark, it achieves top performance in five of the eight tasks and second-best performance in three tasks.

Critical strengths
  1. The approach to the problem is novel. Formulating model merging as a Bayesian game under incomplete information, and using relevant joint beliefs to characterise the private features of other models that are not observable, better captures the issue of interdependence of contributions across models than independent beliefs or purely heuristic weights.
  2. The experiments are fairly comprehensive. They cover 8/14/20 task setups on CLIP ViT-B-32/B-16/L-14 and the GLUE 8 tasks, and include main results, ablation studies, dynamic tasks, and efficiency and hyperparameter analyses; the model outperforms strong baselines in most setups, with particularly notable improvements on the 20-task CLIP dataset.
Critical weaknesses
  1. The results of the equilibrium theory are inconsistent with the core motivation of the paper. Appendix D.4 indicates that, under the sufficient condition (Equation 44), the exact equilibrium strategy is almost independent of the model’s private information, whereas the core of the paper lies in each model dynamically adjusting its contributions using private information. Although the authors explain that strategies obtained through a finite number of training steps may depend on private information, they have not demonstrated that the actual trained strategies approximate such non-constant strategies, nor have they explained how the theoretical guarantees correspond to the actual benefits reported in the paper. This affects the interpretation of the method and its core contribution.
  2. Baseline results are taken directly from the original papers, while BeliefMerge is evaluated over five independent runs. Although the paper reports its own data splits and evaluation metrics, it does not explicitly clarify whether these settings are fully consistent with those used for each baseline.
Questions for authors

Under theoretical conditions, does BeliefMerge degenerate into constant weights independent of smsm? If so, does the actual experimental benefit stem primarily from finite-step approximation, policy network parameterisation, or regularisation?

3 comments
AuthorsOpenReview ↗

Response 1 to Reviewer oMPw

Thank you for your thorough review and insightful comments on our manuscript. We carefully consider your feedback and make the following clarifications to address the concerns raised.

Weaknesses 1/Question 1:The results of the equilibrium theory are inconsistent with the core motivation of the paper...

Thank you for highlighting the connection between our equilibrium theory and the role of private information. Under Eq. (44), exact best-response updates converge to the unique BNE, and the equilibrium weights are constant (cmc_m) with respect to sms_m for fixed public information omo_m. The value of cmc_m can vary across models and change as the public information evolves with the task vectors. This analysis establishes equilibrium uniqueness and convergence of exact best-response updates under the stated sufficient condition (Eqs. (48)–(49)). We have added this clarification to Appendix D.4.

The contribution strategies trained in Appendix D.3 use private information as input and optimize conditional expected utilities through approximate best-response updates (Eqs. (10)–(16)). Our controlled comparisons identify the benefits of private-information inputs and posterior updating within this optimization procedure. In Fig. 5, Prior Only (PO) uses the same strategy network, regularization, private-information inputs, and training budget, while replacing the conditional posterior with the prior. Its accuracy decreases by 3.0 and 3.8 percentage points on ViT-B/32 and ViT-B/16, respectively, demonstrating the additional benefit of posterior updating under matched parameterization and regularization. Fig. 4 further shows that removing task sensitivity, representation relevance, and performance benefit reduces accuracy by 1.5/3.2, 0.9/1.3, and 0.6/0.9 points, respectively. The theory characterizes the exact equilibrium limit and its convergence, while these controlled experiments quantify the performance benefits of private-information inputs and posterior inference in the trained optimizer.

AuthorsOpenReview ↗

Response 2 to Reviewer oMPw

Weaknesses 2:Baseline results are taken directly from the original papers, while BeliefMerge is evaluated over five independent runs...

Thank you for highlighting the need to clarify the experimental settings. Following prior work (Crisostomi et al., 2026; Nguyen et al., 2026b; Yang et al., 2024b), all comparisons use consistent benchmarks and model backbones, as described in lines 328–333. The dataset splits and evaluation metrics are also aligned with prior work. Specifically, Appendix E details the CLIP data partitions adopted from MASS and the Flan-T5 evaluation settings adopted from FusionBench (Tang et al., 2025a). The detailed splits and metrics are provided in Tables 5–6. These settings ensure consistent evaluation protocols across the compared methods. In addition, to clarify differences in information access, Tables 1 and 2 are reorganized into two groups: methods using only public information and methods accessing both public and private information. Appendix C.1 further identifies the specific types of private information used by each method, making the distinctions in information access explicit.

To assess whether the performance gains arise from differences in information access, we add comparisons using the same types of public and private information as model inputs with identical experimental settings. Specifically, BeliefMerge uses only public information and representation relevance when compared with RegMean++, WeMoE, SMILE, and MASS, and only public information and performance benefit when compared with AdaMerging. Accordingly, Eq.(4) is modified to sm=sm,reprs_m=s_{m,\mathrm{repr}} or sm=sm,perfs_m=s_{m,\mathrm{perf}}, respectively. As shown in Table A, BeliefMerge achieves the highest accuracy across all nine representation-relevance settings and all six performance-benefit settings, outperforming the strongest baseline in each setting by 0.1-3.4 and 2.4-15.1 percentage points, respectively. These results demonstrate that BeliefMerge maintains its performance advantage under matched information access and evaluation settings, indicating that the improvements are not simply from additional types of private information. We have added these comparisons and the corresponding experimental details to Appendix C.4.

Table A. Comparisons under the same experimental settings with matched private-information categories.

MethodViT-B/32–8ViT-B/32–14ViT-B/32–20ViT-B/16–8ViT-B/16–14ViT-B/16–20ViT-L/14–8ViT-L/14–14ViT-L/14–20
Representation relevance
RegMean++84.4—74.287.2—76.991.0—82.7
WeMoE88.874.368.289.176.665.088.772.365.0
SMILE-183.275.472.887.881.779.591.286.684.9
SMILE-284.476.474.189.082.780.492.087.185.5
MASS87.082.981.190.687.881.192.990.990.8
BeliefMerge88.985.184.591.288.983.293.691.891.3
Performance benefit
AdaMerging80.1—69.384.9—72.890.8—79.8
BeliefMerge88.2—84.490.7—82.993.2—91.1
AuthorsOpenReview ↗

Thank you & Looking forward to further discussion!

Dear Reviewer oMPw,

Thank you again for your thorough review. We briefly summarize our key clarifications:

  1. Equilibrium Theory and Empirical Benefits (Weakness 1; Question 1): Under the sufficient condition in Eq. (44), exact best-response updates converge to a unique equilibrium (Appendix D.4; Eqs. (48)–(49)). For fixed public information, equilibrium strategies are constant in private information, but their values can differ across models and vary as public information evolves. The approximate optimizer in Appendix D.3 uses private information and conditional beliefs. In Fig. 5, PO replaces the posterior with the prior while retaining private-information inputs, other components, and the training budget, reducing accuracy. Together with the private-information ablations in Fig. 4, this supports the benefits of posterior updating and private information in the trained optimizer.

  2. Evaluation Consistency and Information Access (Weakness 2): We follow consistent benchmarks, backbones, splits, and metrics, adopting MASS’s CLIP partitions and FusionBench’s Flan-T5 protocol (Appendix E; Tables 8–9). Tables 1–2 and Appendix C.1 clarify information access. Furthermore, Table A in our response matches experimental settings and private-information categories: representation relevance for RegMean++, WeMoE, SMILE, and MASS, and performance benefit for AdaMerging. BeliefMerge retains its advantage, showing that the gains are not simply due to additional information categories (Appendix C.4).

We hope these clarifications and additional comparisons address your main concerns and welcome further discussion.

Warm regards,

Authors

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

The paper introduces a game-theoretic framework for model merging that uses the combination of private and public information to construct joint beliefs. The authors formulate the model contribution as a Bayesian game, optimizing the model contributions towards Bayesian Nash Equilibrium. They also provide theoretical guarantees for uniqueness and convergence, and evaluate their method with CLIP and GLUE tasks

Critical strengths
  • The proposed method addresses existing limitations in model merging and proposes a reasonable and principled solution. The method also appears to be effective and is competitive with existing baselines across many of their benchmarks. The paper is also very well-written: the problem is well-motivated, the authors provide a compelling example showing the challenges of model merging, and the method is presented intuitively.
  • The ablations are informative and thorough. I appreciate how the authors broke down every part of their method, such as the amount of information, the choice of prior, the Bayesian Nash equilibrium setup, etc, and presented variations for each of their choices. I think these sections helped me learn a lot about why their method works and understand alternative formulations.
Critical weaknesses
  • The empirical results could be more clear about what sources of information the baselines have access to. Although the paper distinguishes between public vs private methods, there are also other differences such as the type of private information used. Furthermore, there are a few other differences such as whether they use uncertainty modeling, how they combine weights, etc. Since the baselines were taken from prior work, it's difficult to understand how much of the performance gains come from the method itself vs differences in experimental settings.
  • nit: The notation in figure labels can be a bit too dense and requires a lot of jumping between the text and the figure to understand. I think having more descriptive labels or captions could make the paper easier to follow.
Questions for authors
  • Many of the other model merging methods eventually have decreased test accuracy. Is that something you encountered with BeliefMerge when finetuning for longer?
5 comments
AuthorsOpenReview ↗

Response 1 to Reviewer QeFJ

We sincerely thank the reviewer for the valuable insights. We have carefully read the comments and our detailed point-by-point response is provided below:

Weaknesses 1:The empirical results could be more clear about what sources of information the baselines have access to...

Thank you for highlighting the need to clarify the information sources and experimental settings. Following prior work (Crisostomi et al., 2026; Nguyen et al., 2026b; Yang et al., 2024b), the comparisons in Tables 1 and 2 follow consistent task benchmarks, model backbones, data splits, and evaluation metrics. To make information access explicit, Tables 1 and 2 are reorganized into two groups: methods using only public information and methods additionally accessing private information. Appendix C.1 further specifies the types of information available to each method, distinguishing information access from methodological differences such as uncertainty modeling and weight combination.

To further disentangle performance gains from differences in information access, the following comparison is considered with identical experimental settings, where BeliefMerge and the corresponding baselines use the same categories of information. Specifically, BeliefMerge uses public information and representation relevance when compared with RegMean++, WeMoE, SMILE, and MASS. For comparison with AdaMerging, BeliefMerge instead uses public information and performance benefit, matching the information categories available to that baseline. Table A shows that BeliefMerge achieves the highest accuracy in all nine and six evaluated settings, respectively, exceeding the strongest baseline in each setting by 0.1–3.4 and 2.4–15.1 percentage points. These results show that the performance gains persist when information types and experimental settings are matched, supporting the effectiveness of our inference and contribution coordination under uncertainty rather than access to additional types of private information. We have added these comparisons and the corresponding experimental details to Appendix C.4.

Table A. Comparisons under the same experimental settings with matched private-information categories.

MethodViT-B/32–8ViT-B/32–14ViT-B/32–20ViT-B/16–8ViT-B/16–14ViT-B/16–20ViT-L/14–8ViT-L/14–14ViT-L/14–20
Representation relevance
RegMean++84.4—74.287.2—76.991.0—82.7
WeMoE88.874.368.289.176.665.088.772.365.0
SMILE-183.275.472.887.881.779.591.286.684.9
SMILE-284.476.474.189.082.780.492.087.185.5
MASS87.082.981.190.687.881.192.990.990.8
BeliefMerge88.985.184.591.288.983.293.691.891.3
Performance benefit
AdaMerging80.1—69.384.9—72.890.8—79.8
BeliefMerge88.2—84.490.7—82.993.2—91.1

Weaknesses 2:nit: The notation in figure labels can be a bit too dense and requires a lot of jumping between the text and the figure to understand...

Thank you for the suggestion. We have expanded the figure captions and added descriptive annotations to make the figures easier to follow without repeatedly referring to the main text. Specifically, in Fig. 5, we explain LP (Linear Predictor), PO (Prior Only), and IB (Independent Beliefs) directly below the plot. In Fig. 6, we provide the full names of BPI (Bayesian Posterior Inference), NO (Nash Optimization), GPM (Gaussian Posterior Merging), and GNB (Gradient Nash Bargaining). We also explicitly label the two halves of Fig. 6 as absolute accuracy and normalized accuracy to clarify the reported metrics.

AuthorsOpenReview ↗

Response 2 to Reviewer QeFJ

Question 1: Many of the other model merging methods eventually have decreased test accuracy...

We thank the reviewer for this insightful comment. BeliefMerge shows greater robustness to longer fine-tuning, maintaining its peak merging accuracy as several baselines decline. Addressing the decline in merging performance with longer fine-tuning is a central motivation of our work, see Section 2,lines 152–155, and Fig. 2. Additionally, in Fig. 7, all task-specific models are fine-tuned for up to 2,048 steps and merged at successive checkpoints. BeliefMerge reaches 89.2% accuracy at 512 steps and maintains it through 2,048 steps. MASS peaks at approximately 87% at 512 steps, while TSV-M and AdaMerging peak at approximately 83.2% and 80.1% at 256 steps, respectively. By 2,048 steps, their accuracies decrease to 84.6%, 79.4%, and 77.5%, respectively, while BeliefMerge remains at 89.2%, outperforming the strongest baseline by 4.6 percentage points.

This stability can be attributed to BeliefMerge's adaptive coordination of merging contributions in response to changes in task-model interactions during continued fine-tuning. Eq. (10) updates joint beliefs about other models unobserved private information, and Eqs. (15)–(16) coordinate contributions, which are converted into merging weights through Eq. (12). This allows a model weight to respond to changes in other models, even when its own performance continues to improve, helping maintain merging accuracy during longer fine-tuning. Fig. 7 further shows that the RMSE between estimated private information and the corresponding ground-truth values decreases from 0.57 to 0.06, indicating increasingly accurate estimates for contribution coordination.

AuthorsOpenReview ↗

Thank you & Looking forward to further discussion!

Dear Reviewer QeFJ,

Thank you again for your thoughtful review. We briefly summarize our key clarifications:

  1. Information Access and Comparison Fairness (Weakness 1): Tables 1–2 distinguish methods using only public information from those additionally accessing private information, with categories specified in Appendix C.1. Furthermore, Table A in our response matches experimental settings and private-information categories: representation relevance for RegMean++, WeMoE, SMILE, and MASS, and performance benefit for AdaMerging. BeliefMerge retains its advantage, supporting its inference and contribution coordination beyond access to additional information categories (Appendix C.4).

  2. Clarity of Figure Notation (Weakness 2): We expand the captions and annotations in Figs. 5–6, explain abbreviated variant names below the plots.

  3. Robustness to Longer Fine-Tuning (Question 1): Fig. 7 shows that BeliefMerge maintains peak merging accuracy within the evaluated range while several baselines decline. Joint-belief updates (Eq. (10)) and contribution coordination (Eqs. (15)–(16)) adapt merging weights (Eq. (12)) to evolving task-model interactions. Decreasing private-statistic estimation error further supports increasingly accurate estimates for coordination.

We hope these clarifications and additional comparisons address your main concerns and welcome further discussion.

Warm regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Final Remarks

Dear Area Chair and Reviewers,

We sincerely thank Reviewers QeFJ, oMPw, and iMM8 for their constructive feedback. In particular, we appreciate their recognition of the novelty of our method (Reviewers oMPw and iMM8) and the thoroughness of our experimental design (Reviewers QeFJ, oMPw, and iMM8).

BeliefMerge is a Bayesian game-theoretic framework that constructs correlated joint beliefs over unobserved model characteristics and coordinates interdependent contributions via a Bayesian Nash Equilibrium. Across all nine CLIP configurations and the GLUE benchmark, BeliefMerge achieves the highest average accuracy among the compared merging methods.

In addition, we summarize below the experiments and clarifications from our responses, organized by the main concerns raised in the reviews.

  1. Comparison Fairness and Information Access (Reviewer QeFJ: Weakness 1; Reviewer oMPw: Weakness 2): We clarify that the baseline comparisons use consistent benchmarks, model backbones, data splits, and evaluation metrics. Tables 1–2 group methods by information access, and Appendix C.1 specifies the information available to each method. Furthermore, Table A in each response to Reviewers QeFJ and oMPw shows that BeliefMerge retains its advantage under matched private-information categories and experimental settings (Appendix C.4).

  2. Equilibrium Theory and Practical Optimization (Reviewer oMPw: Weakness 1; Question 1): Appendix D.4 clarifies that, under the sufficient condition in Eq. (44), exact best-response updates converge to a unique equilibrium. For fixed public information, each model’s equilibrium contribution is constant in its private information. Its value can differ across models and change as public information evolves with the task vectors. For the contribution strategies trained in Appendix D.3, the Prior Only (PO) comparison in Fig. 5 shows the benefit of posterior updating, and the ablations in Fig. 4 show the benefit of private information.

  3. Robustness, Belief Quality, and Task Combinations (Reviewer QeFJ: Question 1; Reviewer iMM8: Weakness 2; Question 2): Fig. 7 evaluates merging across successive fine-tuning checkpoints, showing sustained accuracy and decreasing private-statistic estimation error. The IB and w/o NO variants in Figs. 5–6 examine correlated beliefs and Nash coordination. Furthermore, Table B in our response to Reviewer iMM8 shows that correlated beliefs and Nash coordination yield larger gains on less directionally similar task subsets. Table C shows that merging accuracy is insensitive to the tested belief-estimation errors.

  4. Calibration and Practical Cost (Reviewer iMM8: Question 3): Belief calibration uses existing public information and local statistics without an additional independent dataset; Appendix D.3 describes cyclic strategy optimization. Tables 10 support using BeliefMerge when merging quality is the priority and local statistics are available, while Task Arithmetic may be preferable when merging time matters more.

  5. Figure Clarity (Reviewer QeFJ: Weakness 2): We expand captions and annotations in Figs. 5–6, explain abbreviations, and distinguish absolute from normalized accuracy.

  6. Comparison with Centralized Optimization (Reviewer iMM8: Weakness 1; Question 1): Fig. 6 compares Nash coordination with centralized average expected-utility maximization while retaining the information structure, beliefs, utilities, and merging rule. In addition, Table A in our response to Reviewer iMM8 compares Centralized BO with full access to true private statistics. Both comparisons support the empirical benefit of individual best-response coordination.

After considering our clarifications and additional experiments, Reviewers QeFJ and iMM8 acknowledged the progress we made in addressing their main concerns.

Again, we sincerely thank the reviewers for their thoughtful feedback, which has improved the clarity and completeness of the manuscript. We welcome further discussion and will incorporate their feedback into the final version.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.