Pith. sign in

REVIEW 3 major objections 78 references

Merging models can stand in for multi-task retraining when pricing private datasets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · grok-4.5

2026-07-12 03:07 UTC pith:2BMJPIN2

load-bearing objection Solid multi-task + privacy data-valuation paper that pairs clean Shapley-error bounds with usable experiments; the second-order merging assumption is the real soft spot but does not sink the contribution. the 3 major comments →

arxiv 2607.03346 v1 pith:2BMJPIN2 submitted 2026-07-03 cs.CL cs.AIcs.CVcs.LG

Efficient Decentralized Multi-task Dataset Valuation via Model Merging

classification cs.CL cs.AIcs.CVcs.LG
keywords dataset valuationmulti-task learningmodel mergingtask arithmeticShapley valuesecure aggregationdata marketplaceprivacy-aware ML
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that data marketplaces for multi-task models need a way to price each seller's dataset without seeing the data or retraining on every possible coalition of sellers. DMVM does this by having each seller fine-tune a shared base model, then treating arithmetic merges of those models as a stand-in for the multi-task model that would have been trained on the combined data. The buyer scores the merged models on its own validation tasks and turns those scores into Shapley-style seller values. A secure aggregation protocol with an unmergeable transformation keeps individual models and raw data hidden. Error bounds link the approximation quality of the merge to the gap from true multi-task Shapley values, and experiments on vision and language tasks show the rankings and numerical values stay close to privileged retraining baselines while avoiding data access and coalition training.

Core claim

Task-arithmetic model merging can serve as a faithful, privacy-compatible proxy for multi-task coalition utility, so that Shapley-style multi-task dataset values can be estimated from merged models alone rather than from retraining on every seller subset.

What carries the argument

DMVM: each seller's fine-tuned model is turned into a task vector; for any coalition the buyer evaluates a scaled sum of those vectors (task arithmetic) as the utility of that coalition, then aggregates the marginals into approximate Shapley values, with secure masks and an unmergeable layer transform protecting individual models.

Load-bearing premise

The method assumes that a simple average of task vectors is close enough to a model trained jointly on the same data that the difference in loss is well controlled by a second-order bound; if that gap is large, the value estimates drift.

What would settle it

On a marketplace where true multi-task Shapley values can still be computed by full coalition retraining, measure whether the absolute gap |true value − DMVM value| stays inside the stated bound (2αn² − 2αn + 2n + α − 1)LC, and whether rank correlation collapses when the second-order remainder is large (e.g., large learning rate or many epochs).

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper proposes DMVM, a framework for multi-task dataset valuation in decentralized marketplaces that approximates coalition utilities via task-arithmetic model merging rather than retraining on raw seller data. Sellers fine-tune a shared pretrained model locally; the buyer evaluates merged models on multi-task validation sets and estimates Shapley values. A secure aggregation protocol with unmergeable layer transformations is introduced so that individual models and raw data are never revealed. Theorems 4.1–4.3 bound the gap between true multi-task Dataset Shapley values and the merging-based (and privacy-aware) proxies under Lipschitz loss and a second-order merging approximation. Experiments on vision and NLP marketplaces report Kendall τ roughly 0.75–0.85 and low MSE against privileged retraining-based ground truth, with MC-DMVM improving the efficiency–accuracy trade-off.

Significance. If the claims hold, this is a useful contribution at the intersection of data valuation, model merging, and privacy-preserving marketplaces. Multi-task dataset valuation is underexplored; most prior work is single-task and assumes centralized data access. Using task arithmetic as a utility proxy is a natural and timely idea, and the paper pairs it with explicit error bounds, a privacy protocol, and systematic vision/NLP experiments including controlled relevance spectra and harmful-data detection. The MC variant and scalability projections make the approach more practical than exponential retraining. Strengths include clear problem formulation, non-circular evaluation against independently computed multi-task Shapley, and concrete algorithmic detail (Algorithm 1). The main significance is practical: a deployable valuation path when raw data cannot be shared and buyers care about multiple tasks.

major comments (3)
  1. Theorems 4.1–4.3 rest on the starting inequality ||θ_merge_S − θ_MTL_S||₂ ≤ s(αs+1)C from Zhou et al., which is invoked only after higher-order remainder O(η³) is declared negligible (Appendix A.4; also stated in Thm 4.1). The subsequent Lipschitz step and convex-combination argument are clean, but the paper never measures ||θ_merge − θ_MTL|| or the realized loss/utility gap on the same coalitions that enter the Shapley sums for the models, optimizers, and short fine-tuning regimes used in §5 and A.10. Without that check, the absolute bounds are not shown to be informative for the experimental setting; if higher-order terms dominate, C no longer controls the gap and the error chain is non-informative. Please either (i) report empirical merge–MTL parameter and utility gaps on representative coalitions, or (ii) clearly reframe the theorems as conditional on that assumption and de-emphasize
  2. §4.2 and Thm 4.3 exclude |S|<2 for privacy, adding a 2R/n term. The unmergeable transformation (permutation on MLP pairs; scalar ρ on consecutive linear/QK layers) is presented as preventing reconstruction by differencing overlapping merges. The security argument is informal: there is no threat model, no analysis of residual leakage under adaptive subset queries, and no discussion of whether other modules (LayerNorm, embeddings, residual paths, attention value/output projections) remain linearly reconstructible. For a claimed privacy-aware marketplace protocol this is load-bearing. Either provide a clearer adversary model and argument that the full architecture is covered, or qualify the protocol as a heuristic defense against simple algebraic unmerging rather than a complete privacy guarantee.
  3. Ground-truth and privileged baselines in §5 / A.10 are computed with only 2–3 epochs of coalition training, while DMVM uses independently fine-tuned seller models merged with α=1/|S|. Short training can make the multi-task optimum closer to a linear combination of task vectors, potentially inflating agreement between merge-based and retraining-based utilities. Tables 2–3 and Figures 2–4 therefore may overstate fidelity relative to a more fully optimized MTL oracle. Please report at least one ablation with longer coalition training (or stronger multi-task baselines) for a subset of n, or explicitly discuss this as a limitation of the evaluation design.

Circularity Check

0 steps flagged

No significant circularity: DMVM proxy and error bounds are not defined in terms of the recovered Shapley rankings, and evaluation uses independently retrained multi-task ground truth.

full rationale

The derivation chain is: (i) define true coalition utility u(S) via multi-task training on raw seller subsets; (ii) define proxy û(S) via task-arithmetic merging of independently fine-tuned models; (iii) bound |u−û| using the external Zhou et al. second-order merging gap plus Lipschitz continuity; (iv) lift that gap through the Shapley linear combination (and Hoeffding for MC) to obtain Theorems 4.1–4.3. None of these steps defines the proxy in terms of the final seller ranking or fits a free parameter to the target Shapley values and then re-labels the fit as a prediction. Empirical claims compare DMVM/MC-DMVM against Dataset Shapley, MC-Dataset-Shapley, and DU-Shapley computed by retraining on raw coalitions—an independent oracle, not a quantity recovered by construction. Self-citations (e.g., Amiri et al. on task-agnostic valuation, Alipour & Amiri on reversible merging) appear only in related work and do not underwrite the error bounds or the numerical rankings. The load-bearing soft spot is an external modeling assumption (higher-order O(η³) terms negligible), which is a correctness risk, not circularity. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The central approximation rests on a second-order merging lemma imported from Zhou et al., Lipschitz continuity of the loss, full-batch gradient descent for the multi-task baseline, and the modeling choice that average validation performance is the right multi-task utility. The free parameter α and the privacy exclusion of singleton coalitions are additional knobs that affect the stated bounds. No new physical entities are postulated; the invented constructs are algorithmic.

free parameters (2)
  • α (task-arithmetic scaling)
    Appears in the merged-model definition and in every error bound; set to 1/|S| in experiments but treated as a free hyper-parameter that can be tuned by binary search.
  • C = (h+2 choose 2) H_max G_max
    Composite constant bounding gradient and Hessian norms over h epochs; not measured, only assumed finite.
axioms (4)
  • domain assumption Difference between task-arithmetic merge and full multi-task GD is O(η³) beyond second order (Zhou et al. lemma).
    Load-bearing starting point of Theorem 4.1; invoked without re-derivation for the architectures used.
  • domain assumption Loss L is L-Lipschitz continuous in model parameters.
    Converts parameter-space distance into utility error; standard but unverified for the concrete losses.
  • ad hoc to paper MTL baseline is full-batch gradient descent.
    Required for the second-order expansion; real training uses Adam/AdamW with mini-batches.
  • domain assumption Average validation performance (or negative loss) is the correct multi-task utility for Shapley.
    Defines u(S); alternative multi-task metrics could reorder sellers.
invented entities (2)
  • DMVM utility proxy û(S) via task-arithmetic merge independent evidence
    purpose: Replace exponential multi-task retraining with parameter-space addition.
    Core algorithmic object; independent evidence is the empirical correlation with true MTL performance reported by Tao et al. and the paper's own tables.
  • Unmergeable transformation (permutation / scalar ρ on consecutive layers) no independent evidence
    purpose: Prevent algebraic recovery of individual models from overlapping merged subsets.
    New privacy gadget; correctness argued by preservation of inference under element-wise nonlinearity, but no formal cryptographic reduction is given.

pith-pipeline@v1.1.0-grok45 · 32248 in / 2933 out tokens · 27678 ms · 2026-07-12T03:07:14.289598+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Efficient Decentralized Multi-task Dataset Valuation via Model Merging." pith.science (2026). https://pith.science/paper/2BMJPIN2

@misc{pith2026260703346,
  author       = {Pith},
  title        = {Pith review of: Efficient Decentralized Multi-task Dataset Valuation via Model Merging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2BMJPIN2}},
  note         = {Machine review of arXiv:2607.03346}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate and efficient dataset valuation is essential for enabling fair and transparent data marketplaces, especially when multiple contributors provide data for training multi-task models. Most existing valuation methods, however, are limited to single-task settings, overlooking scenarios where a buyer aims to optimize performance across multiple downstream tasks. Moreover, traditional valuation approaches, such as Shapley-based or retraining-based methods, are computationally expensive and poorly suited for decentralized environments without a trusted central coordinator and with strict privacy constraints. We propose DMVM (Decentralized Multi-task Valuation via Model Merging), a novel framework that bypasses retraining and data sharing by leveraging task arithmetic to infer dataset contributions directly from model combinations. Instead of retraining or sharing raw data, DMVM quantifies how models trained on different datasets combine in parameter space to infer each dataset's marginal utility across multiple tasks. This formulation yields a valuation process that is scalable, computationally efficient, and explicitly aligned with multi-task generalization behavior. To support decentralized deployment, we introduce a secure aggregation protocol that enables collaborative valuation without revealing individual model parameters or private data. We also provide theoretical error bounds characterizing the approximation quality of DMVM and validate our framework through comprehensive experiments on computer vision and natural language processing tasks.

Figures

Figures reproduced from arXiv: 2607.03346 by Mohammad Mohammadi Amiri, Mohammadsajad Alipour.

Figure 1
Figure 1. Figure 1: Overview of the DMVM framework for decentralized multi-task dataset valuation. Each [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Efficiency-accuracy trade-off of valuation methods on NLP and vision tasks. The [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Convergence behavior of MC-based valuation methods on vision tasks with [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Convergence behavior of MC-based valuation methods on NLP tasks with [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: DMVM-induced shapley values of sellers with different percentages of mislabeled data. As [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Projected wall-clock cost of each valuation method as the marketplace size [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: DMVM performance on vision tasks across the full [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: DMVM performance on NLP tasks across the full [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

78 extracted references · 2 canonical work pages

  1. [1]

    A marketplace for data: An algorithmic solution

    Anish Agarwal, Munther Dahleh, and Tuhin Sarkar. A marketplace for data: An algorithmic solution. InProceedings of the 2019 ACM Conference on Economics and Computation, EC ’19, page 701–726, New York, NY , USA, 2019. Association for Computing Machinery. ISBN 9781450367929. doi: 10.1145/3328526.3329589. URL https://doi.org/10.1145/ 3328526.3329589

  2. [2]

    Mix data or merge models? optimizing for diverse multi-task learning.arXiv preprint arXiv:2410.10801, 2024

    Arash Ahmadian, Seraphina Goldfarb-Tarrant, Beyza Ermis, Marzieh Fadaee, Sara Hooker, et al. Mix data or merge models? optimizing for diverse multi-task learning.arXiv preprint arXiv:2410.10801, 2024

  3. [3]

    Towards reversible model merging for low-rank weights.arXiv preprint arXiv:2510.14163, 2025

    Mohammadsajad Alipour and Mohammad Mohammadi Amiri. Towards reversible model merging for low-rank weights.arXiv preprint arXiv:2510.14163, 2025

  4. [4]

    Fundamentals of task- agnostic data valuation

    Mohammad Mohammadi Amiri, Frederic Berdoz, and Ramesh Raskar. Fundamentals of task- agnostic data valuation. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 9226–9234, 2023

  5. [5]

    Food-101–mining discriminative components with random forests

    Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components with random forests. InEuropean conference on computer vision, pages 446–461. Springer, 2014

  6. [6]

    Chg shapley: Efficient data valuation and selection towards trustworthy machine learning.arXiv preprint arXiv:2406.11730, 2024

    Huaiguang Cai. Chg shapley: Efficient data valuation and selection towards trustworthy machine learning.arXiv preprint arXiv:2406.11730, 2024

  7. [7]

    Towards model-based pricing for machine learning in a data marketplace

    Lingjiao Chen, Paraschos Koutris, and Arun Kumar. Towards model-based pricing for machine learning in a data marketplace. InProceedings of the 2019 International Conference on Management of Data, SIGMOD ’19, page 1535–1552, New York, NY , USA, 2019. Association for Computing Machinery. ISBN 9781450356435. doi: 10.1145/3299869.3300078. URL https://doi.org/...

  8. [8]

    Defending unauthorized model merging via dual-stage weight protection.arXiv preprint arXiv:2511.11851, 2025

    Wei-Jia Chen, Min-Yen Tsai, Cheng-Yi Lee, and Chia-Mu Yu. Defending unauthorized model merging via dual-stage weight protection.arXiv preprint arXiv:2511.11851, 2025

  9. [9]

    Remote sensing image scene classification: Benchmark and state of the art.Proceedings of the IEEE, 105(10):1865–1883, 2017

    Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote sensing image scene classification: Benchmark and state of the art.Proceedings of the IEEE, 105(10):1865–1883, 2017

  10. [10]

    A joint energy and differentially-private smart meter data market.arXiv preprint arXiv:2412.07688, 2024

    Saurab Chhachhi and Fei Teng. A joint energy and differentially-private smart meter data market.arXiv preprint arXiv:2412.07688, 2024

  11. [11]

    Aggarwal, and Yao Ma

    Hongliang Chi, Wei Jin, Charu C. Aggarwal, and Yao Ma. Precedence-constrained winter value for effective graph data valuation. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=tVRVE0OAyb

  12. [12]

    Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean,...

  13. [13]

    Describing textures in the wild

    Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3606–3613, 2014

  14. [14]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. InProceedings of the 2019 conference of the north American chapter of the association for computational linguistics: Human language technologies, volume 1 (long and short papers), ...

  15. [15]

    Think you have solved question answering? try arc, the ai2 reasoning challenge.arXiv:1803.05457v1, 2018

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge.arXiv:1803.05457v1, 2018

  16. [16]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

  17. [17]

    Emnist: Extending mnist to handwritten letters

    Gregory Cohen, Saeed Afshar, Jonathan Tapson, and Andre Van Schaik. Emnist: Extending mnist to handwritten letters. In2017 international joint conference on neural networks (IJCNN), pages 2921–2926. IEEE, 2017

  18. [18]

    Disentangled structural and featural represen- tation for task-agnostic graph valuation.arXiv preprint arXiv:2408.12659, 2024

    Ali Falahati and Mohammad Mohammadi Amiri. Disentangled structural and featural represen- tation for task-agnostic graph valuation.arXiv preprint arXiv:2408.12659, 2024

  19. [19]

    Du-shapley: A shapley value proxy for efficient dataset valuation.Advances in Neural Information Processing Systems, 37:1973–2000, 2024

    Felipe Garrido Lucero, Benjamin Heymann, Maxime V ono, Patrick Loiseau, and Vianney Perchet. Du-shapley: A shapley value proxy for efficient dataset valuation.Advances in Neural Information Processing Systems, 37:1973–2000, 2024

  20. [20]

    Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies

    Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics, 9:346–361, 2021

  21. [21]

    Data shapley: Equitable valuation of data for machine learning

    Amirata Ghorbani and James Zou. Data shapley: Equitable valuation of data for machine learning. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors,Proceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Research, pages 2242–2251. PMLR, 09–15 Jun 2019. URL https://proceedings.mlr. press/v97/gho...

  22. [22]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

  23. [23]

    Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217–2226, 2019

  24. [24]

    Jonathan J. Hull. A database for handwritten text recognition research.IEEE Transactions on pattern analysis and machine intelligence, 16(5):550–554, 2002

  25. [25]

    Editing models with task arithmetic

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Ha- jishirzi, and Ali Farhadi. Editing models with task arithmetic. InThe Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum? id=6t0Kwf8-jrj

  26. [26]

    Private, augmentation-robust and task-agnostic data valuation approach for data marketplace

    Tayyebeh Jahani-Nezhad, Parsa Moradi, Mohammad Ali Maddah-Ali, and Giuseppe Caire. Private, augmentation-robust and task-agnostic data valuation approach for data marketplace. arXiv preprint arXiv:2411.00745, 2024

  27. [27]

    Efficient task-specific data valuation for nearest neighbor algorithms.Proc

    Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nezihe Merve Gurel, Bo Li, Ce Zhang, Costas Spanos, and Dawn Song. Efficient task-specific data valuation for nearest neighbor algorithms.Proc. VLDB Endow., 12(11):1610–1623, July 2019. ISSN 2150-8097. doi: 10.14778/3342263.3342637. URLhttps://doi.org/10.14778/3342263.3342637

  28. [28]

    Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nick Hynes, Nezihe Merve Gürel, Bo Li, Ce Zhang, Dawn Song, and Costas J. Spanos. Towards efficient data valuation based on the shapley value. In Kamalika Chaudhuri and Masashi Sugiyama, editors,Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, volume 8...

  29. [29]

    Pubmedqa: A dataset for biomedical research question answering

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2567–2577, 2019

  30. [30]

    Disrupting model merging: A parameter-level defense without sacrificing accuracy

    Wei Junhao, Yu Zhe, and Jun Sakuma. Disrupting model merging: A parameter-level defense without sacrificing accuracy. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17698–17707, October 2025

  31. [31]

    Understanding black-box predictions via influence functions

    Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. InInternational conference on machine learning, pages 1885–1894. PMLR, 2017

  32. [32]

    3d object representations for fine- grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine- grained categorization. In2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, 2013. doi: 10.1109/ICCVW.2013.77

  33. [33]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

  34. [34]

    Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

  35. [35]

    When is task vector provably effective for model editing? a generalization analysis of nonlinear transformers

    Hongkang Li, Yihua Zhang, Shuai Zhang, Pin-Yu Chen, Sijia Liu, and Meng Wang. When is task vector provably effective for model editing? a generalization analysis of nonlinear transformers. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=vRvVVb0NAz

  36. [36]

    Logiqa: A challenge dataset for machine reading comprehension with logical reasoning.arXiv preprint arXiv:2007.08124, 2020

    Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and Yue Zhang. Logiqa: A challenge dataset for machine reading comprehension with logical reasoning.arXiv preprint arXiv:2007.08124, 2020

  37. [37]

    Dealer: an end-to-end model marketplace with differential privacy.Proc

    Jinfei Liu, Jian Lou, Junxu Liu, Li Xiong, Jian Pei, and Jimeng Sun. Dealer: an end-to-end model marketplace with differential privacy.Proc. VLDB Endow., 14(6):957–969, February

  38. [38]

    doi: 10.14778/3447689.3447700

    ISSN 2150-8097. doi: 10.14778/3447689.3447700. URL https://doi.org/10. 14778/3447689.3447700

  39. [39]

    Data measurements for decentralized data markets.arXiv preprint arXiv:2406.04257, 2024

    Charles Lu, Mohammad Mohammadi Amiri, and Ramesh Raskar. Data measurements for decentralized data markets.arXiv preprint arXiv:2406.04257, 2024

  40. [40]

    Maas, Raymond E

    Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y . Ng, and Christopher Potts. Learning word vectors for sentiment analysis. InProceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142– 150, Portland, Oregon, USA, June 2011. Association for Computational Linguistics. U...

  41. [41]

    Bounding the es- timation error of sampling-based shapley value approximation.arXiv preprint arXiv:1306.4265, 2013

    Sasan Maleki, Long Tran-Thanh, Greg Hines, Talal Rahwan, and Alex Rogers. Bounding the es- timation error of sampling-based shapley value approximation.arXiv preprint arXiv:1306.4265, 2013

  42. [42]

    Merging models with fisher-weighted averaging

    Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. Advances in Neural Information Processing Systems, 35:17703–17716, 2022

  43. [43]

    Can a suit of armor conduct electricity? a new dataset for open book question answering

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. InEMNLP, 2018

  44. [44]

    Scalable data ablation approximations for language models through modular training and merging

    Clara Na, Ian Magnusson, Ananya Harsh Jha, Tom Sherborne, Emma Strubell, Jesse Dodge, and Pradeep Dasigi. Scalable data ablation approximations for language models through modular training and merging. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21125–21141, 2024

  45. [45]

    Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y . Ng. Reading digits in natural images with unsupervised feature learning. InNIPS Workshop on Deep Learning and Unsupervised Feature Learning 2011, 2011. URL http://ufldl.stanford. edu/housenumbers/nips2011_housenumbers.pdf. 13

  46. [46]

    Selectformer: Private and practical data selec- tion for transformers

    Xu Ouyang, Felix Xiaozhu Lin, and Yangfeng Ji. Selectformer: Private and practical data selec- tion for transformers. InThe Thirteenth International Conference on Learning Representations,

  47. [47]

    URLhttps://openreview.net/forum?id=2cF3f9t31y

  48. [48]

    Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. In Gerardo Flores, George H Chen, Tom Pollard, Joyce C Ho, and Tristan Naumann, editors,Proceedings of the Conference on Health, Inference, and Learning, volume 174 ofProceedings of Machine Learni...

  49. [49]

    Parkhi, Andrea Vedaldi, Andrew Zisserman, and C

    Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V . Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012

  50. [50]

    Recognizing indoor scenes

    Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In2009 IEEE conference on computer vision and pattern recognition, pages 413–420. IEEE, 2009

  51. [51]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li,...

  52. [52]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Machin...

  53. [53]

    A value for n-person games

    Lloyd S Shapley et al. A value for n-person games. 1953

  54. [54]

    The german traffic sign recognition benchmark: A multi-class classification competition

    Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: A multi-class classification competition. InThe 2011 International Joint Conference on Neural Networks, pages 1453–1460, 2011. doi: 10.1109/IJCNN.2011.6033395

  55. [55]

    CommonsenseQA: A question answering challenge targeting commonsense knowledge

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. CommonsenseQA: A question answering challenge targeting commonsense knowledge. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4149– 4158, Minneapoli...

  56. [56]

    Fusionbench: A comprehensive benchmark of deep model fusion.Journal of Machine Learning Research, 2025

    Anke Tang, Li Shen, Yong Luo, Enneng Yang, Han Hu, Lefei Zhang, Bo Du, and Dacheng Tao. Fusionbench: A comprehensive benchmark of deep model fusion.Journal of Machine Learning Research, 2025

  57. [57]

    Heterogeneous graph data valuation: A shapley value- based approach

    Xin Tang, Yanzhe Lei, and Chao Li. Heterogeneous graph data valuation: A shapley value- based approach. In2025 40th Youth Academic Annual Conference of Chinese Association of Automation (YAC), pages 2871–2876, 2025. doi: 10.1109/Y AC66630.2025.11150137

  58. [58]

    Merge to mix: Mixing datasets via model merging.arXiv preprint arXiv:2505.16066, 2025

    Zhixu Silvia Tao, Kasper Vinken, Hao-Wei Yeh, Avi Cooper, and Xavier Boix. Merge to mix: Mixing datasets via model merging.arXiv preprint arXiv:2505.16066, 2025

  59. [59]

    Private data valuation and fair payment in data marketplaces.arXiv preprint arXiv:2210.08723, 2022

    Zhihua Tian, Jian Liu, Jingyu Li, Xinle Cao, Ruoxi Jia, Jun Kong, Mengdi Liu, and Kui Ren. Private data valuation and fair payment in data marketplaces.arXiv preprint arXiv:2210.08723, 2022

  60. [60]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 14

  61. [61]

    GLUE: A multi-task benchmark and analysis platform for natural language understanding

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Tal Linzen, Grzegorz Chrupała, and Afra Alishahi, editors,Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353– 355, Bru...

  62. [62]

    Wang, Prateek Mittal, Dawn Song, and Ruoxi Jia

    Jiachen T. Wang, Prateek Mittal, Dawn Song, and Ruoxi Jia. Data shapley in one training run. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=HD6bWcj87Y

  63. [63]

    Model unmerging: Making your models unmergeable for secure model sharing.arXiv preprint arXiv:2509.01548, 2025

    Zihao Wang, Enneng Yang, Lu Yin, Shiwei Liu, and Li Shen. Model unmerging: Making your models unmergeable for secure model sharing.arXiv preprint arXiv:2509.01548, 2025

  64. [64]

    Scaling intelligence through model merging: A comprehensive survey.Authorea Preprints, 2025

    Zijing Wang, Yongkang Liu, Yingfeng Luo, Ming Wang, Zhen Song, Shi Feng, Xiaocui Yang, Dingyang Lin, Daling Wang, Yifei Zhang, et al. Scaling intelligence through model merging: A comprehensive survey.Authorea Preprints, 2025

  65. [65]

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. InInternational conference on machine learning, pages 23965–23998. P...

  66. [66]

    Visual transformers: Token-based image representation and processing for computer vision, 2020

    Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision, 2020

  67. [67]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747, 2017

    Han Xiao, Kashif Rasul, and Roland V ollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.arXiv preprint arXiv:1708.07747, 2017

  68. [68]

    Ehinger, Aude Oliva, and Antonio Torralba

    Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485–3492, 2010. doi: 10. 1109/CVPR.2010.5539970

  69. [69]

    Ties-merging: Resolving interference when merging models.Advances in Neural Information Processing Systems, 36:7093–7115, 2023

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models.Advances in Neural Information Processing Systems, 36:7093–7115, 2023

  70. [70]

    On the inflation of knn-shapley value.arXiv preprint arXiv:2405.17489, 2024

    Ziao Yang, Han Yue, Jian Chen, and Hongfu Liu. On the inflation of knn-shapley value.arXiv preprint arXiv:2405.17489, 2024

  71. [71]

    Data valuation using reinforcement learning

    Jinsung Yoon, Sercan Arik, and Tomas Pfister. Data valuation using reinforcement learning. In International Conference on Machine Learning, pages 10842–10851. PMLR, 2020

  72. [72]

    Language models are super mario: Absorbing abilities from homologous models as a free lunch

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. InForty-first International Conference on Machine Learning, 2024

  73. [73]

    Character-level convolutional networks for text classification.Advances in neural information processing systems, 28, 2015

    Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification.Advances in neural information processing systems, 28, 2015

  74. [74]

    Paws: Paraphrase adversaries from word scrambling

    Yuan Zhang, Jason Baldridge, and Luheng He. Paws: Paraphrase adversaries from word scrambling. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298–1308, 2019

  75. [75]

    Toward efficient influence function: Dropout as a compression tool.arXiv preprint arXiv:2509.15651, 2025

    Yuchen Zhang and Mohammad Mohammadi Amiri. Toward efficient influence function: Dropout as a compression tool.arXiv preprint arXiv:2509.15651, 2025

  76. [76]

    On task vectors and gradients.arXiv preprint arXiv:2508.16082, 2025

    Luca Zhou, Daniele Solombrino, Donato Crisostomi, Maria Sofia Bucarelli, Giuseppe Alessio D’Inverno, Fabrizio Silvestri, and Emanuele Rodolà. On task vectors and gradients.arXiv preprint arXiv:2508.16082, 2025. 15

  77. [77]

    Multi-task code llms: Data mix or model merge?arXiv preprint arXiv:2601.21115, 2026

    Mingzhi Zhu, Boris Sobolev, Rahul Krishna, Raju Pavuluri, Stacy Patterson, and Michele Merler. Multi-task code llms: Data mix or model merge?arXiv preprint arXiv:2601.21115, 2026. A Appendix A.1 Algorithm Algorithm 1DMVM: Decentralized Multi-task Valuation via Secure Model Merging 1: Input:Buyer queries a potential group of sellers [n], base model weights...

  78. [78]

    Moreover, the Shapley coefficients form a convex combination, i.e., 1 n P S⊆[n]\{i} 1 ( n−1 |S| ) = 1

    + 1)C, respectively, using inequality (4). Moreover, the Shapley coefficients form a convex combination, i.e., 1 n P S⊆[n]\{i} 1 ( n−1 |S| ) = 1. Therefore, applying Lemma A.1 with b= 1 yields inequality (a). 18 A.5 Error bound of our privacy-aware method Assuming that we want to exclude subsets of size less than k. Let R be the maximum marginal utility, ...