When Shared Knowledge Hurts: Spectral Over-Accumulation in Model Merging

Yayuan Li, Ze Peng, Jian Zhang, Jintao Guo, Yue Duan, Yinghuan Shi
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:69355-69375, 2026.

Abstract

Model merging combines multiple fine-tuned models into a single model by $\textit{adding}$ their weight updates, providing a lightweight alternative to retraining. Existing methods primarily target resolving conflicts between task updates, leaving the failure mode of over-counting shared knowledge unaddressed. We show that when tasks share aligned spectral directions (i.e., overlapping singular vectors), a simple linear combination repeatedly accumulates these directions, inflating the singular values and biasing the merged model toward shared subspaces. To mitigate this issue, we propose Singular Value Calibration (SVC), a training-free and data-free post-processing method that quantifies subspace overlap and rescales inflated singular values to restore a balanced spectrum. Across vision and language benchmarks, SVC consistently improves strong merging baselines and achieves state-of-the-art performance. Furthermore, by modifying only the singular values, SVC improves the performance of Task Arithmetic by 13.0%. Code is available at https://github.com/lyymuwu/SVC.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-li26ep, title = {When Shared Knowledge Hurts: Spectral Over-Accumulation in Model Merging}, author = {Li, Yayuan and Peng, Ze and Zhang, Jian and Guo, Jintao and Duan, Yue and Shi, Yinghuan}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {69355--69375}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/li26ep/li26ep.pdf}, url = {https://proceedings.mlr.press/v306/li26ep.html}, abstract = {Model merging combines multiple fine-tuned models into a single model by $\textit{adding}$ their weight updates, providing a lightweight alternative to retraining. Existing methods primarily target resolving conflicts between task updates, leaving the failure mode of over-counting shared knowledge unaddressed. We show that when tasks share aligned spectral directions (i.e., overlapping singular vectors), a simple linear combination repeatedly accumulates these directions, inflating the singular values and biasing the merged model toward shared subspaces. To mitigate this issue, we propose Singular Value Calibration (SVC), a training-free and data-free post-processing method that quantifies subspace overlap and rescales inflated singular values to restore a balanced spectrum. Across vision and language benchmarks, SVC consistently improves strong merging baselines and achieves state-of-the-art performance. Furthermore, by modifying only the singular values, SVC improves the performance of Task Arithmetic by 13.0%. Code is available at https://github.com/lyymuwu/SVC.} }
Endnote
%0 Conference Paper %T When Shared Knowledge Hurts: Spectral Over-Accumulation in Model Merging %A Yayuan Li %A Ze Peng %A Jian Zhang %A Jintao Guo %A Yue Duan %A Yinghuan Shi %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-li26ep %I PMLR %P 69355--69375 %U https://proceedings.mlr.press/v306/li26ep.html %V 306 %X Model merging combines multiple fine-tuned models into a single model by $\textit{adding}$ their weight updates, providing a lightweight alternative to retraining. Existing methods primarily target resolving conflicts between task updates, leaving the failure mode of over-counting shared knowledge unaddressed. We show that when tasks share aligned spectral directions (i.e., overlapping singular vectors), a simple linear combination repeatedly accumulates these directions, inflating the singular values and biasing the merged model toward shared subspaces. To mitigate this issue, we propose Singular Value Calibration (SVC), a training-free and data-free post-processing method that quantifies subspace overlap and rescales inflated singular values to restore a balanced spectrum. Across vision and language benchmarks, SVC consistently improves strong merging baselines and achieves state-of-the-art performance. Furthermore, by modifying only the singular values, SVC improves the performance of Task Arithmetic by 13.0%. Code is available at https://github.com/lyymuwu/SVC.
APA
Li, Y., Peng, Z., Zhang, J., Guo, J., Duan, Y. & Shi, Y.. (2026). When Shared Knowledge Hurts: Spectral Over-Accumulation in Model Merging. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:69355-69375 Available from https://proceedings.mlr.press/v306/li26ep.html.

Related Material