A Spectral Framework for Universal Proxy Regularization in Model Merging
November 16, 2026
Spectral Framework for Universal Proxy Regularization quantifies parameter interference in model merging using spectral coverage and leakage metrics. This framework provides a theoretical bound for using task-agnostic proxy distributions to mitigate interference without requiring task-specific data. The method enables stable multi-task model construction by ensuring spectral alignment between the proxy and target task distributions.