Not All Task Vectors Need Equal Rank: Energy-Proportional Allocation for Model Merging
SERA gives harder-to-compress task updates more rank while keeping the total merge budget unchanged.
The paper argues that uniform rank allocation wastes capacity because task vectors have different spectral complexity. Its method, Spectral Energy-proportional Rank Allocation, assigns ranks from each task vector’s singular-value energy structure. In standard vision model-merging experiments, the authors report better multi-task performance than existing spectral merging methods at the same total rank budget. They also tie the gains to how concentrated each task’s spectrum is. ArXiv · AI/CL/LG's note
The paper argues that uniform rank allocation wastes capacity because task vectors have different spectral complexity. Its method, Spectral Energy-proportional Rank Allocation, assigns ranks from each task vector’s singular-value energy structure. In standard vision model-merging experiments, the authors report better multi-task performance than existing spectral merging methods at the same total rank budget. They also tie the gains to how concentrated each task’s spectrum is. ArXiv · AI/CL/LG's note
score 4