Model soups need only one ingredient
Researchers have developed MonoSoup, a novel post-hoc method that enhances the balance between in-distribution accuracy and out-of-distribution robustness in large pre-trained models using only a single checkpoint. This approach leverages Singular Value Decomposition (SVD) to analyze layer updates, separating task-specific adaptations from residual signals beneficial for robustness. Experiments on models like CLIP and Qwen demonstrate that MonoSoup offers a practical and computationally efficient alternative to traditional weight-space ensembling methods, which typically require numerous checkpoints. AI
IMPACT Offers a computationally efficient method to improve model robustness without requiring multiple checkpoints.