Tag
FeatCal is a calibration method that reduces performance gaps in post-merging models by layer-wise weight updates without gradient descent, achieving superior results on CLIP and GLUE benchmarks with high sample efficiency.