<ul><li>AdaRank is a model merging framework that adaptively selects beneficial singular directions of task vectors.</li><li>The reliance on manual rank selection in existing SVD-based techniques leads to cross-task interference and suboptimal performance.</li><li>AdaRank dynamically prunes singular components causing interference, achieving optimal information allocation to each task vector.</li><li>Empirical results demonstrate that AdaRank consistently outperforms existing methods, reducing the performance gap between fine-tuned models.</li></ul>

AdaRank: Adaptive Rank Pruning for Enhanced Model Merging

Discover more