Robust Zero-Shot Generalization for Open-Vocabulary Action Recognition via Task Arithmetic

The paper presents a novel approach to Open Vocabulary Action Recognition (OVAR) that enhances zero-shot generalization without requiring domain-specific fine-tuning. By employing model merging and task arithmetic, the authors extract and recombine task vectors from multiple pre-trained models, resulting in a merged model that outperforms the base model in out-of-distribution scenarios. This advancement is significant for practitioners as it reduces the need for costly fine-tuning processes while maintaining robust performance across diverse action recognition tasks.