Transporting Task Vectors across Different Architectures without Training

Researchers have introduced Theseus, a novel training-free method for transporting task-specific parameter updates across different AI model architectures, in a paper published on arXiv on February 26, 2026, addressing the challenge of expensive relearning processes when adapting large pre-trained models to new tasks. The Theseus methodology represents a significant advancement in the field of machine learning, particularly for organizations and developers working with large pre-trained models. Traditionally, when adapting these models to specific downstream tasks, researchers would need to generate task-specific parameter updates for each model variant, a computationally expensive process. Theseus overcomes this limitation by enabling the transfer of these updates across models with different widths and architectures without requiring additional training, potentially saving substantial computational resources and time. The research paper highlights that while previous work had demonstrated the ability to transfer updates between models with identical architectures, transferring them across models of different widths remained largely unexplored territory. Theseus fills this gap by providing a solution that maintains task performance while eliminating the need for retraining, potentially revolutionizing how AI systems are customized and deployed across various industries.