CoTBal algorithm enhances multi-task visual instruction tuning by balancing task contributions and difficulties.
Aligning learning progress across tasks through adaptive group-based task interactions to improve overall multi-task performance while maintaining computational efficiency.
This paper introduces PUB, a novel multi-task optimization algorithm that addresses the "seesaw problem" in multi-task learning by balancing parameter updates instead of gradients, leading to improved performance and robustness across various applications, including recommendation systems and computer vision.