multi-task learning commonly encounters competition for resources among
tasks, specifically when model capacity is limited. This challenge motivates
models which allow control over the relative importance of tasks and total
compute cost during inference time. In this work, we propose s