Training a model on several related tasks at once to share representations and improve generalization, balancing losses with schedulers or task weighting.