Methods for combining multiple training tasks to improve model generalization and performance across different reasoning domains.