algorithm-learning

DART proposes a distributional adversarial training framework for recurrent reasoning models, improving their robustness and performance on structured problems by using a local target distribution instead of single-point supervision.