Tag
Introduces Vector Policy Optimization (VPO) to train models with vector-valued rewards instead of scalar rewards, enabling diverse answer sets for test-time search.