Learning to Reason with Mixture of Tokens

Adit Jain, Brendan Rappazzo

Reinforcement learning with verifiable rewards (RLVR) has become a leading approach for improving large language model (LLM) reasoning capabilities. Most current methods follow var…