Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners
Thinking-Free Policy Initialization (TFPI) connects long Chain-of-Thought distillation and standard RLVR to improve reasoning models. It introduces a 'ThinkingFree' operation that removes thinking content during inference, significantly cutting token usage. Training with this adapted input improves performance and lowers costs, accelerating RL convergence and raising the performance ceiling without complex rewards.