Models like OpenAI's o1 and DeepSeek-R1 have demonstrated powerful reasoning abilities, including planning, reflection, and self-correction, through verifiable rewards (such as the accuracy of solving ...