The thing is that a LLMs can point out a logic error in their reasoning if specifically asked to do so. So maybe OpenAI just slapped an RL agent on top of the

The thing is that a LLMs can point out a logic error in their reasoning if specifically asked to do so. So maybe OpenAI just slapped an RL agent on top of the next-token generator.