[whatever1]

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.