This seems like it's just feeding the output back into the model and using more compute to try and get better answers. If that's all, I don't see how it fundamentally solves any of the issues currently present in LLMs. Maybe a marginal improvement in accuracy at the cost of making the computation more expensive. And you don't even get to see the so called reasoning tokens.