A new training method called Feedback Distillation has been proposed to enhance reasoning models by improving token-level supervision and incorporating external knowledge. This approach shows better performance in theorem-proving tasks compared to traditional methods, suggesting potential advancements in AI reasoning capabilities.