Large Language Models (LLMs) have made impressive strides, but their reasoning abilities still lag behind humans. Imagine an AI that could not only solve problems but also identify and correct its own mistakes along the way. That's the exciting potential of new research exploring "intrinsic self-correction" in LLMs. Researchers are exploring a two-stage process that combines the power of Monte Carlo Tree Search (MCTS) with iterative preference learning. In the first stage, the LLM learns to refine its predictions using only its own self-generated data, effectively bootstrapping its self-correction capabilities. This internally corrected LLM then feeds into the second stage, which leverages step-wise preference learning, similar to how AlphaZero masters complex games. This method enhances the LLM's ability to verify its reasoning at each step, leading to more accurate and robust problem-solving. Experiments on challenging math word problems show promising results, with this new approach outperforming existing LLMs by a significant margin. This innovative combination of MCTS and self-correction opens up exciting possibilities for building more reliable and robust AI systems that can reason more effectively, potentially leading to breakthroughs in areas like automated theorem proving, complex problem-solving, and even creative writing. While the research is still in its early stages, the potential for self-correcting AI is vast, promising a future where AI can learn, reason, and improve itself with minimal human intervention.