Imagine an AI that not only solves complex problems but also learns from its own attempts, refining its reasoning process like a human student. Researchers have long sought ways to make Large Language Models (LLMs) more independent learners, but traditional self-training methods often fall short. These methods typically generate a bunch of answers and simply select the correct ones for further training, overlooking the often flawed reasoning steps that led to the right answer. Enter ReST-MCTS*, a revolutionary self-training approach that not only finds the right solutions but also understands *how* to get there. It uses a tree search method, guided by a reward system, to explore different reasoning paths, much like mapping out all the routes on a map before choosing the best one. The key innovation? ReST-MCTS* automatically figures out the best reasoning steps without needing constant human supervision. By simulating numerous problem-solving attempts, it learns to identify helpful steps that lead to correct solutions, effectively teaching itself to reason more effectively. This breakthrough has significant implications for various domains, from solving complex math problems to cracking scientific mysteries. By training LLMs to learn independently, ReST-MCTS* paves the way for a new era of AI reasoning, where machines can continuously improve their problem-solving abilities and unlock new levels of intelligence. While the research primarily focuses on math problems, future work will explore its applicability to broader tasks like code generation and conversational AI. The researchers are also looking at scaling up the reward model and further refining the self-training process, promising even more impressive results in the future.