Large language models (LLMs) have shown remarkable progress in various tasks, but complex reasoning remains a challenge. While prompting LLMs with a chain of thought (CoT) has improved their reasoning abilities, it often relies on manual effort or struggles to guide the LLM effectively. Researchers are constantly exploring new ways to overcome these limitations, and a novel approach called LBS3, inspired by curriculum learning, offers a promising solution. Think about how humans learn: we start with simple concepts and gradually build towards more complex ones. LBS3 applies this principle to LLMs. Instead of throwing a difficult reasoning problem at an LLM directly, LBS3 first prompts the model to generate easier, related "proxy" problems. Then, the LLM solves these simpler problems, generating a chain of thought for each. These solutions become a kind of training ground for the original, harder problem. The next step involves the LLM generating a set of harder proxy problems, analogous to the original query. Crucially, LBS3 uses the solutions from the *easier* proxy problems as examples to help solve these more complex ones. This progressive approach, moving from simple to complex, allows the model to learn more effectively and avoid the problem of accumulating errors that can occur when tackling a difficult problem head-on. Finally, LBS3 combines all the solutions and chains of thought gathered from both the easy and hard proxy problems and uses them as a comprehensive guide to solve the original query. This "learning by example" method significantly boosts the LLM's reasoning accuracy, outperforming other self-prompting methods in various tests. The research shows that LBS3 consistently outperforms other CoT prompting approaches, especially with more powerful LLMs like Llama 3 and GPT-4. This highlights the importance of carefully designed prompts and a structured learning process for improving LLM reasoning. While challenges remain, particularly in the computational cost of generating and solving multiple proxy problems, LBS3 points toward exciting future directions for LLM development. Could this approach pave the way for LLMs that can truly think and reason like humans? It is still a question for future research.