Imagine teaching a robot to navigate a complex maze. Instead of overwhelming it with the entire challenge at once, what if you could break it down into smaller, manageable steps? That’s the core idea behind new research on using Large Language Models (LLMs) to help AI agents conquer intricate tasks by mastering subtasks. Traditionally, training AI for complex goals has been like trying to solve a giant puzzle without knowing where the pieces fit. This new approach, however, uses a clever method to identify key 'landmarks' or subtasks within a larger problem. Think of it like strategically placing flags along the maze path. These landmarks become the milestones for the AI to achieve, step-by-step. Researchers have developed an algorithm that analyzes successful and unsuccessful attempts at a task, identifying the common states achieved in the successful attempts—these become the landmarks or subtasks. Then, LLMs come into play. The LLM, armed with its understanding of common-sense reasoning and the structure of the task, generates rule templates for achieving each subtask. These rules guide the AI agent, telling it what actions to take when it encounters specific situations. It’s like giving the robot a mini-instruction manual for each part of the maze. This research shows promising results in a simulated 'GetOut' environment where the AI has to collect specific objects in a particular order. The results demonstrate that the AI can effectively learn complex tasks by focusing on these smaller milestones. This approach is significant because it can potentially lead to more efficient and less data-intensive AI training, enabling machines to solve more complex problems in the real world. This work presents a fascinating fusion of symbolic reasoning and LLM capabilities. By integrating these approaches, the research opens doors for tackling intricate real-world problems in robotics, planning, and other domains where step-by-step guidance is crucial.