Imagine navigating a maze, not just by randomly trying paths, but by having a smart guide whispering the best direction at each turn. That's the power of heuristic search algorithms like A*, combined with Large Language Models (LLMs). A* acts as the explorer, while the LLM provides an intuitive sense of direction, a 'heuristic,' about how close any point in the maze is to the exit. This collaboration allows AI to tackle complex problems much faster than traditional brute-force methods. But there's a catch: training these LLMs to be effective guides can be incredibly time-consuming and computationally expensive. New research delves into precisely this challenge, exploring how to create a 'training data recipe' that accelerates the learning process. The core idea is to find the most valuable training examples, a 'coreset,' that teach the LLM the most efficient navigation strategy without needing an enormous amount of data. Surprisingly, the key ingredient in this recipe is focusing on training data near the 'goal' of the problem. This makes intuitive sense for the LLM—it's easier to predict the distance to the exit when you're standing nearby. But more unexpectedly, the research finds that even for A*'s search efficiency, having an accurate heuristic near the goal is more valuable than accurate guesses earlier on. Why? Speculatively, the researchers suggest that wrong turns early on can be corrected if the AI has a strong sense of direction as it gets closer to the destination. This focus on “end-game” data has led to a planner-aware training method that dramatically improves the speed of AI problem-solving in tasks like maze navigation, Sokoban puzzles, and even the classic sliding tile puzzle. Results show up to a 15x reduction in the time it takes A* to find a solution! While tested on classic puzzles, the researchers believe these insights into data optimization and training can be applied to more complex AI planning problems, leading to faster, more efficient solutions in a variety of real-world applications. The future of smarter, faster AI planning lies not just in bigger models, but in finding the right ingredients to train them.