Large language models (LLMs) have shown impressive abilities, but complex reasoning, especially in math, remains a challenge. Current methods often rely on providing specific examples (like showing a student how to solve a particular math problem). However, this approach struggles with generalization – what happens when the LLM encounters a new problem type it hasn't seen before? New research introduces HiAR-ICL, a system that moves beyond simply providing examples. Instead, it teaches LLMs higher-level “thinking patterns” using a technique called Monte Carlo Tree Search (MCTS). Imagine teaching a student general problem-solving strategies instead of just memorizing formulas. HiAR-ICL identifies fundamental reasoning actions (like analyzing the problem, breaking it down into sub-problems, and refining solutions). It then uses MCTS to explore different combinations of these actions, creating “thought cards” that represent effective reasoning strategies. When faced with a new problem, HiAR-ICL selects the most relevant thought cards based on the problem's complexity and guides the LLM through the solution process. The results are impressive: HiAR-ICL significantly boosts the performance of LLMs, especially smaller models, on complex reasoning benchmarks. It even helps some smaller models surpass the performance of much larger, closed-source models like GPT-4 on challenging math problems. This approach not only improves accuracy but also reduces the computational cost associated with traditional tree search methods, creating a more efficient way for LLMs to reason. This shift from memorization to strategic thinking is a major step forward in developing truly intelligent AI systems. Future research will explore even more sophisticated reasoning paradigms and extend these techniques to other complex domains.