CoT-Tree: An Efficient Prompting Strategy for Large Language Model Classification Reasoning

Prompt engineering in Large Language Models (LLMs) has become pivotal for task adaptation, especially in complex reasoning tasks where traditional methods fall short. This paper introduces an innovative method to optimize prompt efficiency by constructing a decision tree of the Chain of Thought (CoT) for classification tasks, namely CoT-Tree. Initially, we demonstrate two advantages of using CoT as features for decision-tree prompts: task adaptability and semantic dissimilarity between features. These advantages theoretically ensure that the CoT-Tree prompting strategy achieves higher performance in classification tasks. Subsequently, we propose the CoT-Tree algorithm and apply it on three classic classification scenarios. By leveraging the CoT-Tree prompting strategy, we select distinctive and efficient prompts that maximize task relevance while minimizing redundancy and token consumption, allowing for a dynamic pruning mechanism, effectively controlling the depth and complexity of the prompts. The results indicate that our method achieves higher accuracy with fewer LLM interaction rounds. Additionally, a case study is conducted to further demonstrate the practicality of our method.