Diverge-to-Induce Prompting Enhances Zero-Shot Reasoning in LLMs

A new framework called Diverge-to-Induce Prompting (DIP) improves zero-shot reasoning in large language models by generating multiple diverse rationales before synthesizing a final plan. Unlike standard Chain-of-Thought prompting, which relies on a single reasoning path, DIP first prompts the LLM to produce several high-level rationales, elaborates each into a detailed draft plan, and then induces these into a final plan. Experiments show DIP outperforms single-strategy methods without resource-intensive sampling. The work addresses instability in unguided reasoning paths and demonstrates the effectiveness of multi-plan induction.