SQuARE
Research paper: SQuARE uses self-interrogation to decompose questions, outperforming chain-of-thought prompts.
SQuARE introduces a self-interrogation paradigm for prompting LLMs. Instead of directly answering a question, the model generates and resolves a sequence of auxiliary questions before tackling the main query, encouraging a more comprehensive decomposition and exploration of the topic. This approach builds on chain-of-thought prompting and aims to better leverage the model’s internal reasoning. We evaluated SQuARE with Llama 3 and GPT-4o across multiple QA datasets and reported that it significantly surpasses traditional chain-of-thought prompts and existing rephrase-and-respond methods. We provide code at the Intel Labs GitHub repository.
