Hypothesis Generation with Large Language Models

This paper describes HypoGeniC, a new method that uses large language models (LLMs) to generate and refine data-driven hypotheses. The process begins by creating initial hypotheses from a small sample, then iteratively updating them based on a reward function inspired by multi-armed bandits, which helps balance exploring new ideas and leveraging effective ones. The generated hypotheses can then be used to create interpretable classifiers that often outperform traditional supervised learning and few-shot methods, and these hypotheses are shown to generalize well across different LLMs and datasets, even uncovering novel insights in real-world tasks. The authors highlight that the success lies in the LLM's ability to generate and combine concepts that are then validated through the exploration-exploitation process.