Hypothesis-Conditioned Query Rewriting for Decision-Useful Retrieval

This paper introduces Hypothesis-Conditioned Query Rewriting (HCQR), a method that improves how AI systems retrieve relevant information when they need to choose between competing answers. Instead of using a single search query that finds generally relevant documents, HCQR creates three targeted searches designed to support a leading hypothesis, distinguish it from alternatives, and verify key details—making the retrieved information more directly useful for making the final decision. Tests on medical question datasets show this approach significantly outperforms traditional retrieval methods.