Hypothesis-Conditioned Query Rewriting for Decision-Useful Retrieval
arXiv:2603.19008v1 Announce Type: new Abstract: Retrieval-Augmented Generation (RAG) improves Large Language Models (LLMs) by grounding generation in external, non-parametric knowledge. However, when a task requires choosing among competing options, simply grounding generation in broadly relevant context is often insufficient to drive the final decision. Existing RAG methods typically rely on a single initial query, which often favors topical relevance over decision-relevant evidence, and therefore retrieves background information that can fail to discriminate among answer options. To address this issue, here we propose Hypothesis-Conditioned Query Rewriting (HCQR), a training-free pre-retrieval framework that reorients RAG from topic-oriented retrieval to evidence-oriented retrieval. HCQR first derives a lightweight working hypothesis from the input question and candidate options, and then rewrites retrieval into three targeted queries that seek evidence to: (1) support the hypothesi
arXiv:2603.19008v1 Announce Type: new Abstract: Retrieval-Augmented Generation (RAG) improves Large Language Models (LLMs) by grounding generation in external, non-parametric knowledge. However, when a task requires choosing among competing options, simply grounding generation in broadly relevant context is often insufficient to drive the final decision. Existing RAG methods typically rely on a single initial query, which often favors topical relevance over decision-relevant evidence, and therefore retrieves background information that can fail to discriminate among answer options. To address this issue, here we propose Hypothesis-Conditioned Query Rewriting (HCQR), a training-free pre-retrieval framework that reorients RAG from topic-oriented retrieval to evidence-oriented retrieval. HCQR first derives a lightweight working hypothesis from the input question and candidate options, and then rewrites retrieval into three targeted queries that seek evidence to: (1) support the hypothesis, (2) distinguish it from competing alternatives, and (3) verify salient clues in the question. This approach enables context retrieval that is more directly aligned with answer selection, allowing the generator to confirm or overturn the initial hypothesis based on the retrieved evidence. Experiments on MedQA and MMLU-Med show that HCQR consistently outperforms single-query RAG and re-rank/filter baselines, improving average accuracy over Simple RAG by 5.9 and 3.6 points, respectively. Code is available at https://anonymous.4open.science/r/HCQR-1C2E.
Executive Summary
Hypothesis-Conditioned Query Rewriting (HCQR) is a training-free pre-retrieval framework proposed to improve Retrieval-Augmented Generation (RAG) methods. HCQR reorients RAG from topic-oriented retrieval to evidence-oriented retrieval by deriving a lightweight working hypothesis and rewriting retrieval into three targeted queries. Experiments show HCQR consistently outperforms single-query RAG and re-rank/filter baselines, with average accuracy improvements of 5.9 and 3.6 points over Simple RAG on MedQA and MMLU-Med datasets, respectively. HCQR's ability to retrieve decision-relevant evidence aligns with answer selection, allowing generators to confirm or overturn initial hypotheses. This framework has significant implications for applications requiring nuanced decision-making, such as medical diagnosis and question-answering systems.
Key Points
- ▸ HCQR is a training-free pre-retrieval framework that enhances RAG methods.
- ▸ HCQR reorients RAG from topic-oriented retrieval to evidence-oriented retrieval.
- ▸ HCQR derives a lightweight working hypothesis and rewrites retrieval into three targeted queries.
Merits
Improved Retrieval Accuracy
HCQR's evidence-oriented retrieval approach enables generators to more accurately confirm or overturn initial hypotheses, leading to improved average accuracy over Simple RAG.
Demerits
Limited Generalizability
HCQR's performance on MedQA and MMLU-Med datasets may not generalize to other domains or tasks, limiting its practical applicability.
Expert Commentary
HCQR's innovative approach to RAG methods demonstrates a sophisticated understanding of the complexities involved in decision-making tasks. By reorienting retrieval towards evidence-oriented evidence, HCQR addresses a critical limitation of existing RAG methods. However, further research is needed to explore HCQR's limitations, particularly its generalizability to diverse domains and tasks. Moreover, HCQR's potential vulnerability to adversarial attacks highlights the need for robustness and security evaluations in AI-powered decision-making systems.
Recommendations
- ✓ Develop HCQR-based systems for high-stakes applications, such as medical diagnosis and question-answering systems.
- ✓ Investigate HCQR's limitations and potential vulnerabilities to ensure its safe and effective deployment.
