Hit-RAG: Learning to Reason with Long Contexts via Preference Alignment

ABSTRACT

Hit-RAG addresses a key challenge in long-context retrieval systems: attention dilution caused by large volumes of retrieved evidence. The framework introduces a multi-stage preference alignment pipeline that teaches models to prioritize the most relevant information. Training progresses through supervised context learning, discriminative preference alignment, and reinforcement-style policy optimization. This layered alignment strategy helps models resist distractors and focus on critical evidence. Benchmarks demonstrate improved reasoning accuracy across long-context QA tasks.