ReasonRank: Boosting Passage Ranking with Reasoning Models

A new paper introduces ReasonRank, a listwise passage reranker that leverages large reasoning models (LRMs) to improve ranking performance on complex queries. The authors propose an automated framework to generate reasoning-intensive training data by sourcing queries and passages from diverse domains and using DeepSeek-R1 to produce high-quality labels. A two-stage training approach is employed: cold-start supervised fine-tuning followed by reinforcement learning. The method aims to address the scarcity of reasoning-intensive training data and enhance the ranking ability of rerankers in complex scenarios.