RADAR: Retrieval-Aligned Attribution and Dual-Perspective Reasoning for Conversational Query Rewriting

Conversational Query Rewriting (CQR) plays a vital role in modern information retrieval by transforming ambiguous user queries into self-contained versions that can be effectively processed by standard retrieval systems. Although essential for maintaining conversational context, current approaches tend to depend heavily on explicit dialogue history, leading to poor performance in situations where contextual information is limited or when topics shift abruptly. Consequently, the limited ability to look beyond explicit history hinders the generation of diverse and effective queries when historical cues are insufficient or misleading. To address these issues, we propose a design motivation grounded in dual-perspective reasoning. We posit that effective rewriting requires not only resolving historical dependencies but also proactively capturing the user's underlying search intent. Guided by these dual perspectives, we introduce RADAR, a novel feedback-driven framework that leverages the reasoning capabilities of Large Reasoning Models (LRMs) to generate and refine queries. Specifically, we design a two-stage framework. First, we construct an initial corpus of dual-perspective Chain-of-Thought (CoT) reasoning via In-Context Learning (ICL). However, generic reasoning paths can be noisy and retriever-agnostic. To bridge this gap, our second stage introduces the Attribution-based Iterative Reasoning-Rewrite Refinement (AIR3) algorithm. AIR3 leverages passage feedback to identify and quantify the fine-grained contribution of each token and sentence to the retrieval success. This attribution drives an iterative refinement process of pruning redundant reasoning and regenerating rewrites, ensuring the final query is aligned with retriever preferences. We evaluate RADAR on QReCC and TopiOCQA using both sparse and dense retrievers. Results show that it consistently improves various LRM backbones and achieves competitive performance across both retrieval settings, demonstrating the effectiveness of integrating intrinsic reasoning with extrinsic retrieval feedback.