MACRDR: Enhancing Interest Discovery and Diversity in News Recommendation via Multi-Agent Reflection

<title>Abstract</title> <p>Personalized news recommendation systems often suffer from filter bubbles due to excessive dependence on historical user preferences, making them insensitive to emerging interests and dynamic intent shifts. Existing diversity-aware recommendation methods mainly improve diversity through heuristic constraints or item-level diversification strategies, but rarely model the underlying evolution process of user intent.In this paper, we propose MACRDR, a Multi-Agent Collaboration and Reflection framework for intent-calibrated and diversity-aware news recommendation. Different from conventional reranking-based diversification methods, MACRDR reformulates recommendation as a dynamic intent calibration process consisting of base retrieval and reflection-guided reranking. Specifically, we introduce a multi-agent collaboration mechanism with a predictive-reflective feedback loop, where the Action Agent estimates users' potential interests based on historical profiles, while the Reflect Agent compares predicted interests with actual interactions and generates explicit cognitive adjustment patches. To support long-term intent evolution, a version-controlled dynamic profile memory module is designed to preserve users' evolving preferences and reflective corrections across interaction periods. During online recommendation, the Action Agent performs intent-aware reranking by jointly considering base relevance scores and reflective guidance.Extensive experiments on the MIND dataset demonstrate that MACRDR achieves substantial improvements in interest discovery and serendipity while maintaining competitive recommendation accuracy. The results verify that modeling user intent evolution through multi-agent reflection provides an effective and interpretable solution for mitigating filter bubbles and enhancing recommendation diversity.</p>