ONeRec: Towards Openness-Aware and Adaptive Proactive News Recommendation

Proactive news recommendation seeks to guide users over extended interaction sessions towards a cultivated interest in targeted news, thereby shaping public opinion and contributing to social stability. Conventional news recommendation algorithms, by contrast, are largely passive: they rely solely on a user's historical preferences, a practice that exacerbates filter-bubble effects and opinion polarization. To mitigate these drawbacks, proactive news recommendation strategically adjusts the sequence of suggested articles so that users gradually cultivate an interest in a target. This paradigm, however, presents three central challenges: (i) accurately modeling a user's receptiveness to novelty; (ii) tracking evolving interests across multiple rounds of proactive recommendation; and (iii) selecting intermediary articles that balance immediate relevance with long-term target guidance. To tackle these challenges, we introduce ONeRec, a novel framework towards user Openness-aware and adaptive proactive News Recommendation. ONeRec steers users towards target news by adaptively recommending target-relevant intermediate news items according to the user's openness and current interest. ONeRec incorporates two personalized mechanisms: an openness coefficient, derived from reading history, that models a user's tolerance for novelty and balances interest matching with target guidance; and an evolutionary coefficient, which dynamically updates user interest as they engage with recommended news. To support offline training and evaluation, we further employ a Large Language Model agent to simulate user feedback. Extensive experiments on the public MIND dataset demonstrate that ONeRec consistently outperforms strong baselines in proactive news recommendation scenarios.