This paper describes team DArgk’s approach to the ACM RecSys Challenge 2024, which focused on effective and responsible online-news recommendation. The method models users from their implicit reading behavior while accounting for the rapidly changing news agenda and the short lifetime of news items. A transformer-based architecture encodes each user’s reading history and ranks articles according to estimated click preferences, while also considering beyond-accuracy behavior for users whose interests differ from the average. The submitted system placed second in the academic track, illustrating the value of sequence-based user-history modeling for news recommendation under highly dynamic conditions.