A new arXiv paper addresses how generative recommender systems can keep up with user preferences that shift over time. The authors propose a self-evolving memory mechanism intended to let these models update their behavior as tastes change. The work is categorized as a cross-list submission under arXiv cs.AI.
A revised arXiv paper introduces CRAMER, a technique for controlling sequential recommendation models through request-aware masking. The approach aims to let such models respond more flexibly to users' immediate, time-sensitive requests rather than only their long-term behavior patterns. The paper appears on arXiv's cs.AI and cs.LG listings as a cross-list replacement.
A revised arXiv preprint introduces a multimodal recommendation method that combines collaborative signals with visual and textual item features. It uses modality-guided mixture of structured experts, where entropy-based routing decides how much to rely on each evidence source per user-item interaction. Diagnostic probes trained on individual modalities are used to partition held-out interactions and guide the routing behavior.
A new arXiv preprint proposes treating safety as a constraint when fine-tuning a large language model used as a recommender system. Standard recommenders are trained only to predict the next item a user will engage with, not to justify the prediction, so the authors add self-explanation as a training objective. The goal is to give users personalized reasons for suggestions without letting the generated explanations violate safety requirements.
A revised arXiv paper introduces a contextual scalarisation Thompson sampling approach for recommender systems that must balance several competing goals at once. The authors use public service media editorial decisions as a motivating case, where audience reach, cultural values, public service duties, and operational limits all compete. The method aims to support decision-making when objectives cannot be easily reduced to a single score.
A new arXiv paper introduces TATK, a framework aimed at the mismatch between text-generation-style next-item prediction and full-catalog top-K ranking in LLM-based sequential recommenders. It combines Top-K Learning with knowledge-grounded verification to better align the model's training objective with ranking tasks. The work is presented as a preprint and no production deployment is reported.
A new arXiv preprint examines how a standard remedy for class imbalance — assigning each label a positive-class weight equal to the ratio of negatives to positives — alters the ranking of the top K items in one-vs-rest rankers. Drawing on Elkan's identity, which shows the weight shifts a label's log-odds, the authors characterize the resulting slippage in top-K results and propose a way to diagnose and repair those errors.
A new arXiv paper presents a coordinated multi-agent approach for manipulating recommender systems by simulating user behavior to steer ranking outcomes. The authors report that the method can influence recommendations effectively while keeping resource costs manageable. The work highlights a security concern for platforms that depend on user-interaction data to personalize content and purchases.
A revised arXiv paper cross-listed in AI and machine learning presents a gamified recommender that uses a 20-questions style interaction to support cybersecurity education. The authors argue that traditional security training often leaves learners disengaged and position the game-based tool as a more explanatory alternative as threats grow more sophisticated. The v2 entry appears in both the cs.AI and cs.LG listings.
A research paper presents a method for timing scheduled batch jobs on large asynchronous computing platforms, where tasks such as promotional notifications and recommender pre-computation must run at the right moment. The approach personalizes execution times across very large job volumes, aiming to improve how promptly information is delivered or refreshed.
A new arXiv preprint analyzes how regularization shapes the behavior of linear recommendation models. The work situates these simpler models against deep-learning-inspired recommenders that currently lead on standard recommendation benchmarks. It appears aimed at clarifying when and why regularization choices matter for ranking performance.