On the Memorization Behavior of LLMs in Generative Recommendation: Observations, Implications, and Training Strategies

Collection of recent IR and recommender-systems papers (arXiv 2026-06-17) focused on LLM-based generative recommendation, reranking, temporal and diversity-aware retrieval, failure analysis, and domain-specific LLM information extraction. Key contributions: (1) analysis of LLM memorization in generative recommendation and IIRG, a training strategy to reduce one-hop memorization and teach multi-hop collaborative and semantic relations; (2) RSRank, a reranker using representational-shift alignment for calibrated relevance scoring; (3) TPOUR/TRPO, a method for temporal preference optimization in