TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems

This collection of recent IR/ recommender-system papers presents multiple industrial-scale advances: TokenMinds (SID-based discrete user tokens + dense embeddings served asynchronously at billion-user scale), Recommendation-as-Generation for on-demand personalized video generation, S2-CAR for complexity-adaptive sequential recommendation, Adaptive Re-Ranking for cost-aware routing, and AutoRelAnnotator for calibrated model cascades in offline relevance annotation. Other contributions include extreme meta-classification for zero-shot retrieval (EMMETT/IRENE), TheoremGraph bridging informal and