NextPosRec: Next-Positive Token Generation Over Long User Sequences

Abstract We present NextPosRec, an industrial-grade generative recommender that pioneers true end-to-end “vector-to-vector” recommendation—eliminating both the fragmented multi-stage ranking pipeline and the error-prone semantic ID abstraction. In stark contrast to existing generative methods that force items into discrete tokens via quantization, NextPosRec operates natively in continuous space: at inference, it directly generates a d-dimensional prediction vector in the same semantic space as precomputed item embeddings, and retrieves top candidates via efficient Approximate Nearest Neighbor (ANN) search. This design completely avoids quantization loss and ensures perfect training-inference consistency. To achieve industrial scalability, we fully decouple item representation learning into a frozen, upstream dual-tower multi-task system, while the main model employs a lightweight, item-centric tokenizer that fuses external vectors with learnable interaction-type embeddings. Furthermore, a hierarchical sequence construction strategy and a novel Multi-Task InteractType Mask enable effective modeling of ultra-long user histories while minimizing cross-task interference. NextPosRec is currently undergoing large-scale online A/B testing in a real-world product serving millions of users, demonstrating a practical, deployable path toward efficient, stable, and scalable generative recommendation. This manuscript focuses on the model architecture and system design; empirical results from ongoing validation will be reported in a follow-up publication. Keywords: Generative Recommendation, End-to-End Recommendation, Vector-to-Vector Generation, Semantic ID Elimination, Continuous-Space Modeling, Industrial Recommender System, Transformer, ANN Retrieval