EchoRec uses cycle-consistent AI to predict multiple items

A team of researchers from the National University of Singapore, Tsinghua University, and Peking University has introduced EchoRec, a novel generative recommendation system that leverages Multi-Token Prediction (MTP) with cycle-consistent preference alignment to better predict future user behavior. The system consists of two core modules: Horizon-aware Preference Generation (HPG) and Verifiable Holistic-Preference Alignment (VHA). HPG sequentially chains lightweight auxiliary branches on top of a base recommender, where each branch conditions on its predecessor to respect the natural evolution of user preferences. VHA then consolidates these predictions into a holistic preference representation and enforces cycle consistency through projectors, ensuring that the preference signals are verifiable and robust against rank-collapse—a common failure mode in recommendation systems.

The researchers demonstrate that EchoRec outperforms existing baselines on three benchmark datasets while introducing negligible online serving overhead, as all auxiliary components are discarded at inference time. The model naturally acquires multi-item generation capabilities, enabling it to predict sequences of items rather than single recommendations. By addressing the decay of semantic echo in future behaviors and suppressing spurious alignments, EchoRec achieves more accurate and holistic preference modeling, setting a new standard for generative recommendation systems.