EchoRec: Cycle-Consistent Preference Alignment for Generative Recommendation

A new paper on arXiv (2608.14011) introduces EchoRec, a method that enhances generative recommendation systems by leveraging multi-token prediction (MTP) for dense supervision. The authors observe that future user behaviors carry a semantic echo of current behavior, which is informative but order-dependent, decaying over time due to intent transitions. EchoRec comprises two modules: Horizon-aware Preference Generation (HPG) and cycle-consistent alignment, which together align preferences across multiple horizons. The approach aims to improve the efficiency and accuracy of generative recommendation by using future behaviors as supervision signals. The paper is a cross-type submission and is available on arXiv.