EchoRec Enhances Generative Recommendation with Multi-Token Prediction and Cycle Consistency

_reachsumit · x · 2026-08-17

Core Method: EchoRec leverages Multi-Token Prediction (MTP) to forecast user actions and uses cycle-consistent preference alignment to verify the model's absorption of future behavior signals.

Architecture: Comprises two modules: 1) Horizon-aware Preference Generation (HPG): Chains auxiliary branches to model preference evolution sequentially. 2) Verifiable Holistic-Preference Alignment (VHA): Consolidates multi-horizon preferences and suppresses spurious alignment via cycle-consistent projectors.

Insight: Analysis reveals that future behaviors act as a "semantic echo" of current ones, decaying over intent transitions, serving as informative yet order-dependent supervision signals.

Original post →

More from Research

Research channel →