HoloRec: Holistic Encoding and Interleaved Reasoning for Generative Recommendation

HoloRec is a novel generative recommendation model that addresses the limitations of traditional cascade architectures by integrating hierarchical semantic representations with an endogenous chain-of-thought mechanism. It utilizes a multi-granularity nested residual quantization approach optimized by a holistic reconstruction loss, offering two inference modes: a lightweight non-thinking mode for fast predictions and a thinking mode that generates reasoning steps on-the-fly. Experimental results show that HoloRec outperforms existing models, particularly in sparse data scenarios, making it a valuable tool for practitioners seeking to enhance recommendation systems with integrated reasoning capabilities.