New generative recommender system boosts Recall@1 by 29%

A new paper from researchers Sixiao Zhang, Mingrui Liu, and Cheng Long introduces a fully generative conversational recommender system that overcomes the limitations of existing approaches. Traditional systems either decouple recommendation from dialog generation or rely on retrieval-based pipelines, which limits integration and leads to suboptimal user intent modeling. The proposed method represents items as discrete semantic IDs and integrates them directly into the generation process, enabling joint prediction of items and responses via next-token modeling.

The system introduces a structured generation paradigm that factorizes conversational recommendation into a sequence of interdependent decisions. The model first predicts the response intent and the recommendation target, then generates the response conditioned on those predictions. This design enables end-to-end optimization, enforces a coherent dependency structure, and supports faithful item generation through constrained decoding. Extensive experiments show consistent improvements, achieving gains of up to 29% on Recall@1 over strong baselines while maintaining competitive dialog quality. The work bridges a key gap in conversational AI, potentially enabling more natural and accurate recommendation interfaces.