Researchers propose better LLM recommendations with personalized language

A research team led by Changjiang Han from six institutions (including McGill University and Ant Group) has developed a framework to improve LLM-based generative recommendation systems by addressing a critical flaw in semantic-ID (SID) generation. Current systems struggle because SIDs—discrete codes representing items—often fail to encode collaborative signals (e.g., user-item interactions), leading to misaligned embeddings and suboptimal recommendations. The proposed solution, detailed in a paper submitted to arXiv, introduces personalized natural language as a channel to restore collaborative signals during inference without retraining or altering the backbone model.

The framework works by attaching analyzable links between collaborative patterns and their audiences using natural language, guiding the model to generate more accurate SIDs. Unlike existing approaches that rely on costly multi-round training or explicit reasoning—which often yields minimal gains—the method leverages hierarchical collaborative cues dynamically. In experiments, this approach delivers consistent improvements in recommendation accuracy by grounding generation in collaborative structures at inference time, bypassing the limitations of traditional SID-based systems.