Restoring Collaborative Signals in Semantic-ID Generative Recommendation via Personalized Natural Language

2026-07-30

AI Summary

Existing generative recommender models based on Semantic IDs (SIDs) struggle to jointly capture content and collaborative signals due to misalignment between textual and SID embedding spaces, thereby limiting recommendation performance. This work proposes an inference-time optimization approach that requires neither architectural modifications nor retraining of the SID model. By leveraging personalized natural language prompts, the method dynamically injects hierarchical collaborative cues during inference, establishing—for the first time—an explicit, interpretable link between collaborative patterns and individual users at test time. This strategy effectively recovers missing collaborative information, significantly improves recommendation accuracy, and seamlessly integrates collaborative structure into the generative process, all while avoiding the computational overhead associated with explicit reasoning or multi-stage training.

Abstract

Making LLM-based generative recommendation models stronger and more personalized through natural language and explicit reasoning is a widely anticipated yet still unsolved goal. Such models cast recommendation as autoregressively generating an item's semantic-ID (SID), a short tuple of discrete codes, so that recommending well reduces to emitting the right SID. In this setting the model verbalizes its knowledge poorly, and text and SID tokens live in misaligned embedding spaces. Deep reasoning therefore rarely turns into a correct SID, and enabling explicit "thinking" often gives no gain or even hurts. The deeper cause is that a compact SID cannot hold content and collaborative signal at once: the two compete, and collaboration loses. Because a mis-predicted SID is a wrong recommendation, this caps accuracy directly. Costly multi-round training barely helps, and few methods try to supply the missing signal at inference time. What is missing is a reliable channel that carries collaborative signal into SID generation. We therefore propose a framework, guided by personalized natural language, that adds hierarchical collaborative cues as the model generates, without altering the backbone or retraining the SIDs. Rather than mapping language onto SIDs directly, it uses language to attach analyzable links between collaborative patterns and their audiences, restoring the collaborative signal that SIDs miss. The result is consistent gains in recommendation accuracy, grounding generation in collaborative structure at inference time rather than relying on explicit reasoning or retraining.

Problem

• generative recommendation
• semantic-ID
• collaborative signal
• personalized natural language
• LLM-based recommendation

Innovation

• generative recommendation
• semantic-ID
• collaborative signals
• personalized natural language
• inference-time augmentation