Preference-Aligned Semantic IDs for Generative Recommendation
Abstract
Semantic IDs (SIDs) are becoming a practical interface for generative recommendation. Existing pipelines usually build them by quantizing frozen content embeddings, implicitly assuming that content similarity is the right geometry for recommendation. This assumption is often too strong: items that look similar in content space may play different roles in user behavior, while behaviorally related items may be far apart in text embedding space. We propose a simple pre-quantization alignment step for SID construction. Starting from frozen content embeddings, we train a lightweight content-to-item adapter inside a sequential next-item prediction model. The adapted vectors replace the standard item embedding table and are optimized with the recommender's cross-entropy objective. After training, the sequential model is discarded, and the adapted item vectors are quantized using a standard SID tokenizer such as RQ-KMeans. The resulting codes preserve the content-based SID interface while aligning the underlying continuous space with user behavior. This design separates two roles that are usually coupled: learning a recommendation-useful item geometry and assigning discrete codes. It is easy to add to existing SID pipelines and requires no changes to the tokenizer or the downstream generative recommender. Experiments show that aligning embeddings before quantization improves ranking quality over content-only SIDs, collaborative-only SIDs, and compression baseline.