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

Generative recommendation models, which represent items as semantic IDs (SIDs), often struggle to balance content-based representations with collaborative filtering signals. While these models are designed to leverage natural language, explicit reasoning or "thinking" often fails to improve recommendation accuracy because SIDs are primarily optimized for content, leaving collaborative co-occurrence patterns under-represented. The authors investigate whether a reliable, training-free channel can be established to inject these missing collaborative signals into the generation process without retraining the backbone model.

The proposed framework operates in three distinct phases:

1. Collaborative Bridge Construction: The authors use hierarchical Poisson factorization on item-item co-view matrices to extract collaborative latent factors. These factors are then aggregated into "prefix signatures" that correspond to the hierarchical levels of the SIDs.
2. Natural Language Mapping: An LLM summarizes user history into audience and behavior tags. A learned bridge maps these tags into the same collaborative factor space, allowing the model to represent a user's preference as a vector that is directly comparable to the item prefix signatures.
3. Level-Differentiated Reranking: During inference, the model uses these collaborative vectors to rerank candidates at each level of the SID generation. The residual signal is injected at the coarse (s_a) and mid-level (s_b) stages to promote items that align with the user's collaborative profile, while the final level (s_c) is constrained by inventory availability.

The framework consistently improves recommendation accuracy across different model scales (1.7B and 8B parameters). The authors find that the performance gains are primarily driven by the mid-level (s_b) decision, where the collaborative signal effectively differentiates between content-similar items that have different audience appeal. Furthermore, the natural-language-reconstructed queries are shown to be effective proxies for behavioral data, successfully differentiating user audiences without requiring access to raw interaction history at inference time.