Can Generative Recommendation Reach Cold Items? A Temporal Perspective on Semantic-ID Generation

This paper investigates whether Semantic-ID (SID) based generative recommendation models can truly perform 'open-token' cold-start induction—that is, recommending new items that were not present during training. While SID models are praised for their ability to recombine tokens to form new item representations, the authors question whether this 'closed-world' compositionality translates into the ability to reach items that contain entirely new atomic tokens or unsupported SID paths.

The authors introduce an absolute-time temporal protocol using sliding windows to separate training data from future test data. This setup ensures that 'cold' items are those that genuinely emerge after the training period, preventing the data leakage common in standard leave-one-out evaluation. The researchers then perform a token-level diagnosis, categorizing cold items by their SID support (e.g., all-token-seen, any-token-unseen, or prefix-seen) and using oracle-prefix probing to determine whether failures occur at the coarse semantic bucketing stage or the fine-grained path-completion stage.

The study reveals that SID-based generation functions as a hierarchical semantic bucketing process. Early tokens in an SID sequence select coarse semantic regions, while later tokens refine the specific item path. Consequently, models can successfully reach cold items only if the required tokens and prefixes were observed during training. When a cold item requires an unseen atomic token or an unsupported path, the model's performance degrades significantly. This demonstrates that current SID-based generative recommendation is compositional but not truly open-ended.

This research clarifies the fundamental limits of generative recommendation. By identifying that SID generation is a constrained search over a learned discrete space rather than a flexible generative process, the authors provide a roadmap for future improvements. They suggest that to overcome these cold-start boundaries, researchers should focus on developing more independent SID spaces, integrating scoring-based interfaces, and incorporating dynamic textual context to better handle items that fall outside the learned token vocabulary.