A new paper on arXiv revisits the comparison between generative retrieval (GR) and dense retrieval for sequential recommendation. The authors note that GR has become a strong candidate generator as language models and semantic-item identifiers improve, but they argue it is not a uniformly stronger replacement for dense retrieval.
The study introduces a "static value prior" and shows that this prior reverses the gap between the two approaches. While the abstract does not specify the direction of the reversal in detail, the key claim is that the relative performance of GR and dense retrieval depends on this prior.
The finding suggests that the choice between generative and dense retrieval should be informed by the presence of static value signals, rather than assuming GR is always superior. Further details of the experiments and results are not available from the abstract alone.