Generative recommendation models that use semantic IDs treat retrieval and ranking as a single autoregressive decoding task. Unlike discriminative models, however, they cannot directly look up rich side features of the target item while generating the output. That structural difference can limit their ability to exploit item attributes.
UniSGR-V2, described in a new arXiv preprint, addresses this by introducing stepwise target-side feature anticipation and crossing. The idea is to let the decoder anticipate the relevant side features of the target item at each step and cross them with the semantic-ID sequence, so the generative process can make use of the same information that discriminative models access directly.
The paper's abstract does not provide experimental details, but the proposed mechanism is positioned as a way to bring generative recommenders closer to the capabilities of discriminative ones. The work is an extension of the earlier UniSGR line, with the 'V2' indicating a second version of the framework.