Generative recommendation systems that rely on semantic IDs (SIDs) retrieve items by making successive token-level decisions. A new paper on arXiv argues that personalization at each decoding step is essential for accurate retrieval, yet existing methods do not sufficiently account for this. The paper proposes SPAR (Stepwise Personalized Autoregressive Recommendation) to fill that gap.
SPAR uses prefix-balanced semantic IDs to support personalization throughout the autoregressive decoding process. By balancing the prefixes of semantic IDs, the method aims to ensure that every step of the retrieval path is tailored to the user, rather than only the final decision. The abstract does not include experimental results, but the approach directly targets a known weakness in current SID-based generative recommenders.