Do SID Alignments Help Generative Recommendation? A Research and Practice Note on OpenOneRec Diagnostics

Abstract

Semantic-ID-based generative recommendation represents each item as a short token sequence and trains an autoregressive model to predict the next item. This note asks whether Semantic ID tokens should be aligned with natural language or specialize into collaborative codes. We report two studies. First, representation probes on OpenOneRec variants show that SID-to-text explanations are template-sensitive, while SID tokens form relatively clean subspaces separated from ordinary text. Second, on sampled offline logs from a large-scale industrial short-video platform, we run a controlled ablation that holds architecture, data, and training budget fixed and varies only the SID embedding initialization: a Xavier-initialized SID-only model obtains lower validation loss and stronger mid-/large-K HitRate than its alignment-initialized counterpart, when the input contains only a system prompt and a SID sequence. These results indicate that SID alignment is conditional rather than universally beneficial.