引导语义ID粗粒度层级使细粒度层级可学习
原标题：Guiding the coarse levels of semantic IDs makes the fine levels learnable
AI 摘要
该论文提出 Guided SID 方法，通过确定性监督索引分配，强制 RQ-VAE 的粗粒度语义 ID 层级编码预定义的、文本可接地且任务相关的类别属性，同时保持码本可学习。在 6.54 亿至 7.41 亿标识符的生产目录上，引导式检索器在各列表长度上均提升 recall@，MRR 从基线提升至更高值，并更常预测预定义属性。对照实验表明增益来自重构后的粗粒度编码，而非简单地将属性作为额外 token 前置。
正文节选
Guiding the Coarse Levels of Semantic IDs Makes the Fine Levels Learnable Abstract Generative retrieval represents each item by a short Semantic ID (SID)—a sequence of discrete codes from a residual-quantized autoencoder (RQ-VAE)—and casts recommendation as autoregressive generation of that sequence. Because the tokenizer is trained independently to reconstruct an item embedding, its codes are aligned with neither the downstream LLM (they enter the vocabulary as opaque tokens) nor the end task.