PinGER：基于语义ID的Token高效生成式检索
原标题：PinGER: Token-Efficient Generative Retrieval with Semantic IDs
AI 摘要
Pinterest 提出并部署了生成式检索模型 PinGER，通过交错 SID 与稠密 token、非对称注意力掩码，使模型在训练时对每个 SID 位置进行监督，而在服务时每个历史物品仅需一个 token。该方法结合目标条件生成、强化学习和前缀多样性惩罚来提升候选多样性，并配套高效宽束解码服务栈。在 Related Pins 场景上线后，新 Pin 的收藏率与整体收藏率均有提升，端到端延迟降低，并消除了所替代的最近邻搜索成本。
正文节选
PinGER: Token-Efficient Generative Retrieval with Semantic IDs Abstract. Generative retrieval replaces nearest-neighbor search with token-by-token decoding of an item’s semantic ID (SID). But existing SID-based methods primarily use levels of tokens sequentially to represent each item in the input history. As a result, a fixed context window holds times fewer items than it would with one embedding per item. We present PinGER, a generative retriever deployed at Pinterest that retains SID outputs