BRIDGE：弥合生成式推荐中的语义量化鸿沟
原标题：BRIDGE: Bridging the Semantic Quantization Gap for Generative Recommendation
AI 摘要
论文提出生成式推荐中此前被忽视的「语义量化鸿沟」（SQG）：RQ-VAE 量化阶段利用码本质心、邻域几何与层级熵等丰富连续信息构建语义 ID，但解码器仅获得离散标识符，导致语义锚定缺失、邻域信息丢失与计算分配失当三类问题。作者提出 BRIDGE 框架，通过码本锚定的潜在推理、特权 EMA 自蒸馏与难度感知循环三个耦合组件弥合该鸿沟。在三个 Amazon 基准上，Recall@10 与 NDCG@10 平均提升 10.9% 和 9.7%（最高 12.6% 和 10.7%），并在某大型电商广告平台 5% 流量的为期一周在线 A/B 实验中使 eCPM 提升 0.65%。
正文节选
BRIDGE: Bridging the Semantic Quantization Gap for Generative Recommendation Abstract. Generative recommendation retrieves items by autoregressively generating Semantic IDs (SIDs), where an RQ-VAE quantizes each item into a hierarchical SID sequence that the decoder generates level by level. We identify an overlooked bottleneck in this paradigm: the Semantic Quantization Gap (SQG). Quantization derives SIDs from rich continuous signals, including codebook centroids, local neighborhood geometry,