生成式推荐中 Beam Search 的首步覆盖瓶颈与预算重分配
原标题：Where Beam Search Loses Top-Ranked Items:First-Step Coverage in Generative Recommendation
AI 摘要
该论文研究生成式推荐中 beam search 的排序损失定位问题，发现 TIGER 推荐器在六个全测试设置下，所有被剪枝的精确 top-10 路径均首次丢失于解码第一步，即首步覆盖瓶颈。作者提出将 beam 宽度预算重新分配到首步，在 Beam-50 不变预算下将 Search Recall@10 提升 0.96–10.90 个百分点，并在全部 77,373 个 RQ-VAE 测试用户上以约一半前缀容量预算复现 Beam-100 的有序 top-10，同时降低 6.35–6.39% 的解码延迟，无需重训练或修改标识符与分数。
正文节选
Where Beam Search Loses Top-Ranked Items: First-Step Coverage in Generative Recommendation Abstract. Generative recommendation can reproduce a wider beam’s ranked output with roughly half the prefix-capacity budget by reallocating retention capacity across semantic-ID positions. We identify a first-step coverage bottleneck: across six full-test settings of the TIGER generative recommender spanning three domains and two identifier constructions, every pruned exact top-10 path is first lost at the