ReGen：基于引导式 token 细化的可控生成式推荐
原标题：Controllable Generative Recommendation via Guided Token Refinement
AI 摘要
该论文提出 ReGen 框架，将生成式推荐从自回归逐 token 解码改为并行迭代细化（掩码填充），使推理时可直接对高层语义 token 施加控制信号而无需重训练。为应对激进 logit 干预导致的无效物品问题，引入拓扑约束树与置信度感知细化，在 beam 推理中强制合法码本路径。实验显示 ReGen 在适度引导下取得更优的准确率-多样性权衡，并保持生成 Semantic ID 的高有效性，代码已开源。
正文节选
Controllable Generative Recommendation via Guided Token RefinementThanks: † Corresponding author.Thanks: This study was funded by the National Natural Science Foundation of China (Grant No. 72401232), the Natural Science Foundation of the Jiangsu Higher Educational Institution of China (Grant No. 23KJB520037), the XJTLU Research Development Fund (RDF-21-01-053), and Liaoning Liaohe Lab (Grant Nos. LLL24ZZ-02-01 and LLL24ZZ-02-02). Abstract Generative recommendation over discrete Semantic IDs has