Hierarchical Residual Policy Optimization for Generative Recommendations
Published in KDD 2026 Research Track, 2026
Hierarchical Residual Policy Optimization (HRPO) is a post-training framework for semantic-ID generative recommendation. It converts item-level logged outcomes into dense, token-aligned residual credits and optimizes the decoder with conservative token-wise policy updates.
Recommended citation: Kaifeng Guo, Yiming Yang, Jingtong Gao, Guolei Zeng, Fukang Yang, Yukang Liang, Peng Jiang, Qingpeng Cai, and Xiangyu Zhao. (2026). "Hierarchical Residual Policy Optimization for Generative Recommendations." KDD 2026 Research Track.
Download Bibtex