Kuaishou Tech

Apr 16, 2026 · Artificial Intelligence

Hierarchical Semantic RL Tackles Dynamic Action Spaces in Recommendations

Researchers from Kuaishou, Fudan and Tianjin University introduce the Hierarchical Semantic Reinforcement Learning (HSRL) framework, which maps high‑dimensional, dynamic item spaces into a fixed‑size semantic action space via semantic IDs, employs a hierarchical policy network and multi‑level critics, and demonstrates 13‑18% gains on public datasets and an 18.4% ad spend lift in billion‑scale online tests.