STEP-HRL enables efficient long-horizon decision-making through subtask planning and compact local progress tracking, avoiding the need to condition on an ever-growing interaction history while reducing token usage.
About Me
Hello, I am Shuai Zhen (甄帅), a master's student in Computer Science and Technology at Beijing University of Posts and Telecommunications (BUPT). I received my Bachelor's degree from Tiangong University in 2024.
My research interests lie in Reinforcement Learning (RL) and Large Language Models (LLMs), with a particular focus on:
- Improving the sample and training efficiency of RL for LLMs.
- Exploring novel applications of RL to enhance LLM capabilities.
- Developing more efficient and effective attention mechanisms.
Selected Publications
ACL 2026 · Main ConferenceHierarchical Reinforcement Learning with Augmented Step-Level Transitions for LLM Agents
A 1B-parameter hierarchical recurrent language model trained on 40B unique tokens, outperforming Llama 3.2 3B, Gemma 3 4B, Qwen 3.5 2B, and OLMo 3 7B on ARC-C, DROP, GSM8K, and MATH.
Integrating reflection symmetry into PPO and SAC to improve sample efficiency in continuous control.
Education
-
2024.09 – Present
Beijing University of Posts and TelecommunicationsMaster, Computer Science and Technology, advised by Prof. Yanhua Yu.
-
2020.09 – 2024.07
Tiangong UniversityBachelor, Information Management and Information Systems.
Experience
- 2026.07 – 2026.09, Summer Internship, ASI Lab, National University of Singapore (NUS), advised by Prof. Dianbo Liu.
- 2023.09 – 2024.02, Exchange, Universitat Politècnica de Catalunya (UPC).
Internships
-
2025.02 – 2026.07
Sapient Intelligence
-
2024.02 – 2024.08
MeiTuan