Hi, I’m Mingyi Deng (邓明熠), an MStat student at the University of Hong Kong. I got my B.S. in Statistics (Data Science + Applied Economics) from Renmin University of China.
I work on LLM-based agents — especially interactive and user-centric ones: how agents should recognize ambiguity, ask good clarifying questions, and learn from multi-turn user feedback. I’m currently an LLM research intern at Meituan Beam.
Reach me at dengmingyi1219@163.com.
Publications
- InteractComp: Evaluating Search Agents With Ambiguous Queries · ICML 2026 · first author A benchmark for evaluating whether search agents can recognize query ambiguity and interact to resolve it.
- InfoPO: Information-Driven Policy Optimization for User-Centric Agents · ICML 2026 Turns per-turn information gain into a dense signal for multi-turn RL on user-centric agents.
- ReCode: Unify Plan and Action for Universal Granularity Control · arXiv 2025 Unifies planning and action as executable code with universal granularity control.
Education
- MStat, The University of Hong Kong · 2025 – 2026
- B.S. in Statistics (Data Science + Applied Economics), Renmin University of China · 2020 – 2024
Internships
- Meituan Beam · LLM Research Algorithm Intern · Apr 2026 – Present
- DeepWisdom (深度赋智) · Research Intern · Jul 2025 – Feb 2026 LLM agent research: InteractComp, InfoPO, ReCode.
- Nanjing Xuming Private Fund · Quant Research Intern · Feb 2025 – May 2025 Tree-model / NN composite quant models; agent-driven strategy exploration.