I am currently a full-time Research Assistant at City University of Hong Kong, where I work with Prof. Shuang Qiu on training algorithm optimization for LLM MOPD. I received my M.S. in Computer Science from Rice University in 2026 and my B.E. in Software Engineering from Zhejiang University in 2023. My research focuses on training and evaluating LLM agents so that they can reason reliably, use tools, and make long-horizon decisions in interactive environments. Previously, I worked with Prof. Zhisong Zhang on RetailBench and interned at Alibaba on agentic LLM training; these experiences shaped my interest in bridging realistic agent benchmarks, model training, and practical agent systems.

My research interests include:

  • LLM Agent
  • Agentic RL
  • GUI Agent
  • Data Synthesis
Email: zlh20011228@gmail.com Google Scholar LinkedIn GitHub Resume

Education

  • 2024 – 2026, Rice University, M.S. in Computer Science, Houston, USA
    • Relevant Coursework: Computer Systems Architecture (A+), Natural Language Processing (A+), Design & Analysis of Algorithms (A+)
  • 2019 – 2023, Zhejiang University, B.E. in Software Engineering, Hangzhou, China
    • Overall GPA: 3.79/4.0, Major GPA: 3.87/4.0
    • Outstanding Engineer Scholarship, Third-Class Scholarship

Publications

ARR 2026 (Under Review)

Linghua Zhang, Zhisong Zhang, et al. RetailBench: Evaluating Long-Horizon Autonomous Decision-Making and Strategy Stability of LLM Agents in Realistic Retail Environments. [arXiv]

Research Experience

City University of Hong Kong

Jun 2026 – Present On-site Research Assistant | Advisor: Prof. Shuang Qiu

  • Reproduced OPD training pipelines and extended them to code, math, instruction following, and ScienceQA tasks.
  • Ran multi-domain experiments on 6 H200 GPUs to study optimization conflicts and performance trade-offs across tasks.
  • Designed experiments around MOPD training objectives and student sampling strategies for improving multi-domain Agentic RL.

City University of Hong Kong

Dec 2025 – Mar 2026 Research Assistant | Advisor: Prof. Zhisong Zhang

  • Led RetailBench, a realistic supermarket simulation benchmark with 96 products and 7 modules covering inventory, shelves, supply chains, customers, reviews, news, and finance.
  • Designed a 180-day business simulation, oracle baselines, and diagnostic metrics to evaluate long-horizon decision-making, tool use, and strategy stability across 7 LLM/agent frameworks. Published on arXiv, under ARR review.

Zhejiang University, CAD Lab

Dec 2022 – Jun 2023 Undergraduate Thesis | Advisor: Dr. Fan Yan

  • Built a ChatGPT-based financial text analysis and video visualization platform.
  • Proposed a logic-relation-driven framework for financial data visualization.

Work Experience

Alibaba Local Services

May 2025 – Nov 2025 LLM Algorithm Engineer Intern | Agentic LLM Training

  • Built an SFT + workflow rollout + GRPO training pipeline for data-analysis agents, connecting trajectory sampling, reasoning rewrite, data construction, RL fine-tuning, and evaluation.
  • Constructed 8k multi-turn reasoning/tool-use cold-start samples and a 2k vertical RL task set from expert knowledge documents.
  • Conducted GRPO fine-tuning for a 32B model on 64 H20 GPUs, improving both general benchmarks and vertical data-query tasks over the SFT baseline.

Alibaba DingTalk (DingDing)

Jul 2023 – Aug 2024 Software Engineer

  • Worked on DingTalk Docs internationalization, including multilingual text extraction, translation platform integration, and frontend display-flow refactoring.
  • Built commercialization entries, collaboration permission controls, analytics instrumentation, and responsive adaptation for document editing and collaboration scenarios.

Skills

  • LLM Training: Python, PyTorch, Transformers, SFT, GRPO, Agentic RL, multi-turn trajectory rewriting
  • Agent & Evaluation: LLM agents, tool use, workflow rollout, benchmark design, long-horizon decision-making evaluation, diagnostic analysis
  • Engineering: TypeScript, React, frontend engineering, internationalization, responsive design, permission control, analytics instrumentation
  • Languages: English (Fluent, TOEFL: 100), Mandarin (Native)