About Me

I am a Ph.D. student in Computer Science at Simon Fraser University, advised by Zhengjie Miao. Prior to my Ph.D., I received my B.Eng. in Computer Science and Technology from Sichuan University (2019–2023), graduating with Honors.

Research Interests

• Reinforcement Learning for Data Science
• Agentic RL

Experience

• Aug 2026 – Present: Research Intern, Microsoft (Vancouver, BC, Canada — Remote)
  • Mentors: Zhongqi Yue, Danrui Qi
  • Working on post-training.
• May 2025 – Aug 2025: Research Intern, Microsoft (Beijing, China)
  • Mentor: Haoyu Dong
  • Post-trained tabular QA model via GRPO in a multi-node, multi-GPU environment using verl and Ray.
  • Developed a multi-stage agent pipeline with LangGraph to handle large spreadsheet QA tasks.
  • Built specialized modules for dynamic context selection, data preview and query verification, allowing the agent to navigate and synthesize information from tables that were too large to encode fully.
• Jun 2022 – Sep 2022: Research Intern, Mitacs (DataPrep.ai) (Burnaby, BC, Canada)
  • Mentor: Jiannan Wang
  • Developed a module to generate performance reports with Gantt charts, CDF charts, and tables to diagnose slow SQL queries.
  • Visualized statistics like table scan history to help identify and remove unnecessary indexes.
  • Implemented error detection for common query issues (e.g., missing columns) before compilation.
  • Implemented query tuning techniques to optimize slow query bottlenecks.

Preprints

• SPA: A SQL-Plan-Aware Reinforcement Learning Framework for Query Rewriting with LLMs — Submitted to VLDB 2027
  • Developed a plan-aware GRPO reward strategy, enabling Qwen-32B (LoRA) to outperform GPT-5.4 and Gemini-2.5-Pro in query rewriting by incentivizing plan change over superficial lexical changes.
  • Extended GRPO with multi-stage rewards, with probability-gated adaptive reward for query-level curriculum learning.
  • Designed adaptive reward shaping and on-policy self-improvement to mitigate sparse rewards, reduce slowdown rewrites, and improve runtime/tail-latency on IID and OOD workloads.
• Structure-Grounded Knowledge Retrieval via Code Dependencies for Multi-Step Data Reasoning
  • Engineered a novel retrieval framework that organizes domain knowledge via a graph induced by function-call dependencies, specifically for complex code generation tasks.
  • Implemented a dependency-path retrieval logic that identifies critical links between semantic input/output tags, providing a more precise context extraction than embedding-based similarity.
  • Validated performance on DABStep, demonstrating that grounding retrieval increases accuracy by 109.1%–283.3% over traditional and graph-based RAG methods while retrieving substantially less context.
• EVE-Bench: Diagnostic Evaluation of LLMs’ Capabilities in SQL Reasoning
  • Developed EVE-Bench, an execution-grounded framework for SQL equivalence verification, utilizing counterexample generation as a diagnostic lens to pinpoint LLM reasoning bottlenecks.
  • Utilized schema abstraction to identify LLMs’ potential reliance on semantic hints, detecting when models bypass structural reasoning in favor of lexical cues.
  • Benchmarked 8 LLMs across 2,900+ query pairs, discovering that the most advanced models outperform traditional formal methods by up to 155% while exhibiting complementary failure modes.

Education

• Ph.D. in Computer Science, Simon Fraser University, 2024 – Present
  • GPA: 4.00/4.33
  • Advisor: Zhengjie Miao
  • Research area: RL for data science, agentic RL
  • Mitacs Globalink Graduate Fellowship
• B.Eng. in Computer Science and Technology, Sichuan University, 2019 – 2023
  • GPA: 3.89/4.00
  • Graduated with Honors

News

• Sep 2025 — Transferred to Ph.D..
• Jan 2024 — Started MSc. in computing science at Simon Fraser University.