About Me

Welcome to my academic page! I am an undergraduate student currently pursuing a B.S.E in Computer Science at the University of Michigan, Ann Arbor, and a B.Eng. in Mechanical Engineering at Shanghai Jiao Tong University.

My primary research interest is in AI Agents, with a particular focus on Web Agents and their memory. My goal is to develop highly capable and intelligent systems that can reliably perceive, reason, and act in complex web environments. I am particularly interested in evaluating, benchmarking, and developing robust MLLM-based agentic systems.

Research Experience

• Agent Memory Decay Under Controlled Web Drift (Jan 2026 – Present) [Progress Report]
  Advisor: Prof. Honglak Lee (University of Michigan)
  Investigating how test-time agent memory (insights vs. workflows) degrades when website UIs change. Developed an evaluation framework measuring Experience Transfer Gap (ETG) and Decay Rate (EDR) across 6 controlled drift variants (e.g., surface, structural, content). Comparing the robustness of different memory abstraction levels under version shift to uncover critical failure modes and inform adaptive agent architectures.

• Efficient Inference for Embodied Foundation Models (Nov 2025 – Present)
  Advisor: Dr. Jiachen Liu (University of Michigan)
  Profiled Video Action Models (VAM) for robotic policy, implementing cross-attention KV caching and token compression to accelerate inference in dynamic environments. Explored quantization and speculative decoding on VLA models to optimize real-time decision-making.

• Gravitational Effect on Swarming Behavior of Microorganisms (Sep 2024 – Aug 2025)
  Advisor: Prof. Zijie Qu (Shanghai Jiao Tong University)
  Trained a U-Net model for colony boundary segmentation; diagnosed systematic failure modes, then adopted a SAM-based pipeline that achieved reliable automated detection, replacing manual annotation.

Selected Projects

• Probabilistic Motion Planning for Redundant Robots (Nov 2025 – Dec 2025)
  EECS 465 (Intro to Algorithmic Robotics) course project
  Evaluated RRT-Connect and PRM on a 7-DOF Franka Panda; optimized sampling strategies in high-dimensional state spaces, achieving a 75% reduction in planning latency.

Technical Skills

• Frameworks: PyTorch, Hugging Face, CUDA, Docker, Linux
• Languages: Python, C/C++, Java, SQL, LaTeX, Typst