CV

Education

• Peking University, Beijing
  • Undergraduate in Intelligence Science, Aug 2023 to Present
• Jiangsu Dongtai High School, Yancheng, Jiangsu
  • Sept 2020 to Jun 2023

Research Experience

• Agentic E-commerce Assistant with Function Calling (In Progress)
  • Joint University-Enterprise Innovation Workstation Project, Beijing / Shanghai / Hangzhou
  • May 2025 to May 2026
  • With support from Xiaoduo AI Co., Ltd.
  • Working on enhancing LLM agent capabilities in e-commerce customer service, especially tool use and function calling.
  • Introducing a long-term memory unit and expanding the e-commerce customer service benchmark.
  • Built an end-to-end LLM agent post-training pipeline, EComBench, integrating supervised fine-tuning and reinforcement learning.
  • Designed rollout strategies to construct high-quality reasoning data from successful agent trajectories.
  • Proposed a dual-path RL data construction framework with step-level semantic evaluation and trajectory-level preference optimization.
  • Introduced reward variance-driven sampling and group filtering to stabilize RL training.
  • Addressed state misalignment, long-context training, and sparse reward signals.
• Research Intern
  • Prof. Chen’s Lab, Peking University
  • Feb 2025 to Present
  • Assisted in spike camera data collection for SpikeStereoNet: A Brain-Inspired Framework for Stereo Depth Estimation from Spike Streams.
  • Currently working on VLA efficiency by designing a conditional SSM-based action denoiser in a flow-matching framework.
  • Reducing cross-modal interaction overhead through low-dimensional conditional summarization.
  • Exploring few-step ODE solvers and distillation for a better efficiency and success-rate tradeoff.
• Collaborator
  • TarraVerse Group, Peking University
  • Dec 2025 to Feb 2026
  • Contributed to TerraVerse, an IROS 2026 submission on a large-scale vision-language terrain dataset.
  • Participated in data collection, cleaning, annotation, and scalable pipeline construction.
  • Worked on cross-dataset aggregation, Mask-to-Patch transformation, VLM-based parallel annotation, and multi-stage quality filtering.
  • Improved annotation scalability and consistency across heterogeneous terrain datasets.

Research Interests

• Embodied Intelligence
• Cognitive Architectures for LLMs
• Model Interpretability

Publications

SpikeStereoNet: A Brain-Inspired Framework for Stereo Depth Estimation from Spike Streams

Zhuoheng Gao, Yihao Li, Jiyao Zhang, Rui Zhao, Tong Wu, Hao Tang, Zhaofei Yu, Hao Dong, Guozhang Chen, and Tiejun Huang. (2026). "SpikeStereoNet: A Brain-Inspired Framework for Stereo Depth Estimation from Spike Streams." ICLR 2026.

Hobbies

• Cue sports
• Molecular gastronomy
• Karaoke
• Learning Cantonese and Japanese through multilingual pop songs