M.S. Candidate, Department of Automation, Tsinghua University
LEAP Lab · advised by Prof. Gao Huang
_{he/him  ·  Beijing, China  ·  george.tsinghua@gmail.com}

I work at the intersection of large language models, agent systems, and embodied intelligence. My recent research is organized around three questions:

1. How should multiple agents coordinate so that emergent behavior is reliable, not merely plausible?
2. What is the right inference substrate for an agent-native workload — one where prefixes, tools, and policies, rather than tokens, dominate the cost surface?
3. Can hierarchical, experience-accumulating planners take LLMs out of the simulator and onto real construction sites?

In parallel, I founded SeamLess AI, where the same questions are stress-tested against industrial robotics, judicial agents, and CAx software at scale.

• Agent Systems — multi-agent coordination, agent memory, prompt-free workflows, A2A / MCP protocols
• Agent-Native Inference — KV-cache governance, stable-prefix protocols, parallel scheduling on Ascend (MindIE / CANN)
• Embodied Intelligence — hierarchical planning (TSP × LLM × RL × PDDL), sim-to-real for industrial robots
• Retrieval and Long-Form Reasoning — score-gated termination, access-driven materialization for RAG

• CostRAG: Free Signals Are Enough — Pareto-Improving Long-Form RAG via Score-Gated Termination and Access-Driven Materialization.. First author.
• HEAT: Hierarchical Thinking with Long-Term Experience Accumulation and Test-Time Evolvement in Real-World Robotic Environments. First author.
• Z. Wang et al. A Frequency-Domain Scheme for High-Speed Telemetry Downhole Wireline Communication. IMCCC 2015 (EI). First author.
• Vertical Magnetic Field and Its Analytic-Signal Applicability in Oil-Field Underground Pipeline Detection. Journal of Geophysics and Engineering, 2015 (SCI).

Cooragent— an open-source multi-agent collaboration framework from LEAP Lab. Designed and shipped end-to-end; covered as a SOTA framework by Synced (机器之心) and roughly ten other outlets.

Telos — an agent-native inference engine targeting Ascend hardware. Combines fleet parallelism, a stable-prefix protocol (PIN → FOLD → DROP), and token-efficient tool interfaces; integrated with MindIE / CANN.

HEAT (Robotic Plastering) — joint work with Country Garden / Bright Dream Robotics: the first LLM-driven plastering robot deployed on real construction sites. A three-level planner — layout-TSP × cross-room LLM × in-room LLM + RL + CV + PDDL — yields +34.7% over LLM+P and +11.5% over Lamma-P on VirtualHome, with a 90% task-completion rate in field deployment.

Autonomous Cardiac-Ultrasound Robot — with Lingshi Tech: a full perception – decision – control loop. Featured as a Nature Biomedical Engineering Research Highlight by Editor-in-Chief Rita Strack.

• M.S., Automation, Tsinghua University, 2022 – 2026. LEAP Lab, advised by Prof. Gao Huang (DenseNet; CVPR 2017 Best Paper).
• M.S., Information & Communication Engineering, China University of Petroleum (Beijing), 2013 – 2016. Ranked 1st in cohort; Outstanding Graduate Student of Beijing Municipality.

• Dream-Launch / Kunpeng-Ascend Seed Program, Tsinghua University × Huawei
• Global Runner-up, Tsinghua Sanchuang Entrepreneurship Competition (Seed Group)
• Champion, Tsinghua Guoqiang Institute Dual-Innovation Competition (Seed Group)
• Bronze Award, Huawei Ascend AI Innovation Competition
• Outstanding Graduate Student, Beijing Municipal Commission of Education

Before returning to research I spent close to a decade building large-scale systems: kernel-level network virtualization at H3C → a 100k+ VM cloud platform at iQiyi → online-serving and ranking infrastructure at Xiaohongshu (billion-plus daily requests) → heterogeneous edge – cloud computing at the CASC Ninth Academy. That trajectory — systems → platform → algorithms → agents — is what I now bring to academic work on agent infrastructure.

_{This page is intentionally terse. Full CV available on request. Last updated: May 2026.}