Curriculum Vitae

Research Interests

Long-horizon LLM agents, including multi-agent collaboration and memory for long-term interaction. Previous work on reliable LLM evaluation (LLM-as-a-judge, multi-agent debate) and multimodal benchmarking.

Education

University of Central Florida
Ph.D. Student in Computer Science (Aug. 2026 – Present)

• Advisor: Prof. Song Wang

University of Science and Technology of China
B.S. in Computer Science and Technology (Sept. 2021 – Jul. 2025)

• Cumulative GPA: 85/100
• Honors & Awards: Outstanding Student Scholarship – Silver (Top 5%), 2022; Yongman Yang Scholarship

The University of Sydney
Exchange Student in Computer Science (Feb. 2024 – Jun. 2024)

• Overall result: Distinction

Publications

Full list on Google Scholar.

• Multi-Agent Debate for LLM Judges with Adaptive Stability Detection
  T. Hu, Z. Tan, S. Wang, H. Qu, and T. Chen · NeurIPS 2025 · Paper

• MemRouter: Memory-as-Embedding Routing for Long-Term Conversational Agents
  T. Hu, W. Lin, W. Zhang, J. Ma, and S. Wang · arXiv 2026 · Paper

• PPTBench: Towards Holistic Evaluation of Large Language Models for PowerPoint Layout and Design Understanding
  Z. Huang, X. Liu, T. Hu, K. Zhang, and Y. Liu · arXiv 2025 · Paper

Research Experience

Long-Horizon LLM Agents: Multi-Agent Search and Memory
Advised by Prof. Song Wang, University of Central Florida (Jul. 2026 – Present, Orlando, FL)

• Studying multi-agent collaboration for long-horizon web search, where parallel LLM agents work on the same hard question and share passage-cited findings as they search.
• Built the experiment harness for running teams of LLM search agents.
• Showed that aggregating the team’s pooled findings with an LLM significantly outperforms majority voting across benchmarks.
• Currently training a learned delivery controller that decides which shared findings to push to each teammate, and extending the study to other backbones and mixed-model teams.
• Earlier developed MemRouter, an embedding-based memory router for long-term conversational agents that is both more accurate and faster than LLM-based memory management.

LLMs and Knowledge Graphs for EHR-Based Diagnosis
Advised by Prof. Song Wang, University of Central Florida (Jul. 2025 – Jul. 2026, Remote)

• Investigated combining LLMs, retrieval-augmented generation (RAG), and medical knowledge graphs to support diagnostic reasoning over electronic health records (EHRs).

Multi-Agent Debate for LLM Judges
Advised by Prof. Tianlong Chen, UNITES Lab, UNC Chapel Hill (Feb. 2025 – May 2025, Remote)

• Proposed a multi-agent debate framework for LLM-as-a-judge in which agents collaboratively reason and refine their verdicts, and proved that debate amplifies correctness over static ensembles.
• Designed an adaptive stopping mechanism based on the stability of judge consensus, improving accuracy over majority voting at low computational cost.

PPTBench: Benchmarking LLMs on PowerPoint Tasks
Advised by Dr. Kai Zhang, BDAA Lab, USTC (Sept. 2024 – Jul. 2025, Hefei, China)

• Co-developed PPTBench, a multimodal benchmark evaluating LLMs on PowerPoint detection, understanding, modification, and generation tasks.
• Revealed a substantial gap between LLMs’ understanding of slide content and their visual-layout reasoning.

Quantifying the Influence of Mainstream Media
Advised by Prof. Tristram Alexander, The University of Sydney (Feb. 2024 – Sept. 2024, Sydney, Australia)

• Analyzed 1M+ YouTube comments to quantify how external events shift audience sentiment.
• Preprocessed large-scale comment data and applied LLMs and clustering methods to characterize audience responses.

Other Experience

• Intern, Eigen AI (Aug.–Oct. 2025): model-serving and Docker deployment for an inference platform.
• Research Assistant, Prof. Fuli Feng, USTC (Jul. 2025 – Jul. 2026): frontend for a SASRec recommender.