Tianyu Hu
Building LLM agents that stay reliable over long horizons.
B.S. in Computer Science and Technology, USTC (Jul. 2025)
Ph.D. Student in Computer Science, UCF (2026–)
My current research focuses on long-horizon LLM agents, including memory for long-term interaction. Previously, I worked on reliable LLM evaluation and multimodal benchmarking.
Featured Publications
-
MemRouter: Memory-as-Embedding Routing for Long-Term Conversational Agents
Presents MemRouter, an embedding-based write-side memory router that decides which conversation turns long-term agents should store, without per-turn LLM decoding.
-
Multi-Agent Debate for LLM Judges with Adaptive Stability Detection
Introduces a multi-agent debate framework for LLM judges with adaptive stability detection to improve evaluation reliability.
-
PPTBench: Towards Holistic Evaluation of Large Language Models for PowerPoint Layout and Design Understanding
Presents PPTBench, a benchmark for holistic evaluation of large language models on PowerPoint layout and design understanding.