arXiv preprint, 2026
Knowledge Weaver trains a curator to turn agent trajectories into reusable knowledge with mutual-information-inspired feedback, while keeping the task executor frozen.
I am an undergraduate student in Computer Science at the University of Science and Technology of China (USTC), where I am fortunate to be advised by Prof. Fuli Feng and Prof. Xiangnan He in the Lab of Data Science.
I am looking for PhD positions starting in Fall 2027. Feel free to reach out!
yuyangzhao662[at]gmail.com
Email / Google Scholar / GitHub / X
I work on LLM-based agents, currently on two topics. The first is single-agent reasoning and learning: how an agent reasons and plans through a task, and how it keeps learning from experience over time (continual learning), so that what it learns carries over to new tasks. The second is multi-agent systems: how agents communicate with each other, how they collaborate on shared tasks, and how they behave in game-theoretic settings where their interests differ.
* indicates equal contribution.
arXiv preprint, 2026
Knowledge Weaver trains a curator to turn agent trajectories into reusable knowledge with mutual-information-inspired feedback, while keeping the task executor frozen.
Findings of ACL 2026
A benchmark for evaluating Agent-as-a-Judge with environment-aware verification across search, data systems, and GUI domains.
CIKM 2025 demo
A proactive GUI agent system for multidomain information integration and user assistance.
arXiv preprint, 2026
An analysis of where test-time scaling and post-training fall short in individual stance prediction, identifying four failure modes: incorrect consensus, selection failure, response overfitting, and early plateau.
Core contributor
A modular specification framework for composing embodied agents from reusable perception, memory, reasoning, reflection, action, and learning components under controlled interfaces.
Reviewer, Agent Behavior Workshop (WAB) at COLM
University of Science and Technology of China (USTC),
Hefei, China
B.S. in Computer Science, September 2023 – Present