Persistent agents
Maintaining context, state, and task continuity across long-running human–agent work.
Zhiyu Yang · CS PhD @ UT Dallas
I study how agents persist, learn from experience, and spend intelligence where it matters.
Research experience: SMU · THUNLP · ModelBest · OpenBMB
Research threads
AI agents, real-world evaluation, and human–AI collaboration.
Research
The common thread is continuity: how an agent keeps its bearings, learns from experience, and adjusts as the task, context, and relationship with a person evolve.
Maintaining context, state, and task continuity across long-running human–agent work.
Turning real feedback, corrections, and outcomes into signals that can improve future behavior.
Helping agents adjust their behavior and effort as tasks, feedback, and collaboration with people evolve.
Selected work
EMNLP 2025 Main · Oral
Why Stop at One Error? Benchmarking LLMs as Data Science Code Debuggers for Multi-Hop and Multi-Bug Errors
A data-science debugging benchmark for tracing multi-hop logical errors and resolving multiple bugs in realistic code.
ACL 2024 Findings
Method and Evaluation for LLM-Based Agentic Scientific Data Visualization
A visual-feedback agent that turns natural-language requests into scientific visualizations, then uses the rendered chart to repair its own code.
News
Background
Education and research experience.
CS PhD
M.Eng. in Computer Science and Technology
B.Eng. in Computer Science and Technology