In-context learning
Characterizing how demonstration quality, prompt consistency, and pretraining support govern generalization to unseen tasks.
Xuhan Tong
Tsinghua University · Mathematics & Physics + Software Engineering
Research
Toward stable and reliable inference in language models.
My work develops theoretical and empirical accounts of when an LLM can recover a supported inference path, why it may instead follow a salient shortcut, and how a system can diagnose and repair the failure.
Hallucination mechanisms
Studying when high-frequency associations overpower decisive prompt constraints, producing inference misalignment despite available knowledge.
Detection and intervention
Separating knowledge gaps, context distraction, compute insufficiency, and calibration failure so that mitigation can target the active cause.
Multi-agent reliability
Investigating communication and coordination strategies for agents solving complex tasks under heterogeneous and conflicting constraints.
Publications
Selected work
Experience
Research experience
University of Wisconsin–Madison
My research centers on reliable language-model inference. I have developed theoretical accounts of in-context learning and Chain-of-Thought prompting, studied hallucination as inference misalignment between prompt evidence and pretraining priors, and investigated counterfactual interventions that distinguish knowledge-deployment failures from ignorance.
My current work extends this perspective to multi-agent systems, examining how specialized agents communicate and coordinate when solving complex tasks under multiple constraints.
Background
Education
B.S., Mathematics & Physics + Software Engineering
Tsinghua University, Beijing
Selected honors
- 2025Comprehensive Outstanding Scholarship
- 2024Science & Technology Innovation Outstanding Scholarship
- 2024Academic Excellence Scholarship