MemoHarness: Agent Harnesses That Learn from Experience
Preprint Co-first author * Equal contribution
About Me
I am an incoming Ph.D. student at the University of Notre Dame, joining in Fall 2026 under the supervision of Prof. Fanny Ye. Prior to that, I completed my undergraduate studies in Cyber Science and Engineering at Sichuan University.
During my visiting studies, I was mentored by Prof. Xiangliang Zhang and senior researcher Yue Huang, whose guidance shaped my research interests in trustworthy AI and foundation models. I continue to collaborate closely with Zheyuan Zhang.
My research focuses on trustworthy AI for LLM agents, with an emphasis on reliability, alignment, and evaluation under realistic training and deployment conditions.
News
May 2026
AutoDavis was accepted to the KDD 2026 Datasets & Benchmarks Track. Two papers were accepted to ICML 2026: Capability-Oriented Training Induced Alignment Risk and Drift-Bench.
Apr 2026
Two papers were accepted to ACL 2026 Findings.
Jan 2026
One paper was accepted to WWW 2026 Demo Track. One paper was accepted to ICLR 2026.
Aug 2024
One paper was accepted to AAAI 2025.
Research Interests
-
Reliable LLM Agents and Failure Diagnosis
I study how LLM agents fail across multi-turn interaction, input faults, and cooperative settings, how those failures can be diagnosed, and how execution experience can make agent harnesses more adaptive.
Selected works MemoHarness arXiv'26 Drift-Bench ICML'26 IntraAI WWW'26 Demo -
Alignment under Training and Deployment Shifts
I investigate how alignment risks emerge as models acquire new capabilities or move into constrained deployment environments, together with guardian and advisor mechanisms for safer behavior.
Selected works Capability-Oriented Training Induced Alignment Risk ICML'26 Guardian-as-an-Advisor ACL'26 Findings AI Alignment Breaks at the Edge Position Paper -
Evaluation for Trustworthy Foundation Models
I develop benchmarks and evaluation protocols that test whether generative, vision-language, and domain-specific foundation models remain reliable in realistic use cases.
Selected works TrustGen ICLR'26 AutoDavis KDD'26 D&B PolicyLLM ACL'26 Findings
Publications
Recent Preprints
AI Alignment Breaks at the Edge
Preprint First author
Selected Work
ICML 2026
Drift-Bench: Diagnosing CoopeRative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction
International Conference on Machine Learning First author
ICML 2026
Capability-Oriented Training Induced Alignment Risk
International Conference on Machine Learning Co-first author * Equal contribution
KDD D&B 2026
AutoDavis: Automatic Benchmarking of Large Vision-Language Models on Visual Question-Answering
ACM SIGKDD Conference on Knowledge Discovery and Data Mining Co-first author * Equal contribution
ACL Findings 2026
PolicyLLM: Towards Excellent Comprehension of Public Policy for Large Language Models
Association for Computational Linguistics First author
WWW Demo 2026
IntraAI: Bridging Human-AI Understanding Through Intelligent Interface Design
The Web Conference Demo Track First author
More Publications 9 additional works
Conference Papers
Guardian-as-an-Advisor: Advancing Next-Generation Guardian Models for Trustworthy LLMs
ACL Findings 2026 Association for Computational Linguistics
On the Trustworthiness of Generative Foundation Models: Guideline, Assessment, and Perspective
ICLR 2026 International Conference on Learning Representations Leader of T2I Group
LiD-FL: Towards List-Decodable Federated Learning
AAAI 2025 AAAI Conference on Artificial Intelligence
Journal Articles
CRISP: A Causal Relationships-Guided Deep Learning Framework for Advanced ICU Mortality Prediction
BMC MDM 2025 BMC Medical Informatics and Decision Making
Preprints
SkillGen: Verified Inference-Time Agent Skill Synthesis
arXiv 2026 Preprint
Confidence Laundering in Agent Systems: Why Uncertainty Needs a Latent Carrier
arXiv 2026 Preprint
UXBench: Measuring the Actionability of LLM-Generated UX Critiques
arXiv 2026 Preprint
Education
University of Notre Dame
Ph.D.Ph.D. in Computer Science
Supervisor: Prof. Fanny Ye
Research Areas
Trustworthy AI · LLM Alignment · Foundation Models
Sep 2026 — June 2031 Incoming
Sichuan University
B.Eng.B.Eng. in Cyber Science and Engineering
Sep 2021 — June 2026
Research Experience
Visiting Research Student
Advisor: Prof. Xiangliang Zhang
Trustworthy Generative Models · LLM Alignment · Agent Evaluation