Website · Scholar · X · LinkedIn · Email · CV
Past · Present






I am a third-year PhD student at the University of Notre Dame, advised by Prof. Xiangliang Zhang in the MINE Lab.
Previously at Microsoft Research, MIT-IBM Watson AI Lab, IBM Research, and Tsinghua KEG.
I build self-improvement humane and humanized intelligence: foundation models and agentic systems aligned with human values, informed by human cognition, and supported by reliable infrastructure.
I study how to align foundation models and agentic systems with human values, from individual behavior to collective dynamics. My work uncovers emergent social intelligence risks that agent-level safeguards alone cannot prevent (Emergent Social Intelligence Risks Nature Machine Intelligence). I also benchmark trustworthiness (TrustLLM ICML'24, TrustGen ICLR'26), train honest and helpful models (HonestLLM NeurIPS'24), examine biases in LLM-as-a-judge (Justice or Prejudice? ICLR'25), and prioritize safety in alignment (SPA AAAI'26 Oral, Capability-Oriented Risk ICML'26).
I probe how models reason and perceive, deciding when to use tools (MetaTool ICLR'24), adapting reasoning style to the task (AdaReasoner NeurIPS'25 Spotlight), exposing flaws in social character simulation (TrustSim COLM'25), diagnosing failure modes at scale (ProbeLLM ICML'26), ranking models by question difficulty (RankLLM ICLR'26), and probing visual aesthetic judgment (VAB COLM'26). I also explore emotionally-adaptive generative AI (EmoNest NeurIPS'25 Creative, NARRA-Gym NeurIPS'26, Omelune Mac app).
I build the toolchain that makes the above possible: frameworks for unified synthetic data (DataGen ICLR'25, ChemOrch NeurIPS'25), synthetic interaction data for scalable personalization (PersonaGym KDD'26), foundational guardrails for safer agentic systems (Safiron ICLR'26), a new guardian model paradigm (Guardian-as-an-Advisor ACL'26 Findings), agent infrastructure that turns compute into verifiable improvement and has broken 120 world records in mathematics (Épi Bake AI), and infrastructure for recursive self-improvement (RSI Arena Live at COLM'26).




