Yue Zhao

Research Summary

Research Interests: My research centers on building reliable, safe, and scalable AI systems, with a focus on understanding and mitigating failure modes in modern foundation models and agentic systems. I organize my work into two tightly connected tiers: (1) advancing the scientific foundations of safety and robustness in AI, and (2) translating these foundations into system-level evaluation frameworks and high-impact applications.

Tier 1: Foundations of Reliable & Safe AI
I study why and how modern AI systems fail under distribution shift, uncertainty, and strategic pressure, and develop methods to make their behavior more predictable and reliable. This tier integrates two complementary threads:
LLM & Agent Safety: Analyzing and mitigating failure modes in large language models and agentic systems, including hallucinations, jailbreaks, privacy leakage, model extraction, and multi-agent instability.
Robustness & Failure Detection: Developing algorithms and benchmarks for identifying abnormal or unreliable behavior, grounded in robustness, out-of-distribution generalization, and anomaly detection.

Keywords: LLM Safety, Robustness, Agents, Hallucination Mitigation, Jailbreak Detection, OOD Generalization, Failure Analysis

Tier 2: System-Level Evaluation & Scientific/Societal Impact

I adopt a system-oriented perspective to evaluate, stress-test, and deploy reliable AI in realistic settings, and apply these methods to domains where failures carry high cost. This tier emphasizes two directions:
Evaluation & Benchmarking: Designing scalable evaluation frameworks, benchmarks, and workflows that probe model and agent behavior under realistic and adversarial conditions.
AI for Science & Society: Applying reliable foundation models to climate and weather forecasting, healthcare and biomedicine, and political or social decision-making.

Keywords: Evaluation, Benchmarking, System-Level Analysis, AI for Science, Scientific Foundation Models, Climate & Weather Modeling, AI for Healthcare