Robust evaluation methodology is essential for guiding the training of reliable systems. It not only measures system performance, but also seeks to understand system behavior and, most importantly, continuously refine alignment rubrics so they reflect rational human intentions and can be distilled into the training process.

• Static Evaluation: Design granular yet scalable metrics that capture richer task-specific properties and better match real design goals.
• Human-in-the-Loop Evaluation: (1) Build representative feedback loops under limited budgets; (2) monitor bidirectional risks in human-AI interaction (e.g., humans: over-reliance, manipulation; AI: over-alignment, sycophancy) and develop collaboration paradigms that preserve rationality on both sides.
• Interactive (Agentic) Evaluation: (1) Study the strengths and limits of foundational structures that emerge in agentic systems; (2) test the robustness of cooperative behaviors under adversarial conditions.

I'm especially interested in applying these ideas to AI safety and healthcare.