How can we reliably monitor the reasoning and behavior of advanced AI systems to detect failures and potential misalignment? I investigate the faithfulness and limitations of chain-of-thought monitoring, particularly in detecting deception, scheming, reward hacking, and other misaligned behaviors. My goal is to develop robust monitoring methods that enable scalable oversight of increasingly capable and autonomous AI systems.

What can a model's internal representations reveal about how it reasons, makes decisions, and develops unexpected behaviors? I am interested in white-box interpretability methods that provide insights beyond observable outputs and reasoning traces. My research explores how internal model signals can complement behavioral evaluations, improve the detection of hidden misalignment, and support more reliable safety monitoring.

How do safety risks emerge when autonomous AI agents pursue goals, make decisions, and interact with one another? I study deceptive and misaligned behaviors in agentic systems, including collusion, miscoordination, conflicts of interest, and emergent coordination. I am particularly interested in how these behaviors arise through agent interactions and communication, including latent-space communication, and how they can be detected, monitored, and mitigated.

How do language and distribution shifts affect the reasoning, reliability, and safety of large language models? I investigate multilingual reasoning, cross-lingual generalization, and the robustness of safety monitoring across typologically diverse languages. My goal is to understand and reduce disparities in model performance and monitorability, ensuring that AI systems remain reliable and trustworthy across linguistic settings.