About
I'm a fifth-year PhD student at the University of Pennsylvania advised by Professors Mayur Naik and Eric Wong. I spent the past two summers at AWS AI (Fundamental Research Team) working with Matthew Trager and Stefano Soatto on uncertainty quantification and experience-guided reasoning for agents. My research is supported by the NSF Graduate Research Fellowship Program.
My research aims to make AI systems reason reliably and behave as intended. I study the interface between foundation models and programs: which parts of a task should be left to a model's semantic flexibility, which should be handed to executable structure, and where to look for evidence that the system actually did what its output implies. My work is organized around a hierarchy of static, generated, and agentic programs, moving from fixed solvers and symbolic interfaces, to instance-specific programs and inference-time strategies, to agents whose behavior must be audited across traces. Most recently, I've been using the records agents leave behind to make them safer, more reliable, and cheaper to run.
Research Summary
The Interface Between Foundation Models and Programs
Static programs
Translate unstructured inputs into artifacts that existing solvers can execute.
Generated programs
Generate and check instance-specific programs and adaptive strategies.
Agentic programs
Audit systems whose behavior plays out across tools, state, and traces.
Recent News
-
Released Detecting Safety Violations Across Many Agent Traces, finding widespread cheating on popular agent benchmarks. See the blog post and Meerkat code.
-
Presented PIPS at NeurIPS in San Diego, along with three posters including two spotlights at the Mechanistic Interpretability workshop.
-
Founded the Penn Agentic Lab, leading a team of undergraduates working on software reliability via Agentic Testing.
-
Once Upon an Input: Reasoning via Per-Instance Program Synthesis (PIPS) accepted to NeurIPS 2025.
-
Attended ACL in Vienna to present Towards Style Alignment in Cross-Cultural Translation.
PublicationsFeatured
Pre-Prints
Conference Papers
Workshop Papers
Student Mentoring
- Arthur Wayne (Applying to PhD programs)
Teaching
- TA for CIS 547, Program Analysis (University of Pennsylvania, Fall 2023)
- TA for CIS 500, Software Foundations (University of Pennsylvania, Fall 2022)
- Tutor, Tau Beta Pi (University of California, Los Angeles, 2019)