Pretraining Unlearning with Topic-Conditioned Neuron Masks
Can training make information easier to remove later? Experiments with document- and topic-based neuron masks in a 398-million-parameter language model.
Autonomous agents,
evaluation, and oversight
I’m a graduate student at MIT developing technical methods to evaluate, monitor, and shape autonomous agents in multi-agent and human–AI systems.
I’m interested in how agents coordinate, respond to incentives, and behave under oversight—and how we can design these systems to preserve human agency in institutional decision-making.
Recent projects
2025–2026BAITBENCH: Measuring Agent Reward Hacking
When an AI research agent can take a shortcut to a better score, does it? We built tasks that distinguish apparent success from performance on hidden test data.
Oversight Is Not Compliance
A study of tacit collusion among AI pricing agents: what oversight changes, what it misses, and why an agent’s explanation can differ from its behavior.
More research & writing
2026O*NetBench
Building a predictive model of AI capability scores for individual occupations, grounded in worker evaluations of occupational tasks.
Crashing Waves vs. Rising Tides
Estimated task-duration doubling times by worker-rated quality:
- Rating ≥7 · Minimally sufficient
- 2.7 months
- Rating ≥8 · Average-worker quality or better
- 2.4 months
- Rating 9 · Superior quality
- 1.6 months
Each quality threshold defines success: a response either meets that rating or falls below it. The fitted model estimates how quickly task duration doubles while holding the probability of meeting the threshold constant. Duration is measured in human work time; the estimates are sensitive to the fitted relationship between ratings and duration.
AI, scientific progress, and the digital divide
How unequal access to AI could shape who benefits from its use in science.
A little more about me
At MIT, I study technology and policy and electrical engineering and computer science. My research spans reward hacking in AI research agents, oversight of interacting agents, and worker evaluations of AI capabilities.
Before MIT, I was a machine learning engineer at Coursera, the first engineer at Dreambound, and a project lead at the University of Pittsburgh’s AI and Social Good Lab. I studied economics and computer science at Swarthmore.
Across these projects, I keep coming back to a practical question: when a system appears to be doing the right thing, what evidence should convince us?
Get in touch
I’d be glad to compare notes on research.