I'm a postdoctoral researcher at Meta FAIR, with a PhD in Psychology from Princeton.
I build computational models of how intelligent systems form beliefs and make decisions, and use those insights to evaluate and improve AI. My research connects cognitive science and machine learning; at Meta FAIR, I am extending this work from evaluation toward post-training and multi-agent systems.
Website · Google Scholar · Email
- LLM vigilance — Using Bayesian models to evaluate how language models respond to communicative motives and incentives. NeurIPS 2025, equal contribution. Paper
- Muse Spark Safety & Preparedness Report — Contributor to the sycophancy evaluation. Meta, 2026. Report
- Paths to Persistence — Modeling why people maintain beliefs in the face of disagreement. Psychological Review, 2025. Paper
- HorizonBench — Evaluating how language models track changing user preferences across long conversations. Collaborative work, 2026 preprint. Paper · Code
- Brittlebench — Measuring how benchmark conclusions change under variations in prompt wording. Collaborative work, 2026 preprint. Paper
More research, writing, and background at keremoktar.com.
