Social Intelligence in Minds and Machines

I build computational models of how intelligent systems form beliefs, make decisions, and navigate social interactions, then use those insights to evaluate and improve AI.

At Meta FAIR, I am extending this work from evaluation toward post-training and multi-agent systems.

Selected Work

Muse Spark Safety & Preparedness Report Meta · 2026 · Sycophancy evaluation contributor

The report evaluates Meta's LLM, Muse Spark, across safety and behavioral dimensions; I contributed to the sycophancy benchmark, which measures inappropriate agreement and pushback.

Report · arXiv

Are Large Language Models Sensitive to the Motives Behind Communication? NeurIPS 2025 · Equal contribution · Wu, Liu, Oktar, Sumers & Griffiths

We use a Bayesian model of social learning as a normative benchmark for LLM vigilance: model behavior tracks it in controlled tasks, while intention-and-incentive steering improves the match in realistic sponsored content.

Paper

How Beliefs Persist Amid Controversy: The Paths to Persistence Model Psychological Review · 2025 · Oktar & Lombrozo

We unite theories across the social sciences into four concrete reasons people resist disagreement—distrusting dissenters, treating issues as subjective, anticipating costs of changing, or lacking the cognitive resources to update—and test them across five preregistered studies.

Paper

Full list: CV · Google Scholar.

Recent Research

HorizonBench Can models track a user as their preferences change over months? DetailsClose

Preprint · 2026 · Li, Paranjape, Oktar et al.

HorizonBench generates six-month conversations from structured mental-state graphs, with ground-truth provenance for every preference change; across 25 frontier models, most score at or below chance, with updated-state tracking the central bottleneck. Paper · Code · Data

Brittlebench Do benchmark conclusions survive harmless changes in wording? DetailsClose

Preprint · 2026 · Romanou, Ibrahim, Ross, Oktar et al.

Brittlebench introduces a variance-decomposition framework that separates task difficulty from prompt sensitivity; semantics-preserving perturbations change model rankings in 63% of cases and can explain up to half of performance variance. Paper

Intuitive Theories of Truth How do people decide what can be true, is true, and should be asserted? DetailsClose

Trends in Cognitive Sciences · In press · Equal contribution · Oktar, Handley-Miner et al.

Intuitive Theories of Truth organizes everyday truth judgments around aptness, judgment, and assertion, explaining how people can disagree about truth even when they share the same evidence. Paper

About

I was born and raised in Istanbul (🧿); studied economics and cognitive science at Pomona College, CA (☀️); completed my PhD in Psychology at Princeton, NJ (❄️); and now research social cognition in AI systems at Meta FAIR in Seattle, WA (☔).

My research has won CogSci’s Marr Prize for best paper and SPP’s Poster Prize; Princeton’s Center for Human Values supported my PhD through a Rockefeller Prize Fellowship. Beyond research, I co-designed and co-taught Psychology of Justice at Edna Mahan, a women’s prison in New Jersey, in 2024. I have also found and tamed the Oscar Mayer Wienermobile.

Contact

Feel free to contact me at oktar[dot]research[at]gmail.com with regards to research / collaboration / mentorship /… - I love talking about science.

Here is an anonymous feedback form.

Writing

On the Reasonable Effectiveness of Judgment: Why human judgment matters to mathematics and model training.

Why False Premises Birth Truths: An intuition for material implication.

More writing: Power analysis guide · Grad school application guide