matt hardy

matt at roundtable dot ai ▪
twitter ▪ google scholar ▪ blog

I'm currently working on Roundtable (ycs23) with Mayank Agrawal. Roundtable does behavioral cybersecurity for fraud prevention. We look at signals like mouse movements and keystroke patterns to identify bots and bad actors without disrupting the user experience.

Previously, I completed my PhD in Tom Griffiths' Computational Cognitive Science Lab at Princeton. My research focused on using machine learning and behavioral experiments to improve network design, choice architectures, and decision-making in general.

During grad school, I spent a summer as an intern at Microsoft Research where I worked with Dan Goldstein and Jake Hofman to develop tools for out-of-sample forecasting. I also spent some time building electric scatter, a website devoted to interactive data visualizations.

Before my PhD, I majored in statistics and economics (and minored in math) at the University of Toronto, and spent two summers at MIT working in econometrics with Whitney Newey and Jerry Hausman. I also interned in data science at Via, and spent my earlier years working as a busser, paperboy, lifeguard, farmhand, and assembly line worker.

selected papers

Embers of autoregression show how large language models are shaped by the problem they are trained to solve
Tom McCoy, Shunyu Yao, Dan Friedman, Matt Hardy, and Tom Griffiths (2024). PNAS.

When a language model is optimized for reasoning, does it still show embers of autoregression? An analysis of OpenAI o1
Tom McCoy, Shunyu Yao, Dan Friedman, Matt Hardy, and Tom Griffiths (2024). arXiv preprint.

Improving out-of-population prediction: The complementary effects of model assistance and judgmental bootstrapping
Matt Hardy, Sam Zhang, Jessica Hullman, Jake Hofman, and Dan Goldstein (2024). International Journal of Forecasting.

AI-generated visuals of car-free American cities help increase support for sustainable transport policies
Rachit Dubey, Matt Hardy, Tom Griffiths, and Rahul Bhui (2024). Nature Sustainability.

Resampling reduces bias amplification in experimental social networks
Matt Hardy, Bill Thompson, P.M. Krafft, and Tom Griffiths (2023). Nature Human Behaviour.

Optimal nudging for cognitively bounded agents: A framework for modeling, predicting, and controlling the effects of choice architectures
Fred Callaway, Matt Hardy, and Tom Griffiths (2023). Psychological Review.

How do humans overcome individual computational limitations by working together?
Natalia Vélez, Brian Christian, Matt Hardy, Bill Thompson, and Tom Griffiths (2023). Cognitive Science.

Overcoming individual limitations through distributed computation: Rational information accumulation in multi-generational populations
Matt Hardy, P.M. Krafft, Bill Thompson, and Tom Griffiths (2022). Topics in Cognitive Science.