Invited Talks

Robots that interact with real people in real-world environments must adapt to diverse human preferences, tasks, and constraints, yet achieving this remains a significant challenge. While current approaches rely on collecting massive datasets to build generalizable models, this strategy is expensive, slow, and often brittle. Instead, robots must learn to adapt efficiently, learning "a lot from a little" human input. In this talk, I argue that robots don’t just need more data—they need better data. To learn a lot from a little, we need to rethink how we engage with human input and the structures we build around it. By designing information rich yet effortless human inputs, amplifying sparse data through simulation and LLM priors, and structuring the robot learning problem with strong behavioral abstractions, we can create robots that learn efficiently and align with human needs in diverse environments. This shift from data quantity to data quality represents a key step toward human-aligned robot learning for real-world applications.

Effective models of human attitudes and behavior can empower applications ranging from immersive environments to social policy simulation. However, traditional simulations have struggled to capture the complexity and contingency of human behavior. I argue that modern artificial intelligence models allow us to re-examine this limitation. I make my case through computational software agents that simulate human attitudes and behavior. I discuss how we used this approach, which we call generative agents, to model a representative sample of 1,000 Americans and replicate their attitudes and behavior 85% as well as they replicate themselves two weeks later. Extending my line of argument, I explore how modeling human behavior and attitudes can help us design more effective online social spaces, understand the societal disagreement underlying modern AI models, and better embed societal values into our algorithms.

Machine learning methods provide increasingly powerful tools for generating predictions about human behavior. However, simply using off the shelf methods to generate predictions potentially misses opportunities to benefit from and contribute to the psychological literature. In this talk I will discuss three ways in which theory and data can interact through machine learning: using theories to pretrain machine learning models; using theories to constrain machine learning models; and using unconstrained machine learning models to critique explanatory theories. I will illustrate these cases with examples from the study of human decision-making. 

In this talk I will document two kinds of misunderstanding: people misunderstanding algorithms; and algorithms misunderstanding people. The empirical work is done on LLMs and supervised learning. In both cases, I show how these misunderstandings are consequential; and have implications for how models should be trained differently. 

Recent work has suggested substituting human participation and labor in research and development--e.g., in surveys, experiments, and interviews--with synthetic research data generated by Large Language Models (LLMs). I will provide an overview of our work in which we conducted interviews with 19 qualitative researchers to understand how they define and interact with LLM personas as proxies for research participants, and their perspectives on this paradigm shift. I will contextualize this work within the broader set of initiatives to study the use of GenAI on the ground by end users---an important layer of risk evaluation referred to as "field-testing" or "human uplift studies" in AI policy documents. I will conclude with an agenda for future work in this space.

Early work in AI alignment relied on restrictive assumptions about human behaviour to make progress even in simple 1:1 settings with a single operator. This talk addresses two key considerations for developing more realistic models of human preferences for alignment today. In Part I, we challenge the assumption that values and preferences are universal or acontextual through examining interpersonal dilemmas - what happens when we disagree with one another? I'll introduce the PRISM Alignment Dataset as a key new resource that contextualizes preference ratings across diverse human groups with detailed sociodemographic data. In Part II, we challenge the assumption that values and preferences are stable or exogenous by exploring intrapersonal dilemmas - what happens when we disagree with ourselves? I'll introduce ongoing research on anthropomorphism in human-AI interaction, examining how revealed preferences often conflict with stated preferences, especially regarding AI systems' social capabilities and in longitudinal interactions.