In collaboration with the Social Learning Lab at Boston University's Wheelock College, we are studying how parents provide various kinds of explanations to in conversations with their children about how change takes place. 

Specifically, we are utilizing advances on large language models to gain insight into the structure and content of parent-child conversations in a story-reading context. Parents and children in Iran engaged in a conversation about a number of stories; parents were asked to explain to their children how something happens in a story context. For example, a poor family becomes rich; what happens for this change to take place? In other words, what "causes" this change or what is the "causal mechanism?" 

By using large language models to uncover the structure and content of explanations about causal mechanisms, we hope to provide an example of how these models can improve developmental research and minimize error.

In collaboration with mohammad Atari, an assistant professor in UMass Amherst, this research builds on Kelly and Keil's 1985 work on metamorphoses and conceptual structure ( "The more things change...: Metamorphoses and conceptual structure") by studying how things change or transform in folk stories, both within individual cultures and across cultures in general. We look at what turns into what - for example, when humans turn into animals or objects come to life. Using large collections of folk stories, such as the Thompson Motif-Index, we study these changes in detail. Our goal is to identify patterns of transformation that are unique to each culture's stories, as well as patterns that appear commonly across many cultures. We track both how often certain types of changes occur and what these changes mean in their cultural context. A key focus of our research is understanding the social function of these transformations - whether they serve to promote cooperation, reinforce moral values, or fulfill other social purposes within their cultural contexts. By looking at folk story collections from different regions, we analyze how these transformation patterns correlate with various regional characteristics, including religiosity indices and population demographics. This approach allows us to explore potential relationships between transformation patterns in folklore, their social functions, and broader societal factors across different geographical areas.

In collaboration with Reza Aghajani, a researcher in Google Deep Mind, this research investigates how Large Language Models (LLMs) process and exhibit essentialist thinking when encountering both novel and familiar social groups. In our first study, we assess different components of psychological essentialism by presenting LLMs with various group-related queries, examining how they handle both unfamiliar and familiar social categories. Our second study focuses specifically on how different types of language presentation - generic statements ("Xs are Y"), specific statements ("This member of Group X is Y"), and quantified statements (ex: "Many members of Group X are Y") - influence essentialist thinking about novel groups in LLMs. Through these systematic investigations, we aim to understand how language models process and generalize information about social categories, and how the format of language presentation affects their tendency toward essentialist reasoning. 

The ability to detect and reason about others’ mental states is central to social life, shaping processes such as moral reasoning and prosocial behavior. While this capacity emerges early in development, less is known about the factors that influence how children attribute mental states to others. In this study, we test whether linguistic cues—specifically generic language—can experimentally reduce children’s attribution of mental states to members of a novel social group. We predict this effect because generic language encourages thinking about groups as homogeneous “kinds,” which may conflict with children’s tendency to associate minds with individual identities.

In addition to this main prediction, we explore whether the impact of generic language is amplified when paired with biological content, which has been shown to promote essentialist thinking. Since essentialism is linked to stereotyping, dehumanization, and biased judgments about others’ minds, we investigate whether it serves as a mechanism underlying reductions in mental state attribution.