Research

Language use is a fundamentally social activity, but classical work suggested the connection between language and social cognition is slow, effortful, and error-prone. Our work with Paula Rubio-Fernandez has been challenging this view. We have found that people make real-time social inferences as speech unfolds in real time [2020, PNAS], and people can extract high-resolution mental state information from minimal linguistic cues, matching the precision of mental-state inferences from physical action [2021, Science Advances]. On the production side, we have found that even the basic linguistic choices reveal social reasoning operating at the word level, like when to say this versus that [2024, PNAS], when to use adjectives [2020, JEP:General], or how much to say to jog someone's memory [2025, CogSci]. Consistent with this, we our computational models where social cognition is embedded in language outperform standard models and explain cross-linguistic variation, including why Spanish speakers use fewer adjectives than English speakers [2022, Psych Review]. We recently published a theory paper that reviews all these findings (along with other work), to argue that everyday communication is sustained by continuous word-by-word mind-tracking, rather than secondary language-independent pragmatic reasoning [2025, TiCS].

What computations and representations allow humans to reason about other minds? Our work has helped advanced the view that people infer mental states by inverting a generative model of rational action [2017, NHB]. We have proposed that this is accomplished through a good-enough simplified model of other minds that treats others as if they are utility maximizes who trade off costs and rewards when forming plans [2020, Cog. Psychology]. This approximation gets other people’s knowledge and intent broadly right, making their behavior interpretable and predictable. Empirically, we have found that this system permeates social reasoning from early in childhood, helping us infer competence [2015, Cognition], preferences [2017, JEP:General], and knowledge [2021, Child Dev.], as well as guiding how we understand language [2019, Child Dev.], communicate with others [2019, NHB] and make moral judgments [2015, Psych. Science]. You can find a review paper here [2016, TiCS], and a modeling tutorial here.

How pervasive is social reasoning? Our work suggests that people use social reasoning and Theory of Mind not only when observing agents, but even when looking at objects and physical scenes. People can detect indirect of agency in their environment [2021, JOV] and use it to reconstruct happened [2022, JEP:General], and even young children can do this [2021, CogSci]. We have also shown that this capacity allows people to use the physical world itself as a channel for communication, allowing people across cultures to indirectly communicate through objects [2023, Cognition]. This work suggests that social reasoning is not only a system for extracting mental states from people’s observable action, but also from how people shape their environments [2024, Current Dir. in Psych. Science; 2025, BBS Commentary].

How can people’s capacity to model minds guide the design of socially competent and safe AI? Our computational models of human Theory of Mind offers a path for building machines that understand behavior through structured mental-state representations rather than pattern recognition alone [2019, Curr. Op. in Beh. Sci. Special issue on AI]. In parallel, we have also shown how meta-cognitive architectures, where systems reason about themselves, can give vision models the ability to self-discover their own hallucinations and learn when to not trust what they see [2020, NeurIPS SVRHM; 2024, UAI] (We also think this is a key capacity for human intelligence; [2024, BBS Commentary]. By contrast, we recently found that large language models have sophisticated social mimicry but lack the signatures of abstract model-based representations of other minds [2025, AAAI AI-ToM]. These projects shape how we think about broader problems in alignment, thinking about how architectures inspired by biological intelligence can constraint the design of robust and cooperative AI [2025, arXiv], and how robust social intelligence needs to scale beyond individual agents to institutions [2025, arXiv].

One puzzle in mental-state inference is that reasoning about other minds is computationally costly, but people can easily navigate complex social environments. In collaboration with Dr. Dunham’s lab, we are developing the idea that humans have an institutional stance: a system that co-evolved with Theory of Mind, to interpret behavior in terms of structures systems of roles rather than representations of mental states [in press, BBS]. Precursors of the institutional stance explain how other species coordinate without rich models of other minds. But the institutional stance is unique to humans in its generative capacity, allowing us to create arbitrary institutional structures that led to an explosion of role-based systems in modern societies (and some aspects are embedded in language itself!). In recent work we’ve found that role representations allow people to form instant expectations about what others will do, what they know, and who else could serve the same function [2024, CogSci]. We are also modeling how people infer social structures from sparse observations [2022, CogSci], and how we separate role-based sources of behavior from genuine mental states [2025, CogSci]. We have a lot of active work in this area, so more coming soon!

How does social cognition develop and what does this tell us about the mind more broadly? Our work suggests that children can reason about others’ goals and desires from very early in life [see ‘Computational Basis of Theory of Mind’ section above; 2015, Cognition; 2017, JEP:General; 2021, Child Dev.; 2019, Child Dev.; 2019, NHB; 2015, Psych. Science; 2016, TiCS], but knowing how to infer beliefs and knowledge takes longer. We have found that children as young as four years old already have a structured causal model that connects knowledge to action [Aboody et al., Cognition], but they initially rely on coarse heuristics to infer mental states and don’t show probabilistic reasoning characteristic of adult cognition until around age six [2025a, Child. Dev; 2025b, Child. Dev]. Interestingly, adult belief inference is graded and quantitative [2025, Cognition; 2017, NHB], but not when these inferences requires representing broad hypothesis spaces [2023, Cognitive Science]. We have recently proposed a broader framework that suggests that developing social cognition involves learning to build restricted-scope models of other minds—simplified but useful representations that balance efficiency and expressive power, which we hypothesize emerge, in large part, through parent-child conversations [2025, Ann. Rev. of Dev. Psych.].