Thomas proposes that humans are equipped from infancy with an intuitive theory of social relationships, which supports recognizing distinct types of dyadic relationships and their strengths. In this commentary, we draw out the implications of this proposal for computational models of human social reasoning.

Traditional models of social cognition (or more specifically, Theory of Mind) focus on how people reason about individual agents’ hidden mental states (such as their beliefs and desires), to explain and predict those agents’ observable behaviors (e.g., Ho, Saxe, & Cushman, Reference Ho, Saxe and Cushman2022). The fundamental unit of analysis here is the individual agent, whose actions are interpreted under the principle of rational action (Jara-Ettinger et al., Reference Jara-Ettinger, Gweon, Schulz and Tenenbaum2016). Such models have successfully captured many aspects of human social reasoning, including goal inference, cooperative communication, and false belief (Baker, Saxe, & Tenenbaum, Reference Baker, Saxe and Tenenbaum2009; Goodman & Frank, Reference Goodman and Frank2016; Baker et al., Reference Baker, Jara-Ettinger, Saxe and Tenenbaum2017).

Such models have also been used to characterize social relationships in terms of one agent’s preferences about another agent’s preferences. In particular, cooperative and competitive interactions between agents have been modeled through the concept of adopted utility: the idea that individuals incorporate others’ goals into their own utility functions (Powell, Reference Powell2022). In a cooperative interaction, agents act to help each other reach their goals, and in a competitive interaction, agents act to prevent each other from reaching their goals (Ullman et al., Reference Ullman, Baker, Macindoe, Evans, Goodman and Tenenbaum2009; Kleiman-Weiner et al., Reference Kleiman-Weiner, Ho, Austerweil, Littman and Tenenbaum2016). Such models have been tested primarily in grid-world paradigms, where agents navigate to reach goals in physical locations. Adopted utility models provide a good fit for observers’ judgments of whether one agent is acting to promote or hinder the utilities of another agent (Ullman et al., Reference Ullman, Baker, Macindoe, Evans, Goodman and Tenenbaum2009; Hamlin et al., Reference Kiley Hamlin, Ullman, Tenenbaum, Goodman and Baker2013).

Thomas’s proposal implies that computational models of human social understanding require a qualitatively different approach. Thomas highlights that social relationships go beyond cooperative and competitive interactions defined by adopted utilities. The distinctions between communal sharing, authority ranking, and equality matching relationships cannot be reduced to whether one agent adopts another’s utility: in all of these relationship types, agents can act to promote each other’s goals and preferences. Yet, as Thomas shows, the relationship types differ dramatically in the expectations they generate and behavioral predictions they support.

Even more dramatically, Thomas suggests that the basic representation that people use to understand behavior in relationships is not a representation of one agent with beliefs and desires about another agent. Instead, the key cognitive primitive that people reason over, when reasoning about social interactions, may represent the relationship itself. Humans may rely on a small set of structured priors, such as a set of cognitive primitives specifically designed to recognize relational structures, each corresponding to a certain set of internal dynamics governing intuitive theories of different kinds of social interactions (e.g., Pietraszewski, Reference Pietraszewski2022). These priors can guide learning and inference in new social situations: people can use probabilistic inference to recover how likely each type of relationship is – as well as dimensions important to the specific kind of relationship, like its strength – given observed social interactions.

Incorporating relationship primitives in Bayesian models of probabilistic reasoning can contribute to testable theories regarding the development of relational social cognition. These models can not only capture how each social interaction can be described in terms of multiple kinds of relationships, but also how new social information can revise how people think about certain relationships in the first place. Hierarchical models and Bayesian nonparametrics (Kemp, Perfors, & Tenenbaum, Reference Kemp, Perfors and Tenenbaum2007; Gershman & Blei, Reference Gershman and Blei2012) can capture how infants and children flexibly infer the number and nature of relationship categories (and sub-categories) from a combination of innate priors and social experience, as well as accommodate variability across different social and cultural contexts.

This relationship-focused framework explains how minimal social information can lead to rich generalizations and predictions. Knowing that a social interaction – such as sharing saliva – strongly fits a particular relationship type, in this case a communal sharing relationship, can activate expectations about subsequent interactions without necessarily invoking the computationally costly process of explicit reasoning about individual mental states (Thomas et al., Reference Thomas, Woo, Nettle, Spelke and Saxe2022). Similarly, knowing whether a social relationship is equal or hierarchical leads to strong expectations about whether generous acts will alternate or repeat (Chen & Saxe, Reference Chen and Saxe2023). In fact, Theory of Mind reasoning about individual agents’ beliefs and desires may only become necessary when relational expectations are violated, or in special cases like agents with conflicting beliefs or desires about the same relationships. In relational contexts, Theory of Mind might be reserved for resolving ambiguities or social anomalies, rather than forming the core basis for routine social predictions (Bohl, Reference Bohl2015).

If relationships, and not just individual minds, are the primary units of social understanding, then there are many open questions. One question is what precisely constitutes these relational primitives in the first place. Are they simply about sets of actions that are consistent with different kinds of relationships, and if so, is learning about different kinds of relationships the same as a category learning mechanism? How is learning about relationships different from learning about other social structures that govern different kinds of social norms? How much of early social reasoning can be understood in terms of relational primitives, versus active mental-state reasoning?

Another open question is how we might operationalize the notion of relationships in computational models. What would it take to extend existing paradigms that focus on the individual agent to account for relationships? One approach, proposed by Jara-Ettinger & Dunham (Reference Jara-Ettinger and Dunham2024), is to reify “roles” and their associated “norms” as additional entities in a causal Bayesian theory. Could this approach be applied to model infants’ intuitions about dyadic relationships? And if so, how might we build such models concretely – how might we extend domain-specific programming languages like memo (Chandra et al., Reference Chandra, Chen, Tenenbaum and Ragan-Kelley2025) to capture the semantics of relationship reasoning?

In summary, we propose that computational models of social cognition should incorporate relational representations as primitives. Traditional inverse-planning models, which focus solely on individual utilities, overlook the critical relational structures that organize social behavior and can support fast and efficient social learning.