Humanlike representation of social networks in large language models
Year: 2026
Type: preprint
Abstract: Although large language models (LLMs) have made great progress in their general reasoning capabilities, reasoning about social relationships has remained a challenge. In contrast, humans are naturally skilled at reasoning about how people are connected to one another, even in large, complex social networks. Recent research in cognitive science suggests that a key component of human social intelligence is the ability to represent social networks using multistep abstraction, which integrates knowledge of both direct and indirect social connections, and therefore supports inferences about network members’ social proximity to one another. Here, we test whether LLMs represent social networks using similar computational principles. In Study 1, we converted an existing human social navigation task into text and recorded LLM responses to the same problems humans were tasked with solving. LLMs’ behaviors closely mirrored those of human participants, and were best-explained by Katz communicability, a model of multistep abstraction that captures how humans represent and navigate social networks. In Study 2, we tested whether these multistep representations persisted in a task where they offered no benefit—and were in fact detrimental to task accuracy—and found that LLMs continued to show a multistep signature even in this context. These findings suggest that multistep abstraction may be a core property of how LLMs represent relational information, closely paralleling accounts of human cognitive maps.
Institution: Brown University
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7204796480
Cites:
Cited by:
Related to: