Abstract
Long-standing research suggests that social inferences are captured by a few latent dimensions (e.g., warmth and competence). Others argue that social inferences are more complex but lack sufficient empirical support. Here, we conducted two pre-registered studies to test the high-dimensional properties of social inferences. To maximize generalizability, we computationally sampled diverse naturalistic videos and recruited U.S. representative participants (Study 1, N = 1598). Participants freely described people in videos using their own words. Cross-validation identified 25 latent dimensions which explained only 15% of the variance in the data. Alternatively, a sparse network model representing the unique correlations between inferences better represented the data. The network models informed the dynamics of naturalistic inferences, revealing how different inferences co-occurred and how they unfolded over time from concrete to abstract (Study 1). The network models also indicated cultural differences in how one inference was related to another between samples (Study 2, Asian N = 651, European N = 792). Together, these findings show that the high-dimensional network approach provides an alternative model for understanding the mental representation of social inferences in naturalistic contexts, which provides new insights into the dynamics and diversities of social inferences beyond the static, universal structure found with traditional low-dimensional latent-construct approaches.