Building Social World Model with Large Language Models

ICLR 2026 Conference Submission14749 Authors

Keywords: world model, social learning, large language model

Abstract: Understanding and predicting how social beliefs evolve in response to events, ranging from policy changes to scientific breakthroughs, remains a fundamental challenge in social science research. Given that Large Language Models (LLMs) have demonstrated commonsense knowledge and social intelligence, a natural question arises: Can LLMs be used to model the dynamics of social beliefs following social events? Addressing this problem can deepen our understanding of community dynamics and inform better decision-making in the real world. In this work, we introduce the Social World Model (SocialWM) concept, a general framework designed to capture how social beliefs evolve in response to major events. SocialWM learns state-transition functions for social beliefs by mining temporal patterns in social data and optimizing the evidence lower bound, without the need for explicit causal annotations that link events to belief shifts or expensive census data. To evaluate SocialWM's effectiveness in predicting social belief transitions, we introduce a benchmark, SocialWM-bench, derived from real-world Polymarket data. SocialWM-bench includes over 300,000 data samples for social belief prediction tasks spanning diverse domains such as politics, sports, cryptocurrency, and elections. Our experimental results show that SocialWM significantly outperforms traditional time-series models, achieving a 21.56% reduction in RMSE while offering interpretable insights into the underlying mechanisms of social belief dynamics.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 14749