Building Social World Models with Large Language Models

The paper introduces the Social World Model (SWM), a framework utilizing large language models (LLMs) to model the evolution of social beliefs in response to significant events. SWM learns state-transition functions from temporal social data without requiring explicit annotations, and is evaluated using the SWM-bench benchmark, which includes over 12,000 data points from prediction markets like Kalshi and Polymarket. Experimental results indicate that SWM outperforms time-series foundation models, providing state-of-the-art performance on Kalshi data and competitive results on Polymarket, while also offering insights into social belief dynamics, which is crucial for practitioners analyzing societal impacts of events.