SWM: LLM-powered model predicts social belief shifts from events

A team of researchers (Haofei Yu, Yining Zhao, Guanyu Lin, Jiaxuan You) has introduced the Social World Model (SWM), a novel framework that leverages large language models to predict how social beliefs shift in response to events like policy changes or scientific breakthroughs. SWM automatically learns state-transition functions for social beliefs by mining temporal patterns in social data and optimizing an evidence lower bound (ELBO), eliminating the need for expensive human annotations or census data. This approach taps into LLMs' commonsense and social intelligence to model belief dynamics at scale.

To evaluate SWM, the authors built SWM-bench, a benchmark derived from real-world prediction markets (Kalshi and Polymarket), containing over 12,000 data points across politics, finance, and cryptocurrency. Experimental results show that SWM significantly outperforms state-of-the-art time-series foundation models on Kalshi data while achieving competitive results on Polymarket data. The framework also provides interpretable insights into the underlying mechanisms of belief change, making it a powerful tool for social scientists, analysts, and policymakers to understand public opinion evolution without manual labeling.