ES-MAS simulator models US-China opinion shifts using 14K news articles

Understanding how public opinion shifts in response to geopolitical events is critical for risk assessment, but existing LLM-based multi-agent simulators rely on static rules and fixed datasets, failing to capture the dynamic, event-driven nature of real-world opinion evolution. To address this, researchers from the University of Science and Technology of China and JD.com propose ES-MAS (Event-Steered Multi-Agent Simulator), a framework where significant events and daily news continuously drive opinion evolution through agent interactions. The system introduces a Dual-Stream Data Integration Engine (DSDIE) that aligns simulations with historical timelines via macro-level events while enabling personalized information exposure based on individual agent profiles. Additionally, a News-Driven Dynamic Interaction (NDDI) module adaptively groups agents with shared news interests into localized interaction contexts, facilitating bottom-up consensus formation and preventing information cocoons.

The backbone of ES-MAS is the CURE dataset (China-U.S. Relation Evolution), which spans 20 quarters from 2021 to 2025 and includes 258 major events and over 14,000 daily news articles—providing a comprehensive temporal foundation for modeling. Experimental results on CURE show that ES-MAS substantially outperforms existing simulators in reproducing real-world historical opinion trends. By combining event-driven steering, personalized exposure, and dynamic group interactions, the framework offers a scalable and effective approach for modeling dynamic opinion evolution. This work has implications for geopolitical risk modeling, public sentiment analysis, and policy simulation, enabling more realistic, event-responsive forecasts of international relations.