New LLM framework simulates vaccine opinions with multi-round AI conversations

A new study from researchers Bo Zhang and Na Jiang presents a large language model-driven agent-based modeling framework designed to simulate how vaccine opinions evolve in social networks. The system leverages Alibaba's Qwen3-8B LLM to drive agent decision-making, enabling multi-round conversations where agents update their stances based on memory and prompt diversity modules. The framework models agents with heterogeneous demographic profiles and social connections, then runs simulations to observe macro-level opinion dynamics. By toggling specific cognitive modules on and off, the researchers can isolate how memory (recalling past interactions) versus prompt diversity (exposure to varied perspectives) influence vaccine acceptance over time.

The simulation results reveal a stark contrast: memory and prompt diversity modules produce opposite effects on emergent opinion polarization. Memory-driven agents tended to reinforce existing beliefs, while diversity-exposed agents showed more fluid opinion shifts. Notably, the framework successfully reproduces the non-linear behavior patterns of social influence documented in earlier empirical research, validating its realism. This achievement meets Level 3 validation criteria for agent-based models, meaning the model's outputs align with real-world data patterns at multiple scales. The study demonstrates that LLM-powered agents can go beyond simple rule-based simulations to capture complex cognitive processes, offering a powerful tool for public health policymakers to test messaging strategies and predict vaccine uptake under different social conditions.