Multi-Agent Prompting with In-Context Feedback for Cross-Modal Urban Event Chain Prediction

Urban sensor networks detect events such as rain, accidents, and public gatherings, yet existing systems cannot predict how these events propagate across different sensor domains. For example, heavy rainfall may reduce traffic speed, increase citizen complaints, and elevate accident risk, creating a chain of related events that spans multiple data sources. This study investigates whether domain-specialized large language model (LLM) agents with in-context feedback can predict such cross-modal event chains without model training.

Our framework consists of four prompt-based agents representing traffic, weather, social behavior, and public safety. Each agent receives only domain-specific sensor information and performs independent reasoning. A coordination module then combines their outputs into a time-ordered event chain with confidence estimates. To improve performance, prediction errors from each round are incorporated into subsequent prompts, enabling self-correction through in-context learning without fine-tuning or weight updates.

We evaluate the framework using real highway loop detector data aligned with meteorological records and selected storm events for qualitative analysis. Results show that multi-agent prompting with feedback achieves lower prediction error than both single-agent prompting and traditional baselines. Qualitative evaluation demonstrates improved causal coherence and event coverage. Our findings highlight the value of domain decomposition and feedback-driven reasoning for intelligent analysis of heterogeneous urban sensor data.