Can LLMs support altruistic decisions in fire evacuations? Analysis via an FDS-ABM model with pedestrian assistance

Building fire incidents often result in severe casualties and substantial property losses. Understanding pedestrian behavior during fire evacuations is therefore essential for building design and emergency management. However, the application of Large Language Models (LLMs) to simulate pedestrian decision-making in fire evacuations remains largely unexplored. In this study, we propose a coupled Fire Dynamics Simulator (FDS)–Agent-Based Model (ABM) framework in which LLMs generate dynamic and individualized pedestrian assisting decisions based on both physiological and social factors under realistic fire scenarios. GPT-4 is adopted as the primary decision model, while Qwen3-max is introduced to evaluate the consistency and generalizability of LLM-generated decision patterns. An interpretable machine-learning framework combining XGBoost and SHAP is further employed to classify assisting behavior and quantify the contributions of different influencing factors. The proposed framework is compared with an evacuation model without assisting behavior, and sensitivity analyses are conducted for the severe-injury threshold, smoke-sensitivity coefficient, evacuation response time, and leader proportion. The results indicate that high physiological stress substantially reduces the likelihood of assisting severely injured pedestrians, whereas stronger social influence promotes assisting behavior, consistent with previous psychological findings. Compared with the evacuation model without assisting behavior, the proposed framework improves evacuation efficiency and reduces overall life-safety risks. Although GPT-4 and Qwen3-max exhibit different decision preferences and assisting rates, both models show similar relationships between the influencing factors and helping behavior. Sensitivity analyses further demonstrate that the proposed framework remains stable and consistent under different severe-injury threshold settings, whereas the smoke-sensitivity coefficient has only a limited effect on evacuation performance and decision tendencies. In addition, longer response times reduce both evacuation efficiency and assisting rates, whereas a higher proportion of leaders promotes cooperation and evacuation efficiency but may increase individual risk. Overall, the findings represent computational patterns generated by the adopted LLMs rather than validated human assisting behavior. The proposed framework provides a computational approach for integrating LLM-driven behavioral decision-making into fire evacuation simulations, offering methodological support for future building safety design and emergency management.