MAGR-FI: LLM-Based Multi-Agent Game-Theoretic Reasoning Framework for Fire Investigation

Abstract

Fire investigation relies heavily on expert experience, rendering it susceptible to cognitive biases and inefficient in processing heterogeneous information. This study proposes an LLM-Based Multi-Agent Game-Theoretic Reasoning Framework for Fire Investigation (MAGR-FI), which integrates the Fire Triangle model from combustion science with the Man-Machine-Environment-Management (MMEM) framework from system safety theory to establish a structured mechanism for hypothesis generation and adversarial validation. The framework employs three specialized agents—Proponent, Skeptic, and Arbiter—engaged in a two-phase dynamic game, enabling interpretable reasoning from unstructured investigative texts to high-confidence causal conclusions. Evaluated on a test set of 1,051 real-world cases, MAGR-FI significantly outperforms baseline large language models, improving average scores from 6.95 to 8.51 and achieving a 120% performance gain on complex cases, thereby providing a reliable, transparent, and auditable intelligent decision-support tool for fire investigation and transforming retrospective causal analysis into actionable knowledge assets for crisis learning and proactive prevention.