Title: Improving Operational Safety Through Agentic Dialogue for Hazard Identification Analysis

Abstract: Ensuring operational safety in critical sectors—including autonomous systems, industrial process control, and other safety-critical domains—requires robust hazard identification capabilities. Although large language models (LLMs) have demonstrated potential in automating safety analysis, relying on single-turn, monolithic inference proves fragile. Such approaches fail to replicate the iterative self-correction, deliberation, and contextual refinement that human safety engineers employ. To address this, we present HAZDIAL, a framework designed to assess whether structured agentic dialogue—characterized by multi-agent, multi-turn interactions—can enhance the quality of NLP-based hazard identification compared to single-pass baselines. Our study systematically contrasts two dialogue modalities: adversarial debate and constructive discussion, while introducing an algorithm-based method for optimizing agentic interaction. We benchmark all configurations against a curated golden dataset, employing standard classification metrics such as accuracy, precision, recall, and F1 score, alongside new metrics tailored to dialogue evaluation. This research contributes to the convergence of dialogue systems, multi-agent reasoning, and AI safety, offering empirical support for the efficacy of dialogue-driven hazard analysis.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC