LLM-Driven Graph Chain of Thought Reasoning for Agentic Response to Indoor Fire Risk

ACL ARR 2026 January Submission5068 Authors

Keywords: Language Models, Knowledge Graphs, Multi-agent Collaboration, RRetrieval-augmented Reasoning, obotics for Fire Safety

Abstract: Understanding fire risk and response planning in complex indoor environments requires reliable reasoning over incomplete perceptions and heterogeneous domain knowledge. Although large language models (LLMs) have demonstrated strong reasoning capabilities, enabling them to perform structured, interpretable, and adaptive reasoning over dynamic fire scenes remains a significant challenge. In this work, we present Insights on Graph (IOG), an LLM-driven multi-agent reasoning framework that performs adaptive graph-based chain-of-thought reasoning for early fire-risk detection and response recommendation. IOG constructs a fire-domain knowledge graph by integrating fire safety regulations and robotic emergency response protocols and orchestrates three collaborative agents to reason over the KG. Through the incremental construction of dynamic subgraphs aligned with scene observations, IOG enables traceable reasoning, context-aware decision-making, and adaptation to environmental changes. Extensive simulations and real-world experiments have demonstrated that IOG significantly improves fire-risk understanding and response planning compared to existing baselines, highlighting its effectiveness and robustness in complex, safety-critical environments. Our code is publicly available at https://anonymous.4open.science/r/IOG.

Paper Type: Long

Research Area: Language Models

Research Area Keywords: applications, chain-of-thought, LLM/AI agents, retrieval-augmented generation

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data resources

Languages Studied: English

Submission Number: 5068