Agentic Root Cause Analysis through Evidence-Grounded Reasoning
AgentRCA uses a digital twin and an LLM agent to diagnose industrial faults without fault-specific training.
The paper frames root cause analysis as an evidence-gathering loop rather than a supervised classification problem. Its agent tests competing hypotheses against statistical signals from a model of normal system behavior. In evaluations on a multiphase-flow facility and a chemical plant, it performed competitively with supervised baselines. The authors emphasize that the system also produces reasoning traces linking symptoms to physical causes. ArXiv · AI/CL/LG's note
The paper frames root cause analysis as an evidence-gathering loop rather than a supervised classification problem. Its agent tests competing hypotheses against statistical signals from a model of normal system behavior. In evaluations on a multiphase-flow facility and a chemical plant, it performed competitively with supervised baselines. The authors emphasize that the system also produces reasoning traces linking symptoms to physical causes. ArXiv · AI/CL/LG's note
score 4