G-SHARE Framework Improves Human-Factor Event Diagnosis in Nuclear Plants.

This research introduces G-SHARE, a novel framework designed to enhance the diagnosis of human-factor events, particularly within safety-critical environments like nuclear power plants. The system translates established nine-step diagnostic guidelines into a structured, multi-stage reasoning process. G-SHARE's methodology involves extracting evidence from narrative reports, performing stepwise diagnostic reasoning, and then repairing any inconsistencies post-hoc. This structured approach ensures explicit use of evidence, generates intermediate rationales, and validates diagnostic outputs logically. Evaluations using real-world human-factor event reports from the Chinese nuclear industry demonstrated that G-SHARE significantly surpasses both one-shot large language model prompting and traditional machine learning baselines in accuracy and F1-score, highlighting the critical role of structured reasoning and consistency enforcement.