G-SHARE framework uses structured reasoning to outperform baselines in nuclear event diagnosis

Human-factor event diagnosis in nuclear power plants relies heavily on expert interpretation of narrative reports against formal guidelines. Existing data-driven or one-shot large language model approaches often lack structured reasoning, fail to align with diagnostic guidelines, and can produce logically inconsistent conclusions. To address these shortcomings, researchers introduced G-SHARE, a guideline-based structured reasoning framework that directly operationalizes the CNNP nine-step human-factor event diagnosis guideline into a multi-stage pipeline.

G-SHARE comprises three stages: evidence extraction from narrative reports, stepwise diagnostic reasoning following the guideline, and post-hoc consistency repair to validate logical coherence. The team constructed a dataset from real Chinese nuclear industry events with a gold-standard subset annotated by domain experts. Results show G-SHARE substantially outperforms both one-shot prompting and traditional machine learning baselines in overall accuracy and macro-F1. Ablation experiments highlight that structured reasoning and consistency enforcement are critical, especially under weak prompting conditions. The work demonstrates that transforming expert diagnostic guidelines into auditable reasoning workflows provides a practical pathway for intelligent human-factor analysis in safety-critical industries.