An Efficient Dual-Agent Framework for Generating and Evaluating Synthetic Aviation Safety Reports Using Large Language Models

This study proposes a dual-agent framework leveraging Large Language Models (LLMs) to generate and evaluate synthetic aviation safety reports, addressing challenges like class imbalance and domain-specific requirements. By assigning separate generation and evaluation roles to specialized agents operating concurrently on distinct GPUs, the framework optimizes resource utilization and enhances content quality. Results demonstrate improved processing efficiency compared to single-GPU systems and enhanced content quality through iterative refinement. This research highlights the effectiveness of dual-agent LLM systems in producing high-quality synthetic data, reducing reliance on human evaluations and advancing NLP applications in safety-critical domains.