CAST-Eval: A Domain-Specific Benchmark for Large Language Models in Civil Aviation Safety

Abstract

Verbatim from Semantic Scholar. Not paraphrased, not summarized.

In this paper, we present CAST-Eval, a novel, comprehensive and domain-specific benchmark designed to assess the knowledge and reasoning capabilities of large language models (LLMs) in the civil aviation domain. CAST-Eval covers critical professional knowledge areas, including airlines, airports, air traffic control, and other key sectors, with tasks ranging from basic university-level content to advanced qualification exams issued by the CAAC. To validate its effectiveness, we evaluated major Chinese LLMs, both closed-source and open-source, and publicly released the results. Our experiments show that in the civil aviation domain, closed-source commercial models generally outperformed smaller-scale open-source models, with most achieving scores above 60. We also found that while general-purpose models performed well in areas such as meteorology and navigation, specialized fields like flight operations and dispatch remain particularly challenging for general AI models. Furthermore, newly released 70B-level open-source models exhibited performance comparable to or even surpassing some closed-source models. CAST-Eval can serves as a versatile evaluation tool for fine-tuned language models across other industry-specific applications, supporting improvements in AI model training and optimization.

Authors

• Dongxiao Qiao
• Xianhui Tian
• Yubin Xu
• Jie Yang
• Le Yang
• Xuhui Wang

Keywords

• Engineering
• Computer Science