Evaluating Large Language Models For Requirements Engineering And Verification: A Comparative Analysis Of Gpt-5 And Gemini 3 Pro

Large Language Models (LLMs) are increasingly viewed as viable tools for Systems Engineering, yet empirical data regarding their effectiveness in generating engineering-quality requirements remains limited. This thesis compares the performance of GPT-5 and Gemini 3 Pro in generating system functional requirements for the Miner Guardian Drone, a prototype UAV designed for minefield remediation. Both models were provided with identical prompts and tasked with producing 15 functional requirements. Two independent raters evaluated the outputs using a nine-dimensional quality rubric aligned with IEEE 29148 and INCOSE guidelines to ensure rigorous assessment. The difference between the models was analyzed using descriptive statistics and a Single Factor Analysis of Variance (ANOVA). The results indicate that Gemini 3 Pro achieved a higher mean quality score and demonstrated greater consistency with lower variability compared to GPT-5. While GPT-5 produced high-quality outputs, it exhibited a wider score range and a higher frequency of lower-quality outliers. However, the ANOVA results suggest that the performance difference was not statistically significant at the 95% confidence level, implying that while Gemini showed superior descriptive metrics, the variance in the current sample size precludes a definitive rejection of the null hypothesis. This thesis offers an evidence-based, methodological comparison of state-of-the-art LLMs for requirements engineering and identifies critical future research opportunities, including domain-tuned LLMs and automated traceability support.