The rapid advancement of autonomous vehicle technology has created an urgent need for scalable testing methodologies to ensure safety and reliability. Traditional manual scenario generation is time-consuming and resource-intensive. This research investigates Large Language Models (LLMs) for automated scenario generation, evaluating base model performance, prompt engineering, and fine-tuning techniques. In addition to evaluation, the thesis develops a systematic methodology for fine-tuning LLMs tailored to scenario generation in Advanced Driver Assistance Systems (ADAS), providing a structured framework for future research and industrial adoption. Experiments were conducted using three LLMs (CodeLlama-13B-instruct, Mistral-7B-instruct, and DeepSeek-Coder-6.7B-instruct) across seven scenarios of progressive complexity. Each model was tested under base, prompt-engineered, and fine-tuned configurations, with outputs validated through the Esmini simulator and assessed against multiple validity criteria, including executable code generation, scenario adherence, and geometric validity. OpenAI o3 was additionally evaluated in base and prompt engineering modes to benchmarks for broader LLM capabilities. The results demonstrate that prompt engineering consistently outperformed both base and fine-tuned variants in all scenarios and models. Contrary to the initial hypothesis, fine-tuning delivered limited improvements, constrained by dataset size and computational resources. A strong inverse relationship was observed between scenario complexity and generation success rates. Overall, findings indicate that prompt engineering is the most practical and effective enhancement strategy for LLM-based scenario generation in resource-constrained environments. At the same time, the proposed fine-tuning methodology establishes a foundation for scaling, refinement, and domain-specific adaptation in AI-assisted testing for autonomous vehicle development.