This dissertation dives into a pressing issue in autonomous driving: the lack of reliable ways to test world models, especially when traditional motion planning falls short in unpredictable settings like messy road layouts or sudden obstacles. I’ve been motivated by the absence of consistent benchmarks, which makes it tough to fairly compare different models. To tackle this, I’m proposing a new framework to evaluate two unique approaches: Terra, a deterministic autoregressive model, and DrivingDojo, a stochastic diffusion-based one. My goal is to create a solid method using controlled experiments, with standardized inputs and simulated environments, to see how well these models predict driving actions. I’ll focus on setting clear evaluation standards and using real-world simulation tools, like CARLA, to get a deeper understanding of the context. I hope this work will shed light on the strengths and weaknesses of deterministic versus stochastic methods, laying the groundwork for their use in safety-critical systems. I also plan to pinpoint areas for improvement, such as tackling long-term prediction challenges, and suggest directions for future research. I’d welcome evaluators to look at how reproducible my framework is, how deeply I compare the models, and how my findings might shape the future of autonomous driving, all in line with my aim to push forward standards for testing generative models.