Simulation study demonstrates model- and language-dependent aggressiveness shifts in autonomous driving LLMs, indicating the need for careful prompt and model selection.
Key Points
To evaluate the reasoning patterns, behavioral tendencies, and social biases of large language models when making high-level driving decisions in autonomous vehicle contexts.
Tested three prominent large language models (GPT, DeepSeek, and LLaMA) across 1,500 contextual prompt variations featuring scenarios ordered by aggressiveness.
Evaluated multilingual prompt conditions (English, Chinese, and French) using an Ordered Logit Model alongside thematic analysis of reasoning tendencies.
GPT exhibited more conservative driving choices, whereas DeepSeek and LLaMA made more assertive decisions, particularly during vehicle-to-vehicle interactions.
Prompt language altered decision profiles, with French prompts prompting the most assertive driving responses, followed by Chinese, and English being the least assertive.
All models shifted toward protective, lower-aggression driving when vulnerable road users were introduced to the driving context.