General-purpose LLMs as Models of Human Driver Behavior: The Case of Simplified Merging

In three linesStudy comparing OpenAI o3 and Google Gemini 2.5 Pro as models of human driving behavior in a simplified merging scenario. LLMs reproduce intermittent operational control and tactical dependencies, but fail to capture responses to dynamic velocity cues. Prompt ablations reveal model-specific inductive biases that do not transfer across LLMs.

Summary generated by Claude — human-verified