The world is being quietly rearranged by people who write very long documents.


The title they went with General-purpose LLMs as Models of Human Driver Behavior: The Case of Simplified Merging Noisy translates that to

AI can drive like a human, but only if you tell it what to look for


Researchers tested if large language models could act as human drivers in a simplified merging scenario. The AI models could mimic some human driving behaviors, but failed to react consistently to changing speeds.
Automated vehicle developers need reliable models of human driving behavior to test their systems safely. If AI can act as a human driver in simulations, it could speed up testing and reduce costs. But this paper shows that AI models are still unreliable, and their performance depends heavily on how they are prompted, which means they are not yet ready for real-world safety evaluations.
Watch for future research that shows AI models consistently reacting to dynamic speed changes in driving simulations, or for new methods that make prompt engineering transferable across different AI models.

If you insist
Read the original →