Researchers built an AI system whose entire job is making self-driving car simulators crash.
The framework, called Teach-to-Crash, pairs two large language models in a closed loop. A high-reasoning Teacher LLM monitors how a driving simulation is going and steps in only when collision rates and time-to-collision metrics stall out. A simpler, low-reasoning Student LLM does the actual work, generating test scenarios as strict JSON that the CARLA simulator can execute directly. In CARLA trials across two setups with different ego-vehicle speed policies, the system produced crashes in 90.79% of its runs, with a mean time-to-collision of just 18.31 seconds.
Self-driving companies rely on simulation because real-world crash data is rare and expensive to collect, so automated scenario generation is one of the unglamorous bottlenecks in getting autonomous vehicles road-ready. Teach-to-Crash's selling point is that its crashes are not just frequent but structurally diverse, scoring 0.547 on the paper's diversity metric, and it topped the field on an avoidability-based usefulness score too, at 60.04%, meaning the crashes it finds look like ones a better-driving system could plausibly have dodged rather than physics-defying edge cases.
A rival method called PAFOT actually found more crashes on average, 179.44 versus 136.21, but with far higher variance. Consistency, not raw crash count, is what Teach-to-Crash is actually selling, and whether any of it survives contact with real streets remains untested.