ThrowBench: Benchmarking LLMs by Predicting Runtime Exceptions