LLMs that have been fine-tuned on instruction datasets, generally 
perform well on such "zero-shot" tasks.