How much do LLMs learn from negative examples?