Tag
This paper reveals that while large language models appear robust to task-irrelevant context at the aggregate level, their predictions can flip on individual examples, with performance degrading on some and improving on others, highlighting tail risks that aggregate accuracy conceals.