Large language models are often asked to answer questions using retrieved or user-supplied context. But that context may be factually correct while still being unhelpful—for example, describing a different entity, time period, or location than the one the question is about. A new arXiv paper, MisleadQA, focuses on this failure mode.
The paper introduces a benchmark or diagnostic approach for probing how models respond when context is true but out of scope. Rather than testing factual accuracy alone, MisleadQA targets the subtler problem of relevance: whether a model can recognize that the provided material does not actually address the question at hand.
The significance is practical. Systems that rely on retrieval-augmented generation can easily pull in plausible but irrelevant passages. MisleadQA offers a way to measure and understand these failures, pointing toward the need for models that not only know facts but also know when those facts apply.