Web search is supposed to keep large language models current, yet they sometimes answer with outdated information even when the right evidence is in front of them. A paper posted on arXiv, titled "Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence," takes up this puzzle directly.
The authors say they "construct temporal facts whose values" change over time, presumably to create test cases where the retrieved evidence contradicts the model's internal knowledge. The abstract, however, is cut off before presenting any results, so the specific reasons for the failure are not described in the available text.
What can be said from the source is limited to the motivation and approach: the paper identifies a real gap in how search-augmented models use retrieved content, and it proposes a controlled way to study that gap. Readers looking for concrete conclusions will need to consult the full paper.