Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers
A temporal alignment failure, stale-document poisoning, is identified, in which outdated evidence makes a model wrong despite answering correctly without retrieval, and the same internal components also support broader comparison tasks, suggesting that temporal applicability can recruit a general reasoning mechanism us...
Md Shamim Ahmed, Lukas Galke Poech, R. Röttger
· 0 citations