Don't Pick a Winner: Knowledge Conflicts Need a Prior, Not a Rule.

arXiv:2610.36169v1 Announce Type: new Abstract: A retrieval-augmented language model answers from two sources, the retrieved context and its own knowledge, and the two disagree whenever the context is incomplete or wrong. Existing methods settle the disagreeme