When an AI has two answers, it may need to say neither
A research paper tests a method for choosing between two conflicting answers: a retrieved source and an AI model’s stored knowledge. It can also hold back instead of guessing. The reported test gains partly require knowing the true share of reliable source passages. That share may be unknown in real use, so the scores are not a deployment promise.
What the source reports
Minh Pham Van and Cong Tran study a familiar problem: a new source says one thing, while an AI model seems to know another. Their research paper tests a way to choose between the two or to abstain. The test uses a controlled set of altered knowledge links. The authors report gains over fixed rules, and the digest also notes results from another planned test. The catch is central: part of the result depends on giving the method the true share of correct source passages. A team using it in real work may not know that number. This preprint does not show that every conflict can be settled safely without a person checking the sources.
Original source
- Title
- Don't Pick a Winner: Knowledge Conflicts Need a Prior, Not a Rule
- Author
- Minh Pham Van, Cong Tran
- Publication
- arXiv cs.AI
- Date
- Monday, October 26, 2026