Evidence Integration in Large Language Models

An outside answer can influence a model differently depending on what that model was already likely to believe. In Evidence Integration in Large Language Models, researchers report that models more readily absorb candidate answers resembling their own likely answers, including familiar kinds of mistakes. The same supplied evidence could help weaker models while hurting stronger ones. They also report cases where models incorporated candidates despite internally recognizing their invalidity. These findings suggest that evaluating an evidence source in isolation misses part of the problem: its effect also depends on the receiving model and how it incorporates the information.