Do LLMs Make More Mistakes If They Do Not Believe the Input Data?

Models did not strongly override supplied information merely because it conflicted with their stored knowledge in this study's human-reviewed sample. Researchers asked models to turn structured Czech and Slovak facts, altered facts and fictional information into text in four languages. They then assessed whether the generated text stayed faithful to those inputs. The paper reports only a weak conflict effect.

The choice of evaluator changed the apparent result. An automated judge that aligned well with human assessments found only a small difference between factual and altered inputs; a less suitable judge could exaggerate it. The study therefore makes evaluator validation part of determining whether a model ignored the supplied context.

“Contrary to our expectations, we observe only a weak context-memory conflict on the human-annotated sample.”

“We also find that a suboptimal choice of LLM judge would lead to overestimating the strength of the context-memory conflict.”