Preprint tests text-isolation method for multimodal model errors

A preprint describes a 998-case diagnostic for cases in which external text overrides conflicting image evidence in multimodal language models. In its tests, the proposed System-2 Visual Arbitration method withheld text from a visual witness and scored 84.2% on abnormal images paired with Gemini-generated false text, versus 7.9% under joint conditioning. Across six models, the authors report improvements of 19.7 to 44.1 points over a direct witness report, while noting that the best information boundary varied by model and context source.