# Preprint tests text-isolation method for multimodal model errors

_Wednesday, September 2, 2026 at 12:00 AM EDT · Science, AI · Latest · Tier 2 — Notable_

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.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2609.00067)

---
Canonical: https://techandbusiness.org/newswire/-W_fvsUyJ0172dYvKAHTph
Retrieved: 2026-09-02T11:47:56.631Z
Publisher: Tech & Business (techandbusiness.org)