Multimodal Models Pre-Trained on Visual Genome Enhance Adversarial Robustness in Visual Reasoning

This report synthesises findings from 10 peer-reviewed papers addressing the following research question: Do multimodal models pre-trained on Visual Genome exhibit improved robustness against adversarial visual perturbations in visual reasoning tasks compared to models trained on sparse image-text pairs. 6 claims were extracted from source literature; 6 were independently verified against retrieved documents. An automated multi-reviewer quality assessment produced a score of 7.9/10. This report is a machine-generated literature synthesis and does not constitute original research.

Research goal: Do multimodal models pre-trained on Visual Genome exhibit improved robustness against adversarial visual perturbations in visual reasoning tasks compared to models trained on sparse image-text pairs?

Autonomous literature synthesis. Automated review score: 7.9/10. Full text and citation available at Assignee Research.