Multimodal vs. Text-Only Models in Adversarial Fact-Checking Resilience

This report synthesises findings from 10 peer-reviewed papers addressing the following research question: How do multimodal models (e.g., BLIP-2, Flamingo) compare to text-only models in resilience against adversarial perturbations in fact-checking tasks when evaluated using accuracy metrics like Exact. 9 claims were extracted from source literature; 9 were independently verified against retrieved documents. An automated multi-reviewer quality assessment produced a score of 9.1/10. This report is a machine-generated literature synthesis and does not constitute original research.

Research goal: How do multimodal models (e.g., BLIP-2, Flamingo) compare to text-only models in resilience against adversarial perturbations in fact-checking tasks when evaluated using accuracy metrics like Exact Match or F1-score?

Autonomous literature synthesis. Automated review score: 9.1/10. Full text and citation available at Assignee Research.