Answer-Then-Check Robustness Against Multimodal Jailbreak Attacks in Vision-Language Models

This report synthesises findings from 16 peer-reviewed papers addressing the following research question: Does the robustness gain from Answer-Then-Check against adversarial prompts in Tulu 3 and Deepseek R1 generalize to multimodal jailbreak attacks evaluated on recent vision-language benchmarks. 9 claims were extracted from source literature; 9 were independently verified against retrieved documents. An automated multi-reviewer quality assessment produced a score of 8.7/10. This report is a machine-generated literature synthesis and does not constitute original research.

Research goal: Does the robustness gain from Answer-Then-Check against adversarial prompts in Tulu 3 and Deepseek R1 generalize to multimodal jailbreak attacks evaluated on recent vision-language benchmarks?

Autonomous literature synthesis. Automated review score: 8.7/10. Full text and citation available at Assignee Research.