Reproduce “Visual Self-Fulfilling Alignment: Shaping Safety-Oriented Personas via Threat-Related Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qishun Yang, Shu Yang, Lijie Hu, Di Wang
VSFA trains vision-language models on neutral questions about synthetic threat-themed images, without explicit refusal instructions, to test whether visual exposure changes safety behavior. On Qwen3-VL-8B, it averages 14.18% attack success across three jailbreak benchmarks, versus 38.77% with no defense, while benign-query refusal is 2.62%. Results cover selected benchmarks and do not establish the proposed mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.