Reproduce “As Student Rollouts Continue, Visual Backdoors Keep Growing” 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
Zane Fu, Yanzhao Zheng, Jinze Liu, Ziting Qin, Qianshan Wei, Jihuai Zhu, Yipeng Yu, Baohua Dong and 2 more
The authors investigate whether on-policy distillation (OPD) transfers visual backdoors from a vision-language teacher to a student. The student generates responses online but receives teacher token-probability guidance, which can reinforce refusal behavior through a proposed feedback loop. In tested settings, OPD backdoor success sometimes rose abruptly after remaining low; thresholds vary by model and task, and zero-injection mechanism remains unclear.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.