Reproduce “RuleSafe-VL: Evaluating Rule-Conditioned Decision Reasoning in Vision-Language Content Moderation” 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
Zhifeng Lu, Dianyuan Wang, Yuhu Shang, Zhenbo Xu
RuleSafe-VL introduces a benchmark that tests whether multimodal models follow platform moderation policies beyond predicting labels. It evaluates rule activation, typed relations among rules, evidence sufficiency, and outcomes after context is supplied. Gemini-3.1 reaches 64.8 Macro-F1 on relation recovery, and Qwen3-VL-4B reaches 64.5 on decision-state prediction. Using public policies, it covers three content families; it does not establish deployment performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.