Reproduce “Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal 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
Zeeshan Ahmed, Yang Qin, Hanqing Huang
SJR separates multimedia content understanding from policy judgment by turning ads and landing-page evidence into structured text for a separate policy model. It trains the models iteratively: policy correctness rewards useful summaries, while text-space augmentation creates synthetic examples. On an internal misleading-advertisement task, full SJR reaches 79.25% legitimate-ad F1, versus 64.14% for zero-shot chain-of-thought; this does not establish transfer to other policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.