Reproduce “Metric-Aligned Group Relative Policy Optimization for Explainable Deepfake Detection” 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

Luka Dragar, Peter Peer, Vitomir Štruc, Borut Batagelj

The authors present a deepfake detection pipeline that separates authenticity decisions from explanations. A vision ensemble predicts real or fake, and a vision-language model explains the image conditioned on that fixed label; GRPO trains explanations against sequence-level rewards. On 10,000 matched test images, GRPO raised complex-explanation BERTScore-F1 from 0.686 to 0.712. This measures reference similarity, not whether claims are true.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.