Reproduce “Perturb, Repair, Verify: Self-Play Vision-Language Verifiers for Compositional Understanding” 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

Zijun Huang, Yingjun Du, Cees G. M. Snoek

PROVE trains a shared vision-language model to detect caption-image mismatches by introducing small errors, diagnosing them, repairing captions, and rechecking repairs. On Qwen3.5-2B, its four-group compositional average rises from 83.22 to 86.85. An independent audit finds verifier accuracy improves from 75.8% to 83.6%, but does not rule out self-verification bias or establish gains across broader tasks.

Summary by alphaXiv.

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