Perturb, Repair, Verify: Self-Play Vision-Language Verifiers for Compositional Understanding
PROVE (Perturb-and-Repair Optimization for Vision-Language Evaluation) is introduced, a self-play framework that trains verifiers through a closed perturb-verify-repair-verify loop and improves fine-grained compositional verification across model scales and families.
Zi-Jun Huang, Ying-Jun Du, Cees G. M. Snoek
· 0 citations