README.md

ViVerBench

ViVerBench evaluates whether multimodal models can verify if generated visual outputs satisfy prompt-level constraints.

• Paper: Generative Universal Verifier as Multimodal Meta-Reasoner
• Dataset: comin/ViVerBench

Overview

• 3,594 examples across 16 task categories
• Binary verification target (true / false)
• Inputs can contain multiple images (1, 2, or 8)

Usage

python -m lmms_eval \
  --model <model_name> \
  --tasks viverbench \
  --batch_size 1 \
  --limit 8