A challenging benchmark for evaluating multi-evidence integration capabilities of vision-language models
🎉 HERBench has been accepted to CVPR 2026!
🆕 New: Lite-v2 config. We released a refreshed lite_v2 version of the
Lite split (1,971 questions / 68 videos) in which 9 of the 12 tasks were
regenerated and went through additional manual refinement for higher
quality, while TSO, SVA… See the full description on the dataset page:
https://huggingface.co/datasets/DanBenAmi/HERBench.
