John Kang

Current WSI VQA benchmarks are so contaminated by data leakage that they misrepresent the true capabilities of vision-language models, with peak accuracies aligning with the most compromised datasets.