CRoPS, a training-free hallucination mitigation framework for vision-language models, accepted at TMLR.