Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling | Read Paper on Bytez