POPE

Researchers propose EviAnchor, a training-free framework that reduces hallucinations in large vision-language models by preserving and reactivating visual evidence during generation, achieving consist…