Can Feedback Enhance Semantic Grounding in Large Vision-Language Models?