Navigating Rifts in Human-LLM Grounding: Study and Benchmark