Analyzing Multimodal Interaction Strategies for LLM-Assisted Manipulation of 3D Scenes