Generating 3D scenes that support rich interactions are of growing interest due to its importance in embodied ai and spatial intelligence. However, it remains difficult due to issues like data scarcity and complex spatial reasoning. In this talk, we present our exploration of this direction, which consists of two novel approaches: DirectLayout, a top-down LLM framework that generates open-vocabulary layouts from text using explicit spatial planning; and Pair2Scene, a bottom-up method that learns local pairwise support and functional relations to assemble scenes hierarchically. We examine how LLM-driven semantic flexibility complements learned local geometric priors. By analyzing these divergent strategies, we explore how integrating global language reasoning with physics-aware local rules can yield 3D scenes that are both instruction-aligned and physically plausible beyond training distributions.