Autoregressive models for molecular graph generation typically operate on flattened sequences of atoms and bonds, discarding the rich multi-scale structure inherent to molecules. We introduce MOSAIC (Multi-scale Organization via Structural Abstraction In Composition), a framework that lifts autoregressive generation from flat token walks to compositional, hierarchy-aware sequences. MOSAIC provides a unified three-stage pipeline: (1) hierarchical coarsening that recursively groups atoms into motif-like clusters using graph-theoretic methods (spectral clustering, hierarchical agglomerative clustering, and motif-aware variants), (2) structured tokenization that serializes the resulting multi-level hierarchy into sequences that explicitly encode parent-child relationships, partition boundaries, and edge connectivity at every level, and (3) autoregressive generation with a standard Transformer decoder that learns to produce these structured sequences.