FIRM: Resolving Missing Boundary Conditions in Fragment-Based Molecular Generation

Fragment-based models reduce atom and bond predictions using reusable, chemically meaningful substructures, but fragment adjacency omits attachment sites and bond types that constrain compatible templates and mappings. Existing methods typically encode these details within concrete generative states or resolve them during decoding. FIRM, the Fragment Interface Realization Model, instead generates a complete fragment-identity-free typed interface before conditional realization. Interface Flow generates attachment ports and inter-fragment bond types; Conditional Realization Flow selects compatible templates and mappings for typed assembly. The interface reduces realization entropy by 7.29 nats on MOSES and 4.60 nats on NPGen. Reveal-order controls favor interface-first training for interface-to-realization prediction and completion under direct decoding. Across four benchmarks, FIRM attains 100.0% observed validity under the full generation pipeline and closely matches reference distributions of molecular features and domain-relevant properties. These results support treating the fragment interface as a distinct generative layer between coarse fragment topology and atom-level realization.