Crystal-scaffold: deterministic scaffold compilation for text-conditioned CIF generation

Abstract Large language models can generate crystal structures as text, but direct generation of complete, valid CIFs remains difficult, since CIFs combine strict syntax, symmetry constraints, deterministic closure relations, and multiple valid serializations of the same structure. Here, we introduce Crystal-Scaffold, a canonicalized compiled-scaffold framework for text-conditioned full-CIF generation. Crystal-Scaffold converts raw CIFs into deterministic canonical targets, uses a multi-task parsing warm-up to supervise file-level and target-level symmetry descriptors, stoichiometric closure, and lattice--site structural constraints, and predicts a compact scaffold-core representation from composition and space-group conditions. The remaining closure fields are deterministically compiled into a scaffold-full conditioning context, while the final complete CIF is still generated autoregressively. We further introduce Crystal-Scaffold-HBoN, a hierarchical Best-of-N decoding extension that samples scaffold and CIF candidates and reranks them with a verifier. On the MP-20 benchmark, Base Crystal-Scaffold achieves 93.2% structural validity, 89.3% compositional validity, first-candidate MR@1 of 63.2%, and oracle any-of-20 MR@20 of 78.7% on 9,046 targets. Direct-generation controls improve MR@1 from 56.9% to 58.5% through target canonicalization; Base additionally combines scaffold planning, compiled conditioning, and staged training. HBoN achieves a selected-output MR@1 of 66.5%, compositional validity of 91.2%, and structural validity of 100%. These results support target canonicalization and compiled-scaffold conditioning as a practical formulation for reliable full-CIF generation, with hierarchical candidate selection further improving the final selected output.