Constrained Program Generation for 3D Reaction Animation with a 0.8B Model

arXiv cs.GR details ChemXRG, a domain-specific language framework that treats 3D reaction animation as executable code rather than a one-off visualization. The core idea is to keep the symbolic chemistry and the rendered motion in sync through persistent atom identifiers plus explicit bond, charge, and grouping operations.

For developers, the interesting part is the constraint system: given atom-mapped reactant and product structures, the generator is forced to preserve input-determined facts and only choose from valid actions. That means generation, verification, and rendering all operate on the same program representation, which should reduce the kind of drift that makes scientific animation look plausible but wrong.

The implementation uses a reaction-program corpus and ChemQwen, a trained 0.8B DSL generator. The team says paired and component evaluations improved compiler acceptance and normalized full-program agreement under input-conditioned constraints, while execution checks still caught failures that slipped through generation.

A public browser demo ties the symbolic descriptions to inspectable programs and interactive 3D animations. Beyond chemistry, the broader takeaway is familiar to game tools and technical art teams: if you can express a visual transformation as constrained, executable data, you get validation, reproducibility, and easier iteration for free.