AtomSteer is a test-time inference framework for steering pretrained atomistic generative models with structural constraints. Instead of retraining or changing model architecture, it corrects intermediate structure predictions during sampling so that crystallographic symmetry, structural geometry, and stereochemistry are better satisfied while preserving the geometry the pretrained model already captures.
