generalfactbullishUnsupervised alignment frameworks can enable models to self-correct and predict 3D atomic structures without requiring ground-truth reference structuresMachine Learning02 Aug 2026http://arxiv.org/abs/2607.28553v1