New BiSDG framework uses bi-level optimization to train AI models on single domains that generalize to unseen target environments without target data

BiSDG separates task learning from domain modeling through a two-level optimization approach to improve single domain generalization (SDG)

The framework creates surrogate domains using label-preserving transformations to simulate distribution shifts in training data

A domain prompt encoder generates lightweight modulation signals that adjust features via feature-wise linear modulation for domain-specific context

Inner optimization loop focuses on task performance with fixed prompts while outer loop maximizes generalization across surrogate domains