Cross-Interface Policy Regularization (CIPR)

A self-supervised regularization method for LLM agents that constructs representations of equivalent actions under different interfaces and uses KL divergence to constrain policy consistency, improving agent robustness under interface perturbations.