Anchor Invariance Regularization for LLM Safety Alignment

A new arXiv paper proposes Anchor Invariance Regularization (AIR) to improve safety alignment in large language models by enforcing context-invariant behavior. Current preference-based post-training often fails when harmful requests are reworded adversarially. AIR treats verifiable prompts (e.g., multiple-choice) as anchors to regularize open-ended variants, avoiding performance degradation on reliable signals. The approach aims to make safety decisions depend on underlying intent rather than surface form.