Regime-Conditional Verification: A Lightweight Method to Adapt Safety Classifiers

A new study uploaded to arXiv (2608.14089) introduces Regime-Conditional Verification (RCV), an innovative method that tweaks existing safety classifiers for large language models without needing retraining. RCV addresses two common challenges: classifiers that reflect the training policy rather than the intended deployment policy, and performance drops due to changes in deployment traffic. By evaluating the classifier's internal workings, RCV identifies when a prediction might go against the deployer's policy and makes corrections as necessary. It also provides a way to detect shifts in data distribution without labels, creating a cycle for updating correctness estimations and only fine-tuning the classifier when absolutely essential. RCV improves adherence to the deployer's policy across all tested classifiers, supporting three safety classifiers and two benchmark datasets.