Inference-time Unlearning Using Conformal Prediction

Machine unlearning is the process of efficiently removing specific information from a trained machine learning model without retraining from scratch. Existing unlearning techniques perform unlearning either by updating the model parameters post hoc or re-training specific model components. However, these approaches often lack formal guarantees, allowing the resulting models to leak confidential information, which is risky especially in generative settings. To address these challenges, we propose a light-weight strategy to unlearn from language models at inference-time. In contrast to existing techniques, we encapsulate the unlearning objective within a verifier. This verifier is an LLM-as-a-judge framework that evaluates the quality of generated responses. We iteratively refine the quality of the generated responses using feedback from the verifier. We use conformal prediction to set the hyperparameters of this framework to reduce computational costs and provide distribution-free unlearning guarantees. We perform extensive empirical evaluation on several challenging unlearning tasks and showcase that our approach significantly outperforms existing methods, achieving up to a 95\% unlearning error reduction compared to a vanilla LLM with minimal degradation in overall response quality.