Core unlearning: A multi-modal gradient-efficient architecture for exact and approximate model rewriting

Machine unlearning is important for data security, user confidence, and regulatory compliance in AI systems. Despite the significant achievement, existing techniques have limited generalizability across a broad set of forgetting scenarios — feature, class, task, stream, or catastrophic forgetting, and are devoid of a theoretical base, scalability, or computational efficiency. The proposed Core Unlearning (CU) framework bypasses these limitations by integrating state-of-the-art methods like latent space loss optimization, gradient ascent-augmented updates, Adapter Partition and Aggregation (APA), and Projection-Based Residual Adjustment (PBRA) into a unified structure that supports both Exact Unlearning (EU) and Approximate Unlearning (AU). In EU, Negative Preference Optimization (NPO) is employed, a strategy that treats target data as negative samples to actively suppress their influence during unlearning by penalizing correct predictions on forgotten data. Evaluating across multi-modal datasets like CIFAR-10, CIFAR, 100, IMDB4K, CORA, FEMNIST, and MVTec AD, CU achieves improved performance in forgetting fidelity, model utility, and privacy preservation. The GA+APA+NPO achieves up to 2.3% decreased accuracy loss, with 95.2% retraining equivalence, proving high-fidelity unlearning. In AU mode, our approach gets 92.3% forgetting accuracy, 85.7% utility score, and 90.2% unlearning efficiency, enabling a scalable solution for time-critical applications. With a seamless combination of EU and AU into a single paradigm, CU enables versatile management of the precision-speed trade-off, with support for strong application-specific unlearning. The work in this paper demonstrates an early step toward useful, mathematically robust, and privacy-preserving machine unlearning. Code available at:CoreUnlearning.