New Framework Enables API-Only Black-Box LLM Unlearning

As large language models (LLMs) are increasingly accessed via API services, the need to remove sensitive, copyrighted, or outdated information from their behavior without full retraining becomes critical. Existing machine unlearning methods often require access to the model's internal parameters or logits, which is not feasible for black-box API-only LLMs. Furthermore, these methods struggle to preserve the model's general utility when the data to be forgotten is semantically very similar to the data that should be retained. This paper introduces Controlled Behavioral Divergence (CBD), a novel framework designed to address these challenges. CBD operates entirely through API access, leveraging two auxiliary models to generate a controlled divergence in behavior between retained and unlearning-target inputs. This divergence is then quantified into an unlearning relevance score, which guides the routing of specific prompts away from the target LLM. To enhance discrimination, especially when target and retained data are highly similar, CBD constructs a discriminative basis using gradient statistics and a regularized generalized eigenvalue problem. This ensures the unlearning signal targets specific information rather than broad prompt structures. Experiments show CBD outperforms eleven white-box and gray-box baselines, achieving a better trade-off between unlearning effectiveness and utility preservation. For instance, on the ToFU forget10 dataset, CBD approached the performance of a fully retrained model on the forget set while maintaining high utility, and on WMDP, it significantly reduced hazardous knowledge accuracy while preserving general knowledge.