CBD Enables API-Only Black-Box Unlearning for LLMs.

As Large Language Models (LLMs) are increasingly accessed via API services, the need to remove the influence of sensitive, copyrighted, or harmful data without retraining the entire model becomes critical. Traditional machine unlearning methods often require access to model parameters or internal logits, which is not feasible for API-only black-box LLMs. Furthermore, existing techniques struggle when the data to be forgotten shares strong semantic or structural similarities with data that should be retained. To address these challenges, a new framework called Controlled Behavioral Divergence (CBD) has been proposed. CBD operates entirely through API access, meaning it doesn't require internal model details. It employs two auxiliary models to create a controlled divergence in behavior between retained inputs and the inputs associated with the data to be unlearned. This divergence is then converted into an "unlearning relevance score." Based on this score, CBD intelligently routes prompts related to the unlearning target away from the main LLM. To enhance discrimination, especially when data is highly similar, CBD constructs a discriminative basis using gradient statistics, guiding the unlearning signal specifically towards target information rather than shared patterns. Experiments show CBD outperforms eleven other unlearning baselines, achieving a superior balance between forgetting undesired data and preserving overall model utility, even approaching retrained reference performance on forget sets while maintaining high utility on retained data.