🤖 AI Summary
This work addresses the critical challenge of sensitive knowledge leakage in large language models trained on unfiltered corpora, proposing the first end-to-end prompt-driven forgetting framework that operates without modifying model parameters—making it applicable to closed-source models. The approach decouples knowledge forgetting into an optimization problem over learnable prompts, jointly optimizing a prompt generator and the language model via reinforcement learning. A dynamic alignment strategy is introduced to precisely suppress target knowledge while preserving general capabilities. Notably, the method requires no access to or updates of model weights, supports reversible knowledge removal, and maintains strong performance on non-targeted tasks. This advances beyond existing techniques by overcoming limitations in reversibility and transferability, offering a practical and effective solution for controllable forgetting in deployed language models.
📝 Abstract
Large language models (LLMs) trained on unfiltered corpora inherently risk retaining sensitive information, necessitating selective knowledge unlearning for regulatory compliance and ethical safety. However, existing parameter-modifying methods face fundamental limitations: high computational costs, uncontrollable forgetting boundaries, and strict dependency on model weight access. These constraints render them impractical for closed-source models, yet current non-invasive alternatives remain unsystematic and reliant on empirical experience. To address these challenges, we propose the Controllable Alignment Prompting for Unlearning (CAP) framework, an end-to-end prompt-driven unlearning paradigm. CAP decouples unlearning into a learnable prompt optimization process via reinforcement learning, where a prompt generator collaborates with the LLM to suppress target knowledge while preserving general capabilities selectively. This approach enables reversible knowledge restoration through prompt revocation. Extensive experiments demonstrate that CAP achieves precise, controllable unlearning without updating model parameters, establishing a dynamic alignment mechanism that overcomes the transferability limitations of prior methods.