CAP: Controllable Alignment Prompting for Unlearning in LLMs
For developers and users of closed-source LLMs, CAP provides a non-invasive, controllable unlearning method that addresses computational and accessibility issues of existing approaches.
The paper proposes CAP, a prompt-driven framework for unlearning in LLMs that uses reinforcement learning to optimize prompts, achieving precise knowledge suppression without parameter updates. Experiments show it enables controllable and reversible unlearning, overcoming limitations of prior methods.
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.