CAP Framework Uses Reinforcement Learning for LLM Unlearning

A novel approach known as Controllable Alignment Prompting for Unlearning (CAP) tackles the issue of selectively erasing knowledge in large language models (LLMs). Current methods that modify parameters face challenges such as high computational demands, unpredictable forgetting limits, and reliance on access to model weights, rendering them ineffective for closed-source models. CAP separates the unlearning process into an optimization of prompts that can be learned through reinforcement learning. A prompt generator works alongside the LLM to eliminate specific knowledge while maintaining overall functionality. This method is comprehensive and driven by prompts, providing a non-invasive solution that avoids altering model weights. The research is published on arXiv with the identifier 2604.21251.