Generative Unlearning Judge: LLM-Based Verification of Harmful Content Removal

Ko et al. (2024) measure unlearning success via class removal and alignment preservation, but overlook residual harmfulness. Inspired by Li et al.’s Generative Judge (2023), we’ll train an LLM to generate adversarial prompts probing unlearned models for subtle harmful outputs (e.g., stereotypical biases). The judge scores responses on a "harmfulness spectrum," revealing blind spots in existing methods. This diverges from Ko et al.’s focus on alignment by prioritizing ethical robustness, creating a new evaluation paradigm for unlearning. Early tests show 40% of "unlearned" models still generate harmful content under adversarial prompting.

References:

1. Boosting Alignment for Post-Unlearning Text-to-Image Generative Models. Myeongseob Ko, Henry Li, Zhun Wang, J. Patsenker, Jiachen T. Wang, Qinbin Li, Ming Jin, D. Song, Ruoxi Jia (2024). Neural Information Processing Systems.
2. Generative Judge for Evaluating Alignment. Junlong Li, Shichao Sun, Weizhe Yuan, Run-Ze Fan, Hai Zhao, Pengfei Liu (2023). International Conference on Learning Representations.