Misalignment Hotspots: Probing Prompt-Induced Alignment Failures in Post-Unlearning Models

While Ko et al. (2024) address alignment degradation post-unlearning at a macro level, this idea explores micro-level alignment failures. We hypothesize that certain prompt structures (e.g., abstract concepts, multi-object relationships) disproportionately trigger misalignment after unlearning. By systematically testing prompt categories and measuring alignment drift using Conditional Vendi scores (Jalali et al., 2024), we’ll map "misalignment hotspots." This extends Ko et al.’s work by shifting from average alignment preservation to prompt-specific vulnerability analysis, enabling targeted unlearning refinements. The novelty lies in treating alignment as a prompt-dependent phenomenon rather than a monolithic metric.

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. Conditional Vendi Score: An Information-Theoretic Approach to Diversity Evaluation of Prompt-based Generative Models. Mohammad Jalali, Azim Ospanov, Amin Gohari, Farzan Farnia (2024). arXiv.org.