Latent Personality Alignment: Improving Harmlessness Without Mentioning Harms
For developers of LLM safety systems, LPA provides a sample-efficient defense that generalizes to unseen attack distributions, addressing the vulnerability of current adversarial robustness methods to novel attacks.
Latent Personality Alignment (LPA) achieves comparable attack success rates to methods trained on 150k+ harmful examples using fewer than 100 trait statements, while reducing misclassification rates by 2.6x across six harm benchmarks without ever seeing harmful examples during training.
Current adversarial robustness methods for large language models require extensive datasets of harmful prompts (thousands to hundreds of thousands of examples), yet remain vulnerable to novel attack vectors and distributional shifts. We propose Latent Personality Alignment (LPA), a sample-efficient defense that achieves robustness by training models on abstract personality traits rather than specific harmful behaviors. Using fewer than 100 trait statements and latent adversarial training, LPA achieves comparable attack success rates to methods trained on 150k+ examples, while maintaining superior utility. Critically, LPA generalizes better to unseen attack distributions, reducing misclassification rates by 2.6x compared to baseline across six harm benchmarks -- without ever seeing harmful examples during training. Our results demonstrate that personality-based alignment offers a principled approach to building robust defenses with minimal cost.