Search papers, labs, and topics across Lattice.
This paper introduces Latent Personality Alignment (LPA), a novel approach to enhancing the safety of language models by utilizing adversarial training on a minimal set of 66 harm-agnostic statements derived from psychometric literature. By leveraging personality-anchored representations, LPA effectively constrains the latent space exploited by adversarial attacks, achieving near-zero success rates against various attack methods without exposing the model to harmful content during training. The method is not only robust but also efficient, requiring significantly fewer training examples and completing in minutes on a single GPU, thus maintaining utility on standard benchmarks.
Adversarial training on just 66 carefully chosen statements can achieve near-zero attack success rates, revolutionizing safety alignment in language models.
Current safety methods for large language models are known to be vulnerable to adversarial attacks, motivating research into robust alternatives. Latent Adversarial Training (LAT) is among the most effective defenses, but can degrade utility and requires training on large datasets of harmful prompts. We introduce Latent Personality Alignment (LPA), which replaces explicit harm refusal with adversarial training on just 66 harm-agnostic statements drawn from psychometric personality literature. We hypothesize that personality-anchored representations share latent structure with harm avoidance, so adversarially stabilizing them implicitly constrains the subspace exploited by jailbreak attacks. LPA achieves near-zero attack success rates on HarmBench across direct requests and five jailbreak methods, despite never seeing harmful content during training and no loss of performance on standard benchmarks. Moreover, the training process is lightweight; the entire procedure completes in minutes on a single GPU and uses 75x fewer examples than standard LAT. Extensive ablations demonstrate the robustness, efficiency, and generalization of our method.
Mila