Efficient Safety Alignment of Language Models via Latent Personality Traits
Mohamed Amine Merzouk, Nolan Smyth, Damiano Fornasiere, Linh Le, David Williams-King, Adam Oberman
Why It Matters
What makes this one worth your time
This method offers a potentially more efficient and scalable way to enhance the safety of language models, reducing the need for large datasets of harmful content and extensive computational resources.
LPA aligns language models using personality traits to prevent harmful outputs efficiently.
Summary
The paper introduces Latent Personality Alignment (LPA), a method for aligning language models to avoid harmful outputs by leveraging psychometric personality traits instead of explicit harmful content. The approach uses adversarial training on a small set of harm-agnostic statements, achieving low attack success rates without degrading model performance or requiring extensive harmful data.
Key contributions
- Introduction of Latent Personality Alignment (LPA) for language model safety.
- Demonstration of effective adversarial training using only 66 harm-agnostic statements.
- Achievement of low attack success rates without performance loss on standard benchmarks.
Notable insights
- Using psychometric personality traits as a proxy for harm avoidance in adversarial training.
- Achieving robust safety alignment with significantly fewer training examples and computational resources.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2607.07918v1 Announce Type: cross Abstract: 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.