New Method Aligns LLM Safety Using Latent Personality Traits

Current methods for ensuring the safety of large language models (LLMs) are often vulnerable to adversarial attacks, prompting the search for more robust alternatives. While Latent Adversarial Training (LAT) is effective, it can reduce model utility and requires extensive datasets of harmful prompts. A new approach, Latent Personality Alignment (LPA), offers a more efficient solution. LPA replaces explicit harm refusal with adversarial training focused on just 66 harm-agnostic statements derived from psychometric personality literature. The core hypothesis is that representations anchored in personality traits share a latent structure with harm avoidance, meaning that stabilizing these representations adversarially can implicitly constrain the areas exploited by jailbreak attacks. This method has demonstrated remarkable effectiveness, achieving near-zero attack success rates on HarmBench against various jailbreak techniques, all without ever being exposed to harmful content during training. Crucially, LPA maintains performance on standard benchmarks and is significantly more efficient, completing training in minutes on a single GPU with 75 times fewer examples than traditional LAT.