Efficient Safety Alignment of Language Models via Latent Personality Traits

2026-07-08

AI Summary

Current safety alignment methods for large language models rely heavily on extensive datasets of harmful prompts and remain vulnerable to jailbreak attacks. This work proposes Latent Personality Alignment (LPA), a novel approach that introduces personality psychology into safety alignment for the first time. LPA employs only 66 harmless personality statements for adversarial training, implicitly constraining the model’s vulnerability within a latent subspace susceptible to attacks. Requiring no harmful data and training efficiently in minutes on a single GPU, LPA reduces the data volume by 75× compared to standard Latent Adversarial Training (LAT). It achieves near-zero attack success rates on HarmBench against both direct requests and five categories of jailbreak attacks, while preserving full performance on general capability benchmarks.

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.

Problem

• safety alignment
• adversarial attacks
• jailbreak attacks
• large language models
• harm avoidance

Innovation

• Latent Personality Alignment
• Adversarial Training
• Safety Alignment
• Jailbreak Defense
• Efficient Fine-tuning