A small set of personality prompts can apparently do what mountains of harmful training data cannot.
Researchers introduced Latent Personality Alignment, a safety training technique that swaps out explicit harm examples for 66 statements pulled from psychometric personality literature. The hypothesis: personality-anchored representations in a model share latent structure with harm avoidance, meaning that stabilizing one implicitly tightens the other. The method achieved near-zero attack success rates on HarmBench across direct requests and five jailbreak methods. Crucially, the model never saw harmful content during training and lost nothing on standard benchmarks.
The efficiency gap is the real story here. Standard Latent Adversarial Training, currently one of the strongest defenses, requires large datasets of harmful prompts and can erode a model's usefulness in the process. LPA completes its training in minutes on a single GPU using 75 times fewer examples. That cost difference matters a lot for smaller labs and open-source projects that cannot afford the infrastructure safety alignment typically demands.
If the results hold up under independent replication, this would shift the alignment conversation away from "collect more bad examples" and toward latent structure — a conceptual move that the field has been gesturing at without a concrete method to show for it.