Latent Personality Alignment: Improving Harmlessness Without Mentioning Harms

Authors: Linh Le, David Williams-King, Mohamed Amine Merzouk et al.

arXiv · PDF

Summary

Instead of training language models to avoid harm using thousands of examples of bad prompts and refusals, the authors train on fewer than 100 abstract personality-trait statements (like "I am cautious" or "I value honesty") using adversarial fine-tuning. This "Latent Personality Alignment" matches the robustness of methods trained on 150,000+ harmful examples and generalizes 2.6× better to new attack types—all without ever showing the model a single harmful example during training. The key idea is that teaching broad personality traits transfers better across attack distributions than memorizing specific harm categories.

Main takeaways:

• Training on abstract personality traits (fewer than 100 statements) achieves the same attack robustness as training on 150,000+ harmful prompt-response pairs
• Personality-based training generalizes 2.6× better to unseen attack distributions across six harm benchmarks
• The method never shows harmful examples during training yet still learns to refuse harmful requests
• Latent adversarial training on trait statements maintains model utility better than harm-specific training
• This suggests that high-level behavioral abstractions (personality) are more robust anchors than low-level pattern matching on harm categories

Relevance

Directly relevant to my persona work: this is installation-path equivalence in reverse—they're implanting refusal behavior by training on abstract trait descriptions rather than explicit harm examples, exactly the kind of "what representation installs what behavior" question I'm asking about personas, markers, and steering vectors.

Threat model

Potential threat/caveat for clean result "Fine-tuning one persona on a two-marker chunk and another on the start marker plants the end marker at every donor answer's end, not chained to the start (LOW confidence)": this item discusses robustness, adversarial, benchmark.

Abstract

arXiv:2605.08496v1 Announce Type: new Abstract: Current adversarial robustness methods for large language models require extensive datasets of harmful prompts (thousands to hundreds of thousands of examples), yet remain vulnerable to novel attack vectors and distributional shifts. We propose Latent Personality Alignment (LPA), a sample-efficient defense that achieves robustness by training models on abstract personality traits rather than specific harmful behaviors. Using fewer than 100 trait statements and latent adversarial training, LPA achieves comparable attack success rates to methods trained on 150k+ examples, while maintaining superior utility. Critically, LPA generalizes better to unseen attack distributions, reducing misclassification rates by 2.6x compared to baseline across six harm benchmarks -- without ever seeing harmful examples during training. Our results demonstrate that personality-based alignment offers a principled approach to building robust defenses with minimal cost.