Adversarial Self-Play Framework for LLM Safety Alignment

A recent study published on arXiv introduces Persona-Invariant Alignment (PIA), a self-play adversarial framework designed to protect large language models from persona-driven jailbreak threats. The attack mechanism incorporates Persona Lineage Evolution (PLE), while the defense strategy utilizes Persona-Invariant Consistency Learning (PICL). Grounded in the structural separation hypothesis, PICL applies a unilateral KL-divergence constraint to dissociate safety choices from persona contexts, ensuring safe responses across various persona prompts. This research, conducted by a team of researchers, highlights the susceptibility of existing safety alignment methods to new persona-focused attacks.