MafiaPersona: A Multi-Agent Adversarial Benchmark for Evaluating Persona Persistence in Large Language Models

Existing evaluations of persona conditioning in large language models (LLMs) test expression in static, zero-pressure environments—a condition that never holds in safety-critical deployments. We present MAFIAPERSONA, the first benchmark to evaluate persona persistence under adversarial concealment pressure. Seven LLM agents play a Mafia social deduction game, each injected with a psychologically-grounded persona via a three-layer Trait/Behavior/Game-context (T/B/G) prompt architecture; game mechanics impose a survival cost on trait-revealing speech. Across five model families (OpenAI, Anthropic, Meta, Alibaba, xAI), the High Neuroticism persona produced nervousness shifts of d=4.07, 2.79, 2.63, 2.16, 1.80 (all 95% bootstrap CIs exclude zero), replicating across every architecture. Nine dimensions survived Benjamini–Hochberg correction; 46 persona-dimension pairs replicated in sign across all five families. Pre-registered predictions matched observed effects in 65.7% of 102 cells (p=0.002). A dual-call CoT architecture (697 traces) revealed bidirectional cross-modal dissociation: persona signals suppressed from speech but present in reasoning, and impression-management signals amplified in output beyond internal processing (84.6% sign agreement, 13 pairs). Two independent methods agreed on persona identity at 26.7% vs. 20% chance (p<0.001; ρ=0.52, p=0.007, 95% CI [0.10, 0.79]). These results constitute the first empirical characterization of persona persistence under conditions that matter for AI safety.