Simulating Ideological Deadlock: Persona Stability and Conflict Resolution in LLM-Based Multi-Agent Frameworks

The integration of Large Language Models (LLMs) into Agent-Based Modeling (ABM) has revolutionized the capacity to simulate complex sociopolitical dynamics. However, state-of-the-art models are fundamentally aligned via safety protocols (e.g., RLHF) to prioritize helpfulness and consensus, severely limiting their utility in simulating intractable, zero-sum human conflict. To address this limitation, this study presents a novel sociotechnical stress test designed to measure the boundaries of persona stability and affective polarization under extreme cognitive dissonance. Utilizing the AutoGen multi-agent framework and the Gemini foundation model, we conducted a two-phase, within-subject simulation ( N  = 29 independent trials) featuring a 15-round ideological deadlock based on an irreconcilable historical vignette. Generative agents were assigned radically opposing, uncompromising personas to artificially sustain conflict. Phase 1 established a baseline of unmediated polarization, while Phase 2 introduced a "Synthesis Persona" tasked with linguistic mediation. Turn-by-turn psychometric text analysis utilizing LIWC-22 revealed that the intervention failed to generate statistically significant improvements in cognitive complexity ( p = .312) or reductions in affective polarization ( p = .300). However, analysis of power dynamics demonstrated a highly significant programmatic regression: agents exhibited severe persona drift, abandoning their dominant communicative postures upon intervention ( p < .001, d  = 0.724). These findings quantify the psychological fragility of artificial personas, establishing crucial empirical boundaries for deploying "Artificial Humans" in high-conflict sociological simulations, and demonstrating that foundation models resolve simulated conflict through submission rather than intellectual synthesis.