Educational game for conflict mediation training in wartime conditions using large language models

Interpersonal conflicts increase significantly during wartime, negatively impacting psychological well-being and social cohesion. This research introduces an innovative educational game that teaches mediation skills through interactive dialogue with characters generated by large language models (LLMs). The game features dynamically generated conflicts and personalized responses based on player actions, allowing users to practice mediation strategies in a safe, repeatable environment. We implemented the system using the Gemini 1.5 Flash LLM and conducted experiments to optimize model parameters and evaluate the effectiveness of different mediation strategies. Our results demonstrate that the compensation strategy proves most effective in our generated conflict scenarios. The system provides a quantitative method for evaluating mediation strategies, which has been impossible in real-world settings. This novel approach fills a significant gap in mediation education, offering an accessible tool for training mediators, particularly in conflict-affected regions such as Ukraine.