Recent advances in Large Language Models (LLMs) have significantly expanded their capabilities beyond simple information processing and text generation, toward emulating human-like social communication skills such as emotional empathy, persuasion, and...
Recent advances in Large Language Models (LLMs) have significantly expanded their capabilities beyond simple information processing and text generation, toward emulating human-like social communication skills such as emotional empathy, persuasion, and responsibility-based decision-making. This progress highlights the potential of LLMs to function as autonomous agents in cooperative or conflict-driven situations involving social interaction. Negotiation, in particular, is a representative domain of complex social decision-making where conflicting goals and emotional tension are intertwined, requiring human-level understanding and coordination. However, existing studies have primarily focused on simplified exchange scenarios that can be easily quantified, such as price negotiations, leaving empirical investigations on emotional conflict resolution and strategic mediation capabilities relatively unexplored. Moreover, systematic analyses of prompt-based strategy control and the quantitative effects of third-party mediation remain limited. Motivated by these gaps, this study aims to examine the emotion-aware social interaction abilities of LLMs, focusing on their potential roles as negotiation agents and mediators. Specifically, we first analyzed the ability of modern LLMs to perceive and respond to emotionally charged conversational dynamics. Further, we conducted AI-to-AI negotiation experiments under affectively intense
conflict scenarios using GPT-3.5-turbo, GPT-4, and Claude Opus 4.1. We compared changes in negotiation capability and agreement achievement across identical scenarios based on two factors: (1) the application of refined prompts incorporating emotional expression and strategic guidance, and (2) the inclusion of a third-party mediator LLM. Negotiation outcomes were evaluated using Pareto Efficiency Frontier analysis to assess mutual benefit and Jensen–Shannon Divergence (JSD) to quantify fairness and strategic balance. We additionally measured negotiation persistence through dialogue length and agreement rates. The results demonstrate that more advanced LLMs exhibit higher stability in emotional responsiveness and fairness-oriented decision-making. The presence of a mediator significantly improved negotiation continuity, conflict mitigation, and
agreement rates. Strategic prompting further enhanced both negotiation efficiency and persuasive capability, with measurable differences in persuasive strategies across models. These findings suggest that LLMs can serve not merely as language generators, but as negotiation actors capable of interpreting social context and facilitating conflict resolution. This study contributes to the advancement of LLM-based social decision-making research by experimentally verifying their role in emotional negotiation settings. The results provide foundational insights for future developments in human–AI and AI–AI interaction design, as well as AI-driven conflict mediation systems.