(iPoster) Multi-Agent System for Negotiation Simulation with LLMs and Russia-Ukraine Case

• In Event: Poster Session: Conflict Processes 1

Abstract

Do the subjective perceptions of those in power or the objective conditions created by the common people of their countries play a greater role in shaping state actions in crisis times? In this paper, we integrate theoretical models into a multi-level negotiation simulation system prototype that utilizes Large Language Models (LLMs) as agents to simulate complex international negotiations. To test our analytical framework, we are using the Russia-Ukraine ceasefire negotiations in the future as a case study and constructing an interaction model with teams from three parties: the United States, Russia, and Ukraine. Our implementation is based on Langraph and AutoGen for multi-agent system development, where each agent is operated by a portable Llama-2-7B enabling inter-agent communication and dialogue. The agents' personas are designed to emulate the above three parties' corresponding political elites like President Donald Trump, President Vladimir V. Putin, President Volodymyr Zelensky, and the aides by incorporating their documented leadership characteristics, long-term goals of their administration, specific stances on a particular issue (the Russia-Ukraine war), educational backgrounds, professional experiences, and inferred personality traits. The objective conditions of the three countries like economic capacity, military prowess, political stability and governance, cultural influence, technological advancement, diplomatic influence, and natural resources are also in our simulation and are more relevant to the common people. We present a prototype pipeline for developing and evaluating LLM-powered multi-party negotiations, demonstrating methodologies for model deployment, performance assessment, and limitation analysis through this case study. We first adopt the negotiation simulation to predict the specific terms of the Russia-Ukraine ceasefire deal and will use its actual terms to calibrate our simulation. In the end, we provide a systematic approach to assess model capabilities, system scalability, and performance metrics while understanding both the potential and limitations of LLM applications in diplomatic simulations. This study mainly speaks to Rationalist Models of Crisis Bargaining, particularly the works of Thomas Schelling and James Fearon, which focus on strategic interactions under conditions of uncertainty and emphasize signaling, commitment, and the costs of war. Our multi-agent system simulates the strategic communications between different parties, reflecting how states convey their intentions and capabilities. The simulation may address issues of credibility and commitment, which are central to rationalist models. By incorporating the political and military contexts of the parties, our model acknowledges the incomplete information that can lead to conflict. The focus on predicting and calibrating the terms of a ceasefire deal aligns with the negotiation and bargaining processes studied in these models. By changing the value of each of the above variables in our simulation in turn while keeping all the other variables constant and then observing the change in the results each time, we may also obtain new evidence for the three key theories of International Relations, realism, liberalism, and constructivism. More specifically, if the change in the variables relevant to material power like a country’s strength relative to other countries and the structures of politics such as the global commodity or international state system and financial markets exerted the most obvious effect on the outcome of the negotiation, realism may be supported. When the change in the variables germane to economic interests most noticeably influences the outcome, liberalism seems to work better. If individual leadership traits like leaders' personalities, political goals, identities, and beliefs dominate state actions during the international negotiation, agent behavior in our simulation will reflect constructivist ideas. Hence, the cutting-edge LLMs in our research help us to reimagine how the characteristics of those in power themselves and the conditions created by the people of their countries interact in conflict resolution negotiations.

Authors

• Yen-Chieh Liao, University of Birmingham

• Shuli Zhang, University of Pittsburgh