LLM backed autonomous agents in multi-issue multilateral negotiations

Extending beyond bilateral scenarios, this study explores the capability and limitations of LLM-backed agents in multi-issue multilateral negotiations, examining how structural factors impact outcomes. We propose a framework and employ a controlled simulation environment based on a scored negotiation game to analyze how LLM agents adapt to varying group compositions and procedural settings, thereby isolating the effects of structural design. Our findings reveal that LLM-backed agents are indeed capable of executing autonomous negotiation tasks in multilateral settings. Success and the achievement of desirable results were significantly enhanced by the assistance of designed protocols and prompts. This research demonstrates the feasibility of using LLM agents in complex multilateral negotiations and provides empirical insights crucial for designing robust autonomous systems, suggesting important considerations for decision rules, round limits, number of agents and model size.