Mediators: An Interactive Human-AI Negotiation System for Collective Decision-Making with Multi-Agent Learning

Abstract

Collective decision-making is a core challenge across domains where multiple stakeholders must negotiate under conflicting values, limited resources, and uncertain futures. This paper introduces Mediators, an interactive system and computational framework for studying and augmenting multi-agent decision processes. The system abstracts decision-making as a multiplayer round-table environment, where strategies are trained through multi-agent reinforcement learning (MARL) and enacted through agent-agent discussions. To enrich interaction, large language models (LLMs) enable natural-language negotiation and retrieval-augmented generation (RAG) grounds dialogue in contextual knowledge. A custom physical play table provides a tangible interface for observing deliberations and exploring outcomes, supporting both AI self-play and human-in-the-loop participation. By combining strategy learning, language-based reasoning, and embodied interaction, Mediators contributes (1) a framework for simulating complex multi-agent negotiation, (2) a platform for integrating MARL with interactive discussion, and (3) a novel tool for examining how collective intelligence emerges in hybrid human-AI settings.