From consensus theory to LLM agents: Practical consensus-building for multi-issue negotiation

The increasing performance of large language models (LLMs) encourages research on using LLM-based multiagent systems (MAS) to simulate and predict human activities in the real world, especially in the context of negotiation simulation. Meanwhile, the consensus-reaching process (CRP) research has been developed to simulate people’s perception of consensus in practice, particularly the development of consensus models to address specific consensus-reaching issues. However, current LLM negotiation studies largely rely on prompts and outcome scores, offering limited guarantees on stability, salience-aware behaviour, or meaningful termination, whereas CRP provides explicit consensus indices, update logic, and stopping rules. Bridging these lines of work enables negotiation that is measurable rather than anecdotal and reliable rather than round-cap dependent. This paper argues that a bridge between these lines of work is timely and introduces a systematic framework to bring the language of consensus—quality, stability, fairness, and robustness—into LLM-based MAS, allowing different agent designs to be compared on common standards. We introduce and combine methods of the fuzzy logic theory to quantify people’s preferences. We also design a general workflow for LLM-based agents to reach potential consensus in two predefined CRP scenarios. Finally, we introduce and define several cross-paradigm metrics to evaluate the performance of three different agent designs. The experimental results indicate that incorporating consensus models improves stability and fairness. Overall, the paper reframes LLM negotiation as consensus-aware, stability-measurable negotiation, providing a practical bridge between CRP theory and LLM-based agents, and offering a reproducible toolkit for transparent, comparable assessment of multi-issue negotiation.