level: research
pre-mediation helps people reach better deals but is often skipped because of cost, time, and lack of trained mediators. researchers built an automated mediator using a structured pipeline of large language model modules. the pipeline breaks preparation into separate steps: dialogue, preference prediction, response critique, and summarization. each module handles a specific task, and outputs move forward in a fixed order. this design avoids the problems of using a single prompt for everything.
the system was tested in two controlled experiments with human participants. in the first experiment, 82 subjects negotiated a job contract. those who used the ai pre-mediation reached agreement more often and achieved higher joint scores than those who did not. the second experiment, with 120 subjects, compared the structured pipeline against a single-prompt llm mediator. the pipeline led to more agreements and better joint outcomes, showing the benefit of separating tasks.
the modules are not autonomous agents that interact freely. they follow a set sequence, passing results from one to the next. the dialogue module engages users to uncover interests. the preference module predicts what each side values. the critique module checks responses for quality. the summarization module creates a structured brief. this modular approach makes the process more reliable and easier to inspect than end-to-end generation.
why it matters: automated pre-mediation can make negotiation support scalable and accessible, reducing barriers to fair agreements in hiring, disputes, and everyday deals.