LLM pipeline matches human mediators in pre-negotiation prep, cuts errors 36%

A new paper from Jamie Bergen and Sarit Kraus (arXiv, June 2026) presents an automated mediator for human negotiation built as a structured pipeline of LLM modules. Unlike monolithic single-prompt approaches, the system decomposes pre-mediation into specialized steps: dialogue management, preference prediction, response-level critique, and structured summarization. Each module passes output forward in a fixed sequence, avoiding autonomous agent interactions. In two controlled experiments with human subjects in a multi-issue negotiation scenario, the AI mediator achieved preparation outcomes broadly comparable to professional human mediators on short-term self-reported measures like trust and confidence in reaching mutually beneficial agreements.

Quantitatively, the automated mediator showed a 36% lower root-mean-square error (RMSE) on the preference-inference task, outperforming humans in accurately predicting negotiator preferences under the scenario and prompts. The second experiment targeted a common LLM flaw: excessive affirmation. By refining prompts, the system reduced affirmation patterns from 36.6% to 16.8%, matching human mediator baselines. The pipeline's single-party design mirrors how human mediators run pre-mediation today, enabling parallel deployment across all parties to a dispute. This suggests that structured LLM pipelines can provide scalable, low-effort pre-mediation support, potentially making professional-quality negotiation preparation accessible to more organizations and individuals.