Researchers
Principal Investigator
Co-Investigator
Student
This project is funded by OpenAI.
Project Summary
Large-language models (LLMs) have shown themselves capable of acting in “agent” mode on behalf of a principal, making decisions and interacting with third parties on their behalf. We investigate an important and representative category of agent tasks, specifically that of conducting negotiations with other agents. LLM-powered negotiation agents promise to help individuals explore opportunities for collaborative work, exchange of goods and services, and other kinds of deals across a wide range of platforms and interfaces. Before one would deploy an AI agent to negotiate on one’s behalf, one would wish to have confidence that the agent understands the principal’s circumstance and objectives, and is effective in reaching beneficial deals. From a societal perspective, we have additional acute interests in ensuring that:
- AI negotiators behave in a transparent and non-exploitative manner (e.g., that they do not attempt to defraud);
- interactions between separately trained AI negotiators do not produce unanticipated negative outcomes; and
- AI negotiators are effective in identifying creative deals for mutual benefit.
The central question addressed in this project is how AI models can be configured and trained to negotiate effectively, reliably, and benevolently. To answer that, we need methods to evaluate the above properties in AI negotiators, both current and future. This in turn raises questions of how to formally define and measure desired negotiation behavior, and how to apply these to agents negotiating in novel environments and contexts.
This project contributes to the research literature in autonomous agents and multiagent systems, and evaluation of advanced AI.
