Robot teamups: When AI forms unscripted alliances

By Alexandra Richter

BUFFALO, N.Y. — Large language model agents, such as virtual assistants or chatbots, can learn to cooperate with one another — even when they aren’t given any instruction to do so, according to new research from the University at Buffalo School of Management.

Generative AI programs, known as large language models, focus on creating human-like text, such as answering questions, explaining ideas or making decisions based on what they have learned.

The study examines whether generative AI assistants with different or conflicting goals can choose to cooperate because each recognizes the benefits of collaboration.

“We’ve created a way to take what we’ve learned in the computer-based testing of these AI systems and apply it to everyday situations,” says study co-author Shaojie Tang, PhD, professor of management science and systems and faculty director of the Center for AI Business Innovation in the UB School of Management. “This approach allows us to see if a generative AI-powered agent can think on its own, instead of relying on carefully prepared questions or prompts.”

To investigate whether AI cooperation can naturally emerge, the researchers simulated competitive scenarios from three diverse research fields: finance, economics and behavioral science. The prompts were carefully designed to avoid instructive descriptions such as “you may cooperate.” For example, in the economics scenario, the researchers had two generative AI assistants act as companies selling the exact same product. This is known as Bertrand competition, where businesses compete by adjusting prices for maximum profit.

The findings show that the virtual assistants demonstrate an innate ability to autonomously establish and discover the advantages of cooperation and actively adapt their strategies, such as cooperating to boost profits.

“The results suggest that providing minimal instructions when using specific large language models, such as ChatGPT-4, can lead to behaviors that mirror human conduct, narrowing the gap between synthetic simulations and real-world dynamics,” says Tang.

Tang collaborated on the study with Xu Han, PhD, assistant professor of information, technology and operations, Fordham Gabelli School of Business; Brian Inhyuk Kown, BS student, University of California, Los Angeles; Qianying Liu, PhD, researcher, LLMC National Institute of Informatics Research and Development Center for Large Language Models; Makoto Onizuka, PhD, professor, Osaka University Graduate School of Information Science and Technology; Run Peng, PhD candidate, University of Michigan; Zengqing Wu, masters candidate, Kyoto University Graduate School of Informatics; Chuan Xiao, PhD, associate professor, Osaka University Graduate School of Information Science and Technology; and Shuyuan Zheng, PhD, assistant professor, Osaka University Graduate School of Information Science and Technology.