The U.S. Defense Advanced Research Projects Agency (DARPA) officially launched a new project on Tuesday dubbed "Mathematical Communication for Agent Collaboration" (MATHBAC). The initiative seeks to build a rigorous mathematical foundation to improve communication mechanisms between AI agents, thereby accelerating the pace of scientific discovery.
DARPA has issued a solicitation inviting researchers to submit proposals for this 34-month, two-phase project. Selected teams will be eligible for up to $2 million in funding for the first phase.
Building a Mathematical Cornerstone for AI Collaboration
In its announcement, DARPA noted that while current artificial intelligence excels at searching through solution spaces, it remains limited in its ability to systematically explore hypothesis spaces. Existing AI models often rely on heuristic trial-and-error methods and lack a deep understanding of the logic governing agent-to-agent interaction, leading to inefficient and difficult-to-scale collaboration.
"The goal of MATHBAC is to achieve a breakthrough in scientific reasoning efficiency by facilitating effective communication between AIs, systematically accelerating the discovery of new hypotheses," DARPA stated in the project description.
The first phase of the project will focus on developing mathematical models for designing and understanding agent communication protocols. The research will emphasize not only how agents exchange information, but also the quality of the content being exchanged.
Moving into the second phase, the project will delve deeper into the communication content itself. Researchers will need to evaluate whether agents can extract "principles" from data—essentially distilling generalizable scientific laws or correlations—and translate this information into "memory" modules shared among collaborative agents.
To validate these objectives, DARPA has set an ambitious vision. For instance, the agency hopes that agents will be able to autonomously derive scientific laws, such as Mendeleev’s Periodic Table, from datasets rather than relying solely on pre-programmed logic.
Through this initiative, DARPA aims to transform AI collaboration from its current ad-hoc, interactive state into a scientific process based on rigorous systems theory, ultimately enabling AI agents to work together more effectively to solve frontier technological challenges.