Talk ain't cheap: DARPA offers grants for new AI-to-AI communication protocol

MATHBAC program wants better machine-to-machine chatter for scientific discovery

To supercharge agents' ability to make scientific discoveries, DARPA is looking to improve cross-bot collaboration by developing a "science of AI communication" that will help the models work together to come up with better ideas. 

The Pentagon research arm on Tuesday announced its Mathematics of Boosting Agentic Communication (MATHBAC) program with a solicitation inviting researchers interested in advancing the “foundational mathematics, systems theory, and information theory required to enable/accelerate and understand science discovery by autonomous agents and agent collectives” to toss their hats in the ring for awards of up to $2 million in Phase I funding under a 34-month, two-phase project.

DARPA’s rationale for the program is straightforward: AI development has led to some amazing accomplishments so far, but much of it remains heuristically guided, an ad hoc, trial-and-error process focused on outcomes rather than a full understanding of why those outcomes occur. The same problem applies to agent-to-agent communication: without what DARPA calls a “rigorous mathematical foundation” for understanding how agents communicate and coordinate, those interactions will remain inefficient, inconsistent, and difficult to generalize across domains.

"While AI excels at navigating solution spaces, it struggles to systematically explore hypothesis spaces, which are essential for generating transformative and generalizable scientific insights," DARPA explained in the program announcement. "It is precisely the rate of discovery of these important new hypotheses that MATHBAC aims to systematically accelerate by facilitating AI communication so as to enable breakthroughs in the efficiency of agentic scientific reasoning."

The first phase of MATHBAC will be dedicated to developing the mathematics for understanding and designing agentic communication protocols and improving the content of those communications, which means the project isn't just about how AI agents are communicating, but also what they're communicating about. 

"Beyond evaluating and optimizing interaction protocols, MATHBAC will also focus on features of the content of agentic communication," DARPA noted in describing the second technical area of the project (the first being the actual math behind agent communication protocols).

That second technical area will involve looking at the content of agent-to-agent interactions with a "focus on the discovery of 'principles' (laws, correlations) from data–the extraction of compact, generalizable 'nuggets' that should become part of the common knowledge module ('memory') of cooperating agents," the announcement explains. 

In essence, what the second technical area is trying to do is figure out (in the first MATHBAC phase) if a group of AI agents trained in specific scientific areas can infer general scientific principles, laws, or correlations from a set of data that suggests a generalizable rule but doesn't spell it out. 

As an example, DARPA said that a hard goal for the project (one that's generally considered nigh impossible to meet) would be to start with "the data-driven Mendeleev-level rediscovery of the periodic table for atoms and proceed to a 'multidimensional analog' of a periodic table for molecules." 

Fingers crossed this works

"If successful, MATHBAC will fundamentally change the 'ways of doing,' whether for scientific discovery or for instruction," DARPA said in the announcement. It's got that right: This is not going to be an easy project - DARPA even said that it won't entertain any solicitations for research "that primarily results in incremental improvements to the existing state of practice," meaning that it's seeking some truly revolutionary leaps in AI science through the program if it's to be considered a success. 

Beyond the challenges presented by the first phase of the project, phase two is asking MATHBAC researchers to go even further and focus on the creation of AI tools "that enable systematic evolution/invention of new science." 

That objective will be achieved, DARPA said, by directing AI agents to self-evolve in ways that maximize their ability to solve scientific problems using the previously defined communication protocols. DARPA said that it's possible AI agent coordination at the level it's hoping to achieve may require an entirely new domain language unique to AI agents, which will be explored in the second phase of the program.

"Up to now the evolutionary pressure on AI and agentic platforms comes from evolutionary pressures on the humans involved in their development. MATHBAC wants to move this pressure onto the agents themselves and their cooperation skills," the agency explained. "MATHBAC will systematically explore, understand, and design the best protocols, content, and possibly even language of collaborative AI agent communication."

MATHBAC proposals are due by June 16, with the program planned to start this September. Multiple awardees are anticipated. DARPA didn't respond to questions for this story. ®