The US Defense Advanced Research Projects Agency (DARPA) has launched expMath, a fresh initiative aimed at accelerating mathematical progress with the help of artificial intelligence. Despite mathematics forming the backbone of everything from cryptography to modeling complex systems, its problem-solving methods remain rooted in centuries-old traditions. DARPA envisions ´AI coauthors´ that could tackle large, complex math challenges by breaking them down, thus supercharging discovery and innovation in fields essential to technology and national security.
Recent strides in large reasoning models—like OpenAI’s o3 and Google DeepMind’s AlphaProof—demonstrate significant advances over previous generations of large language models, routinely outperforming humans in competitions like the American Invitational Mathematics Examination (AIME). DeepMind’s AlphaEvolve even managed to discover previously unknown solutions to complex math problems. However, experts note a striking difference between models that excel at solving patterned, rule-based competition problems and the creative, experimental process required for genuine mathematical breakthroughs, such as those found in the Millennium Prize Problems.
Tools like PatternBoost by Meta and AlphaEvolve from DeepMind represent an important shift toward collaborative and experimental discovery in mathematics, supporting researchers in exploring unfamiliar territories. While these systems can identify dead ends and suggest new pathways far faster than humans could unaided, they lack the intuitive leaps and creativity at the heart of mathematical invention. Most mathematicians see these technologies, for now, as advance scouts rather than trailblazing pioneers. The hope is that such artificial intelligence tools will continue to sharpen collaboration and accelerate progress, but transforming how mathematics is fundamentally done remains an open question.
