DARPA expMath targets AI coauthors for advanced math

The U.S. Defense Advanced Research Projects Agency (DARPA) has launched a new research effort called expMath to bring artificial intelligence directly into advanced mathematical work. Officials described the as developing “AI coauthors” that can help mathematicians break difficult questions into smaller, workable parts and speed up progress in areas tied to national security.

According to the program description, expMath is focused on mathematical discovery that underpins fields such as cryptography and systems modeling. The central promise is not simply faster calculation, but a workflow where AI systems contribute to the structure of an investigation by organizing subproblems and helping researchers move through complex chains of reasoning.

DARPA’s expMath plan and where it is aimed

DARPA’s framing connects advanced mathematics to security-relevant applications, including cryptography. In practice, that means the program is aimed at accelerating research that can influence how information is protected and how complicated systems are represented and analyzed.

The initiative is presented as an attempt to move AI tools closer to the day-to-day process of mathematical research, where progress often depends on decomposing broad questions into intermediate lemmas, checking subtle logical steps, and managing many partial results at once.

Large Reasoning Models and the recent capability jump

The timing of expMath is linked to recent gains in so-called Large Reasoning Models (LRMs). These systems have shown stronger performance on standardized, high-level mathematics tasks, with reports highlighting results comparable to silver medalists at the International Math Olympiad .

Unlike earlier large language models that were often characterized by pattern-matching limitations, LRMs are associated with step-by-step reasoning methods. The source material describes this as a meaningful shift in how these models approach complex problems, especially in contexts where a solution depends on multi-stage logic rather than a single short response.

The unresolved gap: contest solutions vs original research Even with improved benchmark performance, experts cited in the source material emphasize a clear separation between solving competition-style questions and producing original, exploratory research. The concern is that strong results on established problem formats do not automatically translate into the open-ended work of discovering new mathematics.

Current models are described as effective at applying known techniques, but their ability to generate ideas that go beyond existing human knowledge is not established. The source material notes that this is a crucial uncertainty for any effort that aims to create true “coauthors” rather than automated assistants that only execute routine steps.

In that framing, expMath’s outcome hinges on whether AI systems can move from completing well-defined tasks to supporting the creative and non-linear patterns that professional mathematical inquiry often requires. For DARPA, the program is positioned as a test of whether automated reasoning can be advanced into a tool that contributes to the research process itself, not just to the final answers.