DARPA expMath: 13 Teams, Goals and 2026 Updates

Quick answer: DARPA’s expMath (Exponentiating Mathematics) is a US research program developing AI collaborators for pure mathematics. It aims to help mathematicians find useful intermediate results, called lemmas, and translate between written mathematics and computer-checkable proofs. Its stated targets: predict and prove lemmas in 90% of evaluation problems and cut solution time 100-fold. DARPA lists 13 university-led teams, selected in March 2026: 11 building the technology and 2 evaluating it.

Sources checked November 22, 2026.

The US Defense Advanced Research Projects Agency (DARPA) funded early work that led to the internet and GPS. In 2025 it set its sights on pure mathematics. Its program expMath, short for Exponentiating Mathematics, aims to build AI that works with mathematicians as a genuine co-author. Here is what the program is, who is involved and what it could change.

expMath at a Glance

Full nameExponentiating Mathematics (expMath)
Run byDARPA (solicited by the Information Innovation Office; now listed under the Information Processing Techniques Office)
Program managerPatrick Shafto
AnnouncedSolicitation HR001125S0010, April 30, 2025; DARPA news story May 5, 2025
TargetsLemmas predicted and proved in 90% of evaluation problems; 100x faster than a graduate student redoing solutions by hand
EvaluationsEvery six months
GoalAn AI co-author “capable of proposing and proving useful abstractions”
Teams selected13 university teams, March 2026 (11 building AI, 2 evaluating it)
KickoffMarch 2026, Arlington, Virginia

Why DARPA Cares About Pure Math

DARPA’s argument is that mathematics drives progress in science and technology, from cryptography to physics, yet the pace of mathematical research has barely changed in a century: one person, a notebook and years of work. AI had shown it could solve olympiad problems; expMath asks whether it can speed up real research, where the hard part is finding the right ideas and intermediate results.

What expMath Is Trying to Build

The program has two technical areas:

  • TA1, the AI co-author. Teams build systems that can (a) decompose a hard problem into smaller, reusable lemmas, and (b) auto(in)formalize: translate natural-language mathematics into formal proofs a computer can check, and back again.
  • TA2, the evaluation. Separate teams measure progress against professional mathematics, such as research papers and graduate textbooks, including studies of whether AI helps people (at least math graduate students) work faster.

The focus is proof-based pure mathematics, not software verification. The idea of “useful abstractions” is key: mathematicians make progress by inventing the right concepts and lemmas, not just by proving the final statement.

A teaching example of a lemma. To prove that \(n(n+1)\) is always even, first prove the lemma “of two consecutive integers, one is even.” Then \(n(n+1)\) has an even factor, so it is even. Breaking a hard goal into steps like this, at research scale, is what expMath’s systems are meant to do. DARPA’s FAQ notes that the human speed comparisons measure the automated systems against people, rather than studying human–AI teamwork itself.

The Teams Selected in 2026

AreaTeams
TA1: building AI co-authors (11)Caltech, Carnegie Mellon, Harvard, Johns Hopkins, MIT, NYU, Princeton, Stanford, UC Berkeley, UCLA, University of Maryland
TA2: evaluation (2)Cornell, UC Irvine

Three teams have published details:

  • UCLA, “ALPHA” (Accelerated Formal Proof Synthesis with Neuro-Symbolic Automation): 5 million dollars over three years, led by Wei Wang with Terence Tao, Andrea Bertozzi, Amit Sahai, Kai-Wei Chang and Nanyun (Violet) Peng. The 5 million dollars supports this one team’s project; it is not expMath’s total budget. It targets PDEs, number theory and complexity theory, using Lean and Isabelle.
  • University of Maryland, “GENIUS”: 2.6 million dollars, led by Mohammad Hajiaghayi with Maryland’s Reza Ghanadan and MIT’s Erik Demaine and Ankur Moitra, building a neuro-symbolic reasoning engine that turns human reasoning steps into reusable modules.
  • NYU: led by Claudio Silva with Chinmay Hegde.

How expMath Fits the 2026 Picture

expMath is a public, university-based effort running alongside fast-moving company research. In 2026, company systems contributed to famous results, from the unit distance conjecture to a Lean proof of the Riemann Hypothesis (see AI math breakthroughs of 2026). expMath’s emphasis on decomposition, formalization and careful evaluation addresses exactly the concerns mathematicians raised: results that are checked, explained and reusable, not just announced. Its tools build on formal verification in Lean and Isabelle.

What Success Would Look Like

  • A mathematician states a goal; the AI proposes a plan of lemmas, proves many of them and flags the ones that need human ideas.
  • Papers can be translated into checked formal proofs quickly, and formal proofs into readable explanations.
  • Measured, not anecdotal, evidence that researchers finish real projects faster.

The solicitation’s targets, 90% of lemmas and a 100-fold speedup, are goals, not achieved results. DARPA calls developing robust ways to measure AI progress on frontier mathematics a key challenge.

Common Misconceptions

“expMath is a military math project.” No. It funds open, university-based research in pure mathematics.

“Its goal is to replace mathematicians.” No. The stated goal is an AI co-author working with people.

“It is the same as company AI math labs.” No. It is a government research program with independent evaluation teams.

Watch: DARPA expMath and the AI Co-Author

If you learn better by watching, these videos are worth your time.

expMath: Exponentiating Mathematics (DARPAtv). DARPA’s own introduction to the program and its goals.
ICM 2026 Public Lectures – Terence Tao (Simons Foundation). Terence Tao, a member of UCLA’s expMath team, on mathematics in the age of AI.

Check Your Understanding

Five quick questions to see how much of this has stuck. Pick an answer and press Submit; you will see a short explanation either way.

Answer each question and press Submit to see your result right away.

1. What does expMath stand for?
Found a bug? Let us know!
2. What is expMath’s stated goal?
Found a bug? Let us know!
3. What does “decomposition” mean in expMath?
Found a bug? Let us know!
4. How many university teams did DARPA select in March 2026?
Found a bug? Let us know!
5. Which Fields Medalist is on UCLA’s expMath team?
Found a bug? Let us know!

Frequently Asked Questions

What is DARPA expMath?

expMath (Exponentiating Mathematics) is a DARPA program announced in April 2025 to accelerate progress in pure mathematics by building AI co-authors that can break problems into lemmas and translate between ordinary math and formal proofs.

What are expMath’s goals?

Its solicitation targets AI that correctly predicts and proves lemmas in 90% of evaluation problems and solves them 100 times faster than a graduate student redoing the solutions by hand, with evaluations every six months.

Who runs expMath?

It is run by DARPA’s Information Processing Techniques Office, with Patrick Shafto as program manager.

Which universities are in expMath?

In March 2026 DARPA selected 13 teams: Caltech, Carnegie Mellon, Harvard, Johns Hopkins, MIT, NYU, Princeton, Stanford, UC Berkeley, UCLA and Maryland to build AI, and Cornell and UC Irvine to evaluate it.

Is Terence Tao part of expMath?

Yes. He is a member of UCLA’s ALPHA team, which received a 5 million dollar, three-year expMath award led by Wei Wang.

What is autoformalization?

Autoformalization is using AI to translate mathematics written in ordinary language into a formal language, such as Lean, that a computer can check. expMath also funds the reverse, turning formal proofs back into readable math.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "DARPA expMath: 13 Teams, Goals and 2026 Updates - Effortless Math"?

No one replied yet.

Leave a Reply