AlphaProof and AlphaGeometry 2 Explained

Quick answer: AlphaProof is a Google DeepMind system for finding mathematical proofs in Lean, a language whose proof checker verifies every step; it learns by reinforcement learning from verified outcomes. AlphaGeometry solves geometry problems with a language model that suggests useful constructions and a symbolic deduction engine that does the rest. At the 2024 International Mathematical Olympiad, AlphaProof and the upgraded AlphaGeometry 2 solved four of six problems for 28 of 42 points, silver-medal standard. People formalized the problems first, and some solutions took days of computing.

Updated November 22, 2026.

In July 2024, an AI system earned a silver-medal score at the International Mathematical Olympiad (IMO) for the first time. It was not one program but two: AlphaGeometry for geometry and AlphaProof for everything else. They work in very different ways, and both show a key idea in AI math: pair a creative model with a strict checker.

AlphaProof and AlphaGeometry at a Glance

AlphaGeometryAlphaProof
AnnouncedJanuary 2024 (Nature); AlphaGeometry 2 in July 2024July 2024; Nature paper November 12, 2025
ProblemsOlympiad geometryAlgebra, number theory, combinatorics
MethodLanguage model plus symbolic deduction engineReinforcement learning on proofs in Lean
Training data100 million synthetic examples generated by the systemAbout 1 million informal problems auto-translated into about 80 million formal Lean statements
CheckingEvery step follows from geometry rulesLean checks every step
Headline result25 of 30 benchmark problems (AlphaGeometry); 42 of 50, or 84%, on a broader 2000–2024 set (AlphaGeometry 2, up from 54%)3 of the 5 non-geometry problems at IMO 2024

How AlphaGeometry Works

Geometry proofs often need one clever extra line or point, an auxiliary construction, after which the rest follows by routine steps. AlphaGeometry splits the work:

  1. A symbolic deduction engine derives every fact it can from the diagram using geometry rules (angles, similar triangles, cyclic quadrilaterals).
  2. If it gets stuck, a language model suggests an auxiliary construction, such as “add the midpoint of \(AB\).”
  3. The engine tries again with the new point, repeating until it reaches the goal.

Its big trick was data. Human-written geometry proofs are scarce, so DeepMind generated 100 million synthetic theorems and proofs by drawing random diagrams and recording what could be deduced. In the January 2024 Nature paper by Trieu Trinh and colleagues, AlphaGeometry solved 25 of 30 IMO geometry problems from 2000 to 2022, close to the average gold medalist’s 25.9. AlphaGeometry 2 (described in a February 2025 paper) used a stronger Gemini-based model, a faster symbolic engine and search trees that share findings. On a broader benchmark of IMO geometry problems from 2000 to 2024 (45 problems formalized as 50 cases), it solved 42 of 50, or 84%, compared with 54% for the original system.

A teaching example of an auxiliary construction. To prove that the base angles of an isosceles triangle \(ABC\) with \(AB = AC\) are equal, add the midpoint \(M\) of \(BC\). Then triangles \(ABM\) and \(ACM\) have three equal sides, so they are congruent and \(\angle B = \angle C\). Finding the point \(M\) is the creative step; AlphaGeometry’s language model plays that role.

How AlphaProof Works

AlphaProof works in Lean, a language in which a computer checks every step of a proof. That makes it possible to learn the way AlphaZero learned chess: try, get checked, improve.

  1. Translate. A language model translated about a million informal math problems into Lean, producing about 80 million formal statements to practice on.
  2. Search. For each problem, its neural network proposes tactics, commands that advance a proof, and a search procedure explores alternatives.
  3. Learn. Lean confirms or rejects each attempt. Successful proofs become training data, and the system gets stronger.
  4. Practice at test time. On a hard problem, it generates many related variants and trains on them while solving; at IMO 2024 this took two to three days per solution.

Because every final proof is checked by Lean, AlphaProof’s answers cannot contain a hidden logical gap, which is why its results needed no grader to trust the logic.

Results at IMO 2024

ProblemAreaSolved by
P1AlgebraAlphaProof
P2Number theoryAlphaProof
P3CombinatoricsNot solved
P4GeometryAlphaGeometry 2, in 19 seconds after formalization
P5CombinatoricsNot solved
P6Algebra (hardest; only 5 contestants solved it)AlphaProof

The total was 28 of 42, silver-medal standard; the gold cutoff that year was 29. There were important caveats: people formalized the problems first (into Lean for AlphaProof and into a specialized geometry language for AlphaGeometry 2), and AlphaProof’s solutions took two to three days of computing, versus the contestants’ 4.5 hours per paper. The result is a medal-equivalent score, not an awarded medal. The proofs were scored by former IMO medalists Timothy Gowers and Joseph Myers.

What Came Next

In July 2025, Google DeepMind’s Gemini Deep Think reached gold-medal level, 35 of 42, working directly in natural language within the time limits, without a Lean translation step. OpenAI reported the same score with an experimental model. The formal approach did not disappear: Lean checking became central to 2026’s research results, from Fermat’s Last Theorem to the Riemann Hypothesis. See AI at the IMO and the AI in math timeline.

Can You Use AlphaProof or AlphaGeometry?

Partly. The original AlphaGeometry repository includes research code and model checkpoints. AlphaGeometry 2’s public repository provides its symbolic engine and examples, but not the full Gemini-based system. The AlphaProof Nature paper directs researchers to apply for hosted access. On November 10, 2026, Google DeepMind also introduced AlphaProof Code, which uses AI agents and Lean to verify software against human-approved specifications; it is a separate system from the olympiad prover described here (see formal verification).

Why These Systems Matter

  • Checkers enable learning. Because Lean or a geometry engine verifies every step, the AI can practice on millions of problems without human grading.
  • Synthetic data works. AlphaGeometry showed that a system can create its own training data in a field where human examples are rare.
  • Creativity plus rigor. A neural model suggests ideas; a strict checker keeps only correct ones. The same pattern drives AlphaEvolve.

Common Misconceptions

“AlphaProof competed like a student.” No. Its problems were translated into Lean by people, and it used far more time than contestants.

“AlphaGeometry understands geometry like a person.” It combines pattern-based suggestions with rule-based deduction; it does not see diagrams the way we do.

“Formal proof systems are obsolete since natural-language AI won gold.” No. Formal checking is how many 2026 research results were verified.

Watch: How AlphaGeometry and AlphaProof Work

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

The AI that solved IMO Geometry Problems (3Blue1Brown and Aleph 0). A beautifully animated explanation of how AlphaGeometry works.
Improving on AlphaProof: IMO 2024 Problem 2 in Lean 4 (David Renshaw). A look at what an AlphaProof-style Lean proof of an IMO problem looks like.

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 score did AlphaProof and AlphaGeometry 2 get at IMO 2024?
Found a bug? Let us know!
2. What language does AlphaProof write proofs in?
Found a bug? Let us know!
3. What does AlphaGeometry’s language model suggest?
Found a bug? Let us know!
4. How many synthetic examples was AlphaGeometry trained on?
Found a bug? Let us know!
5. Which IMO 2024 problem did AlphaGeometry 2 solve?
Found a bug? Let us know!

Frequently Asked Questions

What is AlphaProof?

AlphaProof is a Google DeepMind system that learns to prove math statements in the Lean proof language using reinforcement learning. At IMO 2024 it solved three of the five non-geometry problems, including the hardest one. Its Nature paper was published on November 12, 2025.

What is AlphaGeometry?

AlphaGeometry is a Google DeepMind system that pairs a language model with a symbolic geometry engine. The model suggests auxiliary constructions and the engine deduces the rest. It solved 25 of 30 benchmark IMO geometry problems in January 2024; AlphaGeometry 2 solved 42 of 50 (84%) on a broader 2000–2024 set.

Did AlphaProof win a gold medal?

No. With AlphaGeometry 2, it scored 28 of 42 at IMO 2024, silver-medal standard, one point below gold. Gold-level AI scores came in 2025 from Gemini Deep Think and an OpenAI model.

How is AlphaProof different from ChatGPT?

AlphaProof writes proofs in Lean, where a computer checks every step, and learns from that feedback. Chatbots write natural-language answers that are not automatically checked.

Can I use AlphaProof or AlphaGeometry?

Partly. AlphaGeometry’s research code and checkpoints are public, AlphaGeometry 2’s symbolic engine is public, and researchers can apply for hosted AlphaProof access. AlphaProof Code, announced November 10, 2026, is a separate system for verifying software.

Keep Exploring

Sources and Further Reading

Related to This Article

What people say about "AlphaProof and AlphaGeometry 2 Explained - Effortless Math"?

No one replied yet.

Leave a Reply