Curiosity · AI Model
DeepMind AlphaProof
AlphaProof is Google DeepMind's reinforcement-learning-trained theorem prover, announced in July 2024 when — together with AlphaGeometry 2 — it solved four of six 2024 International Mathematical Olympiad problems for the first silver-medal equivalent performance by an AI system. AlphaProof is a Gemini-based language model fine-tuned to write and verify formal proofs in the Lean 4 proof assistant, using a self-play-style loop where the model explores proof candidates and a verifier supplies reward signals.
Model specs
- Vendor
- Google DeepMind
- Family
- AlphaProof
- Released
- 2024-07
- Context window
- 32,768 tokens
- Modalities
- text, code
Strengths
- First AI to reach silver-medal IMO performance
- Provably correct outputs — Lean verifier catches errors
- Demonstrates viable RL loop on symbolic math
- Pairs with AlphaGeometry 2 for geometry problems
Limitations
- Not publicly released — no API, no open weights
- Proof search is compute-heavy (hours per IMO problem)
- Lean-only — does not emit natural-language proofs
- Weak outside competition-style problems
Use cases
- Formal theorem proving in Lean 4
- Math-olympiad-style problem solving
- Augmenting human mathematicians in formalisation
- Research into neuro-symbolic reasoning
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| IMO 2024 problems solved | 4 / 6 (silver-medal equivalent, w/ AlphaGeometry 2) | 2024-07 |
| miniF2F test | new state-of-the-art | 2024-07 |
Frequently asked questions
What is AlphaProof?
Google DeepMind's reinforcement-learning theorem prover that operates in the Lean 4 proof assistant, trained on millions of synthetic proofs and human libraries.
Did AlphaProof actually win an IMO medal?
Informally, yes — when combined with AlphaGeometry 2 it solved 4 of 6 problems from IMO 2024, matching silver-medal-level human performance.
Can I use AlphaProof?
Not yet — it remains an internal research system. DeepMind has published technical details but no public API or weights.
Sources
- DeepMind — AI achieves silver-medal IMO — accessed 2026-04-20
- International Mathematical Olympiad 2024 — accessed 2026-04-20