Google DeepMind rolls out AlphaProof AI model, which specializes in math reasoning, and AlphaGeometry 2, an updated version of a model focused on geometry
Bloomberg
Context & Ripple Effects
DeepMind had already positioned geometry as a demanding test case: its earlier [[a:848367|AlphaGeometry system was reported to solve Olympiad geometry problems near gold-medalist level]]. AlphaProof and the updated geometry model extend that research line from a single benchmark domain toward a broader math-reasoning portfolio.
The release matters because formal mathematics offers unusually clear ways to evaluate whether AI can generate and verify multi-step reasoning, rather than merely produce fluent answers.
First-order effects
Google DeepMind gains two named systems focused on mathematical reasoning: AlphaProof for math and AlphaGeometry 2 for geometry.
Researchers and evaluators get additional specialized models to compare against human-style contest-math performance and other reasoning systems.
Second-order effects
Rival AI labs face added pressure to show reasoning progress on verifiable mathematical tasks, not only on broad conversational or coding benchmarks.
Math and geometry benchmarks become more consequential for model assessment, because specialized systems can expose gaps hidden by general-purpose evaluations.
Third-order effects
If specialized reasoning models continue to improve, AI development may increasingly pair general models with domain-specific systems for tasks where answers can be formally checked.
Mathematics could become a durable proving ground for AI capability claims, though strong performance on contest-style problems does not by itself establish broad real-world reasoning ability.
The trend: AI labs are treating verifiable scientific and mathematical domains as both capability benchmarks and targets for specialized models.
We have long pioneered the use of these types of neuro-symbolic systems starting with AlphaGo in 2016, through to AlphaZero. We'll be bringing all the goodness of AlphaProof and AlphaGeometry 2 to our mainstream #Gemini models very soon. Watch this space!
We're presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈 It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 https://dpmd.ai/... [image]
Paul Christiano on how he would update if an AI got gold in any International Math Olympiad by the end of 2025. Today, AlphaProof came within one point of achieving that gold medal. [image]
A “bridge” from LLMs to the AlphaZero learning algorithm was created to give an AI reasoning capabilities that match the world's most elite mathematicians. It was trained in only weeks, and it - solved one problem in 19 seconds, and - took 3 DAYS to solve others A pre-trained [im…
Paul Christiano and I previously worked hard to pin down concrete disagreements; one of our headers was that Paul put 8% probability on “AI built before 2025 IMO reaches gold level on it” and I put “at least 16%”.
Google just dropped an elite AI mathematician. It's a neuro-symbolic system that formalizes problems into Lean, a formal language, with a fine-tuned Gemini and uses AlphaZero-style search to solve them. Solves 4/6 on IMO 2024, a Silver medal. The future of Math be AI-assisted. [i…
Very impressive result from @GoogleDeepMind! They convert hard math problems into the formal reasoning language Lean, and then use an AlphaZero-style approach to find solutions. This, combined with their earlier work on AlphaGeometry, scores a silver medal at the IMO. Congrats to
Math programming languages like Lean allow answers to be formally verified. But their use has been limited by a lack of human-written data available. 💡 So we fine-tuned a Gemini model to translate natural language problems into a set of formal ones for training AlphaProof. [image…
Advanced mathematical reasoning is a critical capability for modern AI. Today we announce a major milestone in a longstanding grand challenge: our hybrid AI system attained the equivalent of a silver medal at this year's International Math Olympiad!
For non-geometry, it uses AlphaProof, which can create proofs in Lean. 🧮 It couples a pre-trained language model with the AlphaZero reinforcement learning algorithm, which previously taught itself to master games like chess, shogi and Go. https://dpmd.ai/... [image]
Our system had to solve this year's six IMO problems, involving algebra, combinatorics, geometry & number theory. We then invited mathematicians @wtgowers and Dr Joseph K Myers to oversee scoring. It solved 4️⃣ problems to gain 28 points - equivalent to earning a silver medal. ↓ …