The AI that solved IMO Geometry Problems | Guest video by @Aleph 0

Original Media · 14m · English · YouTube · 3Blue1Brown

TLDR

Google DeepMind's Alpha Geometry solves International Mathematical Olympiad geometry problems by combining a non-AI logical engine (a deductive database of geometry rules plus algebraic reasoning) with a language model that proposes auxiliary constructions. The logical engine alone solved 18 out of 30 problems, near bronze medal level. The language model was trained on synthetic data generated by randomly creating diagrams, deducing all true statements with the logical engine, then erasing parts to create problems requiring those erased parts as auxiliary constructions. The full system solved 25 out of 30 problems, surpassing silver medalists, demonstrating a general creative-logical strategy applicable beyond geometry.

For you

PRO members can sign in and request more information — or their own version of this digest, tuned to what they care about.

Sign in to unlock

Surprising Insights

•

A 25-year-old non-AI technique (DD+AR) already solved 18 out of 30 IMO geometry problems, reaching near bronze medal level, despite IMO problems being considered extremely hard and typically requiring human ingenuity.

•

The language model was trained entirely on synthetic data generated by the logical system itself, not on human solutions, which is counterintuitive because one might assume AI needs human examples to learn creativity.

•

The synthetic data generation method works backwards: it starts with a complete diagram, deduces all true statements, then erases parts to create a problem where the erased parts become the required auxiliary constructions.

Key takeaways

•

A purely logical, non-AI system (DD+AR) with 75 geometry rules and linear equation solving solved 18 out of 30 IMO geometry problems, reaching near bronze medal performance without any machine learning.

•

The fundamental limitation of logical geometry solvers is generating auxiliary constructions—adding extra lines or shapes not in the original diagram—which creates an infinite search space that brute force cannot handle.

•

Alpha Geometry overcomes this by integrating a language model trained to propose auxiliary constructions with the DD+AR logical engine, alternating between creative suggestions and rigorous deductions.

•

The language model was trained on synthetic data generated by randomly plotting geometric configurations, using DD+AR to derive all true statements, then erasing elements to create problems whose solutions require reconstructing those erased elements as auxiliary constructions.

•

This synthetic data generation produced hundreds of millions of proof examples, including 9 million requiring at least one auxiliary construction, with the most complex proof having 247 steps and two auxiliary constructions.

•

The combined system solved 25 out of 30 IMO geometry problems, surpassing silver medalist performance, demonstrating that pairing a creative AI module with a logical deduction engine can achieve high-level mathematical reasoning.

•

The strategy of combining creative generation with logical verification is a general problem-solving approach applicable beyond geometry to domains like science, medicine, and engineering.

•

Recent developments show multiple groups achieving gold-level performance on all IMO problems using natural language models without domain-specific languages, indicating rapid progress in AI for mathematical reasoning.

Notable quotes

“The most surprising part for me is not that AI managed to solve these problems, but it's what happened even before the AI showed up.”

▸ Notable quotes · 0:41

“A 25-year-old technique with no AI at all, just using logic and pure equation solving, was already able to solve 18 out of 25 problems.”

▸ Notable quotes · 0:48

“DD can't solve equations.”

▸ Notable quotes · 4:56

“The true gems in the field always feel like adding in the right idea seemingly came out of left field.”

▸ Notable quotes · 8:19

“Put simply, the language model is the creative brain thinking of clever auxiliary constructions. On the other hand, DD plus AR is the logical brain, using pure logic to deduce new facts from known ones.”

▸ Notable quotes · 9:34

“Alpha geometry can think creatively, but also very logically. This is exciting because this general strategy, using creativity and logic together, is by no means specific to geometry.”

▸ Notable quotes · 12:15

Chapters

•

0:00 — Introduction and Non-AI Baseline: The video introduces Alpha Geometry, an AI that solves IMO geometry problems, and reveals that a non-AI logical system already solved 18 out of 30 problems, near bronze medal level.

•

1:33 — Deductive Database (DD) Approach: Explains how a deductive database of 75 geometry rules, encoded in a specialized geometry language, can prove theorems by chaining facts, solving 7 out of 30 IMO problems on its own.

•

4:22 — Adding Algebraic Reasoning (AR): Shows that DD cannot solve equations; adding algebraic reasoning (linear equation solving) and alternating DD and AR raises performance to 14 out of 30, and with human-coded heuristics to 18 out of 30.

•

6:59 — The Auxiliary Construction Challenge: Describes why logical systems struggle with auxiliary constructions—adding extra lines or shapes—which creates an infinite search space, and why this is the key bottleneck for solving hard geometry problems.

•

9:34 — Alpha Geometry Architecture: Creative + Logical Brains: Details the Alpha Geometry loop: a language model proposes auxiliary constructions, DD+AR deduces consequences, and the process repeats until the problem is solved or time runs out.

•

10:03 — Synthetic Training Data Generation: Explains how training data was generated by randomly plotting points, using DD+AR to deduce all true statements, then erasing parts to create problems requiring auxiliary constructions, yielding hundreds of millions of synthetic proofs.

•

11:41 — Results and Broader Implications: Reports the final result of 25 out of 30 problems solved, surpassing silver medalists, and argues that the creative-logical strategy generalizes to other domains like science and engineering.

•

12:48 — Host Outro: Series Context and Future Work: The host notes this is a guest video by Aditya Chakravarti, mentions recent gold-level AI performance on all IMO problems using natural language, and solicits stories about AI in math research.

Speakers

•

Aditya Chakravarti — Guest presenter

•

Host — Channel host

⋮