Introduction

Geometry is one of the most challenging fields of mathematics, especially at the Olympiad level. Proving
geometric theorems requires sophisticated reasoning skills and creativity. However, a team of researchers
from Google DeepMind has developed an AI system that can tackle these problems with remarkable
success. The system, called AlphaGeometry, is the first of its kind to prove theorems related to Euclidean
plane geometry more effectively than the average International Mathematical Olympiad (IMO) candidate.
In this article, we will explore the fascinating world of Alpha Geometry, an algorithm developed by
researchers at Google Deep Mind and New York University. This algorithm has made significant
advancements in solving complex math problems, particularly in the field of geometry. We will delve into
what Alpha Geometry is, how it works, and its practical performance. We will also discuss the pros and
cons of this algorithm and explore potential areas of improvement for future development
What is Alpha Geometry?
Alpha Geometry is an algorithm designed to solve challenging math problems, specifically those
encountered in the International Math Olympiads. These problems are known for their high level of
difficulty, often requiring intricate geometric proofs with numerous steps and auxiliary constructions.
Alpha Geometry focuses on solving Euclidean 2D geometry problems, such as demonstrating Parallel lines
or proving distance ratios. It aims to provide elegant and interpretable solutions to these problems, setting
itself apart from previous works in the field.
The Scope of Alpha Geometry
While Alpha Geometry excels in solving Ukian plane geometry problems, it does not extend its capabilities
to other areas of mathematics, such as algebra or calculus. Certain subfields of geometry, including
combinatorial geometry or geometric inequalities, are also not covered by the algorithm. It is essential to
recognize the algorithm's limitations and understand its suitability for specific problem domains.
How Does Alpha Geometry Work?
Alpha Geometry leverages a combination of symbolic learning techniques and a Transformer-based
language model to solve geometry problems. The algorithm follows a step-by-step approach, using
deductive database rules and algebraic rules to make deductions from the given set of hypotheses
The symbolic machine within Alpha Geometry employs deductive database rules to make direct
deductions from the hypotheses. These rules, developed based on previous work, facilitate
straightforward and interpretable deductions. However, deductive database rules alone are insufficient for
solving complex geometry problems.
One of the key features of Geometry is that it can construct new geometric elements, such as points, lines,
or circles, to simplify the problem. The language model predicts which elements are most likely to be
useful, based on the given diagram and statement. The deduction engine then adds these elements to the
diagram and infers more facts about the geometry. This process helps the system to find logical solutions
for complex geometry problems.

To overcome this limitation, Alpha Geometry incorporates algebraic rules. By transforming equalities into a
specific form, the algorithm translates the problem's hypotheses into a system of linear equations. It then
applies Gauss-Jordan elimination techniques to deduce equalities among the variables. This combination
of deductive database and algebraic rules enables Alpha Geometry to make comprehensive deductions
and progress towards reaching the desired conclusion.
An essential aspect of Alpha Geometry is training the large language model. Unlike other machine learning
algorithms, Alpha Geometry does not rely on human-made examples for training. Instead, it randomly
samples hypotheses from an exhaustive list and uses the symbolic machine to make deductions. Through
this process, Alpha Geometry generates artificial triplets of hypothesis-proof-deduction, which serve as its
training data.

Additionally, the large language model is trained to generate auxiliary constructions—a crucial component
of solving complex geometry problems. The model learns from examples where auxiliary constructions are
moved from the hypotheses to the proof, providing it with a guide on constructing proofs with auxiliary
constructions.
Practical Performance of Alpha Geometry
Alpha Geometry's practical performance has been evaluated on various sets of geometry problems. It was
trained on 100 million artificially generated geometry problems, including those involving auxiliary
constructions. testing on the latest 30 problems from the International Math Olympiads, Alpha Geometry
achieved an impressive success rate of approximately 83%.
AlphaGeometry has been tested on the IMO-AG-30 benchmark, which consists of 30 classical geometry
problems adapted from the IMO contests. The system achieved a success rate of 25 out of 30 problems,
outperforming previous state-of-the-art geometry theorem provers, such as GPT-4 and Wu’s method.
Moreover, Geometry’s performance is comparable to that of an average IMO gold medallist, who would
typically solve 23 to 27 problems.

AlphaGeometry also produces human-readable proofs, which enhance the interpretability of its solutions.
The system can explain how it arrived at the answer, using natural language and mathematical symbols.
Furthermore, Alpha can generalize its solutions to more abstract versions of the problems. For instance,
the system discovered a more general form of a translated IMO theorem from 2004, which applies to any
quadrilateral, not just a parallelogram.
It is crucial to note that Alpha Geometry's evaluation focused solely on geometry problems, while
competitors in the Math Olympiads are assessed across all areas of mathematics. Furthermore, the
algorithm demonstrated exceptional performance across a larger set of 231 problems, achieving a score of
98.7%.
Pros and Cons of Alpha Geometry
Pros:

Alpha Geometry can solve geometry problems that are challenging even for highly skilled mathematicians.
The algorithm does not require human-made examples for training, making it highly autonomous.
It provides elegant and interpretable solutions, enhancing the understanding and Clarity of proofs.
Cons:

Alpha Geometry is limited to Ukian plane geometry problems and does not extend its capabilities to other
areas of mathematics.
There is still room for improvement in broadening the algorithm's applicability and performance in solving
more general math problems.
Potential Improvements and Future Directions
To further enhance Alpha Geometry's performance, several improvements can be considered. One
potential avenue is dynamically adding deduction rules during the training phase, incorporating discovered
theorems along the way. This approach could potentially accelerate the deduction process by enabling
more advanced deductions and skipping trivial proof steps. Research exploring this direction could unlock
significant advancements in the algorithm's capabilities
Conclusion
Alpha Geometry represents a remarkable advancement in the field of automated math problem-solving,
particularly in the domain of geometry. Its ability to solve complex problems without human-made
examples is a testament to the power of symbolic learning techniques and Large Language Models. While
Alpha Geometry currently focuses on Ukian plane geometry problems and has certain limitations, it sets
the stage for further developments in automating mathematical reasoning and problem-solving.

Further studies and improvements will pave the way for Alpha Geometry to expand its capabilities and
potentially solve a broader range of math problems. As research in this field continues, we can eagerly
look forward to future breakthroughs in automated mathematics and problem-solving.
AlphaGeometry
Google’s DeepMind team published a new approach to solving complex geometry problems called
AlphaGeometry.
This new approach performs really closely to the average level of International Mathematical Olympiad
(IMO) gold medalists. IMO gold medalists on average solve ~26 problems within the standard time limit,
while AlphaGeometry can solve 25.
How it works
AlphaGeometry uses two major components in its system, a Language Model and a Symbolic Engine.

Language models, like ChatGPT or Gemini (Bard), are indeed powerful for generating stuff, but they lack
the precision and accuracy in order to produce a working answer that will solve a complex geometry
problem.
Symbolic engines on the other hand, are more logical and precise, and are able to solve complex geometry
problems, but they are too slow as they try to find every fact and process that can solve problems

Oversimplified, but you get the point.

AlphaGeometry’s approach uses the pros of both components. The symbolic engine solves the problem
logically, and instead of directly searching for all possible facts and steps, the language model provides the
facts instead.
This significantly reduces the time needed by the symbolic engine to solve the problem. After attempting
to solve it, the symbolic engine will send feedback to the language model. The cycle goes on until the
problem is solved.
In a paper published today in Nature, we introduce AlphaGeometry, an AI system that solves complex
geometry problems at a level approaching a human Olympiad gold-medalist - a breakthrough in AI
performance. In a benchmarking test of 30 Olympiad geometry problems, AlphaGeometry solved 25
within the standard Olympiad time limit. For comparison, the previous state-of-the-art system solved 10 of
these geometry problems, and the average human gold medalist solved 25.9 problems.

In our benchmarking set of 30 Olympiad geometry problems (IMO-AG-30), compiled from the Olympiads
from 2000 to 2022, AlphaGeometry solved 25 problems under competition time limits. This is approaching
the average score of human gold medalists on these same problems. The previous state-of-the-art
approach, known as “Wu’s method”, solved 10.

AI systems often struggle with complex problems in geometry and mathematics due to a lack of reasoning
skills and training data. AlphaGeometry’s system combines the predictive power of a neural language
model with a rule-bound deduction engine, which work in tandem to find solutions. And by developing a
method to generate a vast pool of synthetic training data - 100 million unique examples - we can train
AlphaGeometry without any human demonstrations, sidestepping the data bottleneck.
AlphaGeometry adopts a neuro-symbolic approach
AlphaGeometry is a neuro-symbolic system made up of a neural language model and a symbolic deduction
engine, which work together to find proofs for complex geometry theorems. Akin to the idea of “thinking,
fast and slow”, one system provides fast, “intuitive” ideas, and the other, more deliberate, rational
decision-making.

Because language models excel at identifying general patterns and relationships in data, they can quickly
predict potentially useful constructs, but often lack the ability to reason rigorously or explain their
decisions. Symbolic deduction engines, on the other hand, are based on formal logic and use clear rules to
arrive at conclusions. They are rational and explainable, but they can be “slow” and inflexible - especially
when dealing with large, complex problems on their own.
AlphaGeometry’s language model guides its symbolic deduction engine towards likely solutions to
geometry problems. Olympiad geometry problems are based on diagrams that need new geometric
constructs to be added before they can be solved, such as points, lines or circles. AlphaGeometry’s
language model predicts which new constructs would be most useful to add, from an infinite number of
possibilities. These clues help fill in the gaps and allow the symbolic engine to make further deductions
about the diagram and close in on the solution.

AlphaGeometry solving a simple problem: Given the problem diagram and its theorem premises (left),
AlphaGeometry (middle) first uses its symbolic engine to deduce new statements about the diagram until
the solution is found or new statements are exhausted. If no solution is found, AlphaGeometry’s language
model adds one potentially useful construct (blue), opening new paths of deduction for the symbolic
engine .This loop continues until a solution is found (right). In this example, just one construct is required.

AlphaGeometry solving an Olympiad problem: Problem 3 of the 2015 International Mathematics Olympiad
(left) and a condensed version of AlphaGeometry’s solution (right). The blue elements are added
constructs. AlphaGeometry’s solution has 109 logical steps.