In the traditional view of mathematics, progress is a "bottom-up" process: one starts with axioms and uses formal logic to derive theorems. However, the actual practice of mathematics often begins with a "top-down" leap of intuition. A mathematician notices a pattern in numbers or shapes and suspects a deeper relationship exists before they ever attempt a formal proof.
A growing body of research suggests that this "mathematical intuition" can be systematized and scaled using artificial intelligence. By treating mathematical structures as data and the discovery process as a heuristic search, researchers are transforming mathematics into an experimental science.
From Formal Proof to Theory Formation
The early history of automated mathematics focused almost exclusively on deductive theorem proving—the mechanical verification of a statement already known to be true. In Theorem Proving in Type Theory, Peter B. Andrews and Eve Longini Cohen demonstrated that machines could derive complex proofs, like Cantor’s Theorem, by generating higher-order lambda expressions.
However, Douglas B. Lenat argued in Automated Theory Formation in Mathematics that this neglected the most creative part of the field: identifying which concepts are worth defining in the first place. His "AM" system showed that by applying heuristic rules to basic set theory, a machine could autonomously rediscover natural numbers and Goldbach’s conjecture. This shifted the focus from proving theorems to the inductive formation of new theories.
Mathematics as an Experimental Science
Modern researchers have refined this idea by treating mathematical invariants—properties like the volume of a knot or the rank of an elliptic curve—as "Platonic data." In this framework, machine learning acts as a "top-down" mechanism for discovery.
Yang-Hui He describes this in From the String Landscape to the Mathematical Landscape as using AI as a digital surrogate for intuition. Instead of building a proof from first principles, a researcher uses a neural network to find correlations in vast datasets. If a model can predict a complex topological property with 99% accuracy based on simple inputs, it suggests that a clean, formal relationship must exist, even if it hasn't been discovered yet.
This approach was famously validated by Alex Davies, András Juhász, and colleagues in The signature and cusp geometry of hyperbolic knots. By using attribution techniques on a supervised learning model, they identified a previously unnoticed connection between the 3D geometry of a knot and its 4D signature, leading to the formulation and proof of a new theorem.
The Neural Network as a Mathematical Scout
How does a "black box" neural network lead to a rigorous proof? The key is using predictive accuracy as a heuristic signal. In Machine learning detects terminal singularities, Tom Coates and Alexander M. Kasprzyk used neural networks to predict geometric properties with high precision. This success acted as a "scout," signaling to the researchers that a hidden structural pattern existed. This prompted them to look closer, eventually discovering a new theorem that replaced a complex computational bottleneck with a simple global rule.
Similarly, in Learning to be Simple, researchers found that neural networks could "sense" the abstract property of mathematical simplicity in finite groups. By analyzing which features the model prioritized to achieve high accuracy, the team was able to formulate a new theorem regarding the fixed points of group generators.
Automating the Conjecturing Loop
The ultimate goal of this research direction is to close the loop between pattern recognition and formal verification. Systems like the "Ramanujan Machine," described by Gal Raayoni and Ido Kaminer in The Ramanujan Machine, automate the discovery of elegant formulas for fundamental constants by reverse-engineering numerical sequences.
More recently, the Conjecturing-Proving Loop (CPL) developed by Kazumi Kasaura and others uses Large Language Models to autonomously propose and then formally verify theorems in the Lean programming language. This moves AI beyond being a mere "calculator" or "verifier" and into the role of an Agentic Researcher, capable of managing the entire experimental cycle of mathematics: observing data, forming a conjecture, and establishing a proof.