AI and Mathematical Discovery: The Lean Conjecturer and Automated Reasoning

For centuries, the frontier of mathematical discovery has been seen as the exclusive domain of human intuition. It is a world of sudden insights, aesthetic appreciation of elegant structures, and the painstaking labor of constructing rigorous proofs, often spanning hundreds of pages. The classic image of the mathematician—solitary, grappling with abstract concepts on a blackboard—emphasizes the deeply personal, almost artistic nature of the field. However, a profound shift is underway. The arrival of potent Large Language Models (LLMs) and the maturation of formal verification tools, most notably the Lean theorem prover, are reshaping how mathematics is done. We are entering an era of neuro-symbolic collaboration, where the creative, probabilistic leaps of artificial intelligence are anchored by the absolute, binary certainty of symbolic logic. This synergy is not only accelerating the verification of complex theorems but is also opening the door to the automated generation of novel, verifiable scientific hypotheses.

At the heart of this revolution is a fundamental tension in mathematics. Human intuition is powerful but fallible. A brilliant conjecture might be true, but the proof might contain a subtle flaw that goes unnoticed for years. As mathematical theories become more abstract and proofs more complex, the burden on human peer review has become immense, sometimes unsustainable. It took years for the mathematical community to fully verify the proof of Fermat's Last Theorem, and many other critical results exist in a state of 'accepted but not totally formalized' certainty. This is where automated reasoning systems offer a transformative solution, providing a level of trust that human review alone struggle to achieve.

The Guardrails of Rigor: The Lean Theorem Prover

To understand how AI is assisting mathematicians, we must first understand the tool that provides the rigor. Lean is an open-source interactive theorem prover (ITP) and a programming language, primarily developed at Microsoft Research. Unlike computer algebra systems (such as Mathematica or Maple), which manipulate mathematical expressions but cannot construct proofs, Lean is designed to verify that every step of a mathematical argument is logically sound. It is built on a foundation of dependent type theory, a robust logical framework where mathematical statements are treated as types, and a proof is a program that constructs an object of that type.

When a mathematician uses Lean, they are not just typing out equations. They are translating their informal human reasoning into a precise, machine-readable language. This process, known as formalization, is extremely rigorous. Lean's 'kernel' is a small, carefully checked piece of code that does only one thing: verify that a given proof term matches the type of the theorem statement. If Lean's kernel accepts a proof, the community can have near-absolute confidence in its correctness, grounded in the fundamental axioms of mathematics.

The monumental effort of the "Liquid Tensor Experiment" demonstrates the power of Lean. In 2020, Fields Medalist Peter Scholze challenged the mathematical community to formalize a crucial, foundational theorem in his new theory of condensed mathematics. Scholze was "99.9% sure" of the theorem's correctness, but its proof was immensely complex, weaving together functional analysis, homological algebra, and convex geometry. He recognized that traditional peer review might not be sufficient for such a dense argument. A global team of mathematicians, led by Johan Commelin, spent over a year formalizing the theorem in Lean. The project was successful, uncovering minor subtleties in Scholze’s initial argument and simplifying the overall proof structure. More importantly, it provided 100% verification. This success signaled that computer verification of cutting-edge research in abstract mathematics was not only possible but necessary.

The Catalyst of Creativity: Large Language Models

While Lean provides the rigor, it is notoriously difficult to use. Writing formal proofs is labor-intensive, requires specialized knowledge, and often feels like fighting the machine over trivial logical steps that a human would take for granted. This is the gap that Large Language Models are now filling. LLMs, like those powering advanced conversational agents, excel at pattern recognition, linguistic analogy, and logical synthesis within unstructured text. They have 'read' the entirety of arXiv and virtually all digitized mathematical literature. They 'understand' the grammar of mathematical definitions, the structure of lemmas, and the narrative flow of a standard proof sketch.

The synergy between LLMs and Lean creates a neuro-symbolic agent, combining neural network intuition with symbolic logic verification. This collaboration takes several powerful forms:

Autoformalization: The goal is to train an LLM to take an informal proof written in a PDF or textbook and translate it directly into Lean code. While a perfect "math-to-code" translator does not yet exist, LLMs are increasingly capable of generating high-quality Lean proof sketches and filling in routine lemma gaps, significantly reducing the labor of formalization.
Tactic Generation: In Lean, proofs are often constructed using 'tactics'—small programs that automate routine reasoning steps, such as 'simplify this expression' or 'apply induction on this variable.' LLMs can be integrated into the Lean environment to function as an intelligent 'Copilot,' suggesting the next likely tactic or even a sequence of tactics to bridge the gap between a goal and the available premises. Systems are emerging that combine tree-search algorithms with LLMs to explore millions of potential tactic combinations, a form of neuro-symbolic proof search.
Proof Repair: When Lean's verifier rejects a step, it provides an error message. LLMs can be trained to interpret these error messages and suggest corrections, whether they are simple syntactic fixes or alternative logical strategies, iterating on a proof until the kernel accepts it.

The Lean Conjecturer and Automated Discovery

The most provocative aspect of AI in mathematics is not just proving existing theorems, but discovering new ones. This requires overcoming the "triviality trap." It is trivial for an AI to generate an infinite number of true but uninteresting statements (e.g., '1+1=2,' '1+2=3,' and so on). The true challenge is creating novel, interesting, and provable conjectures.

Projects like "The Lean Conjecturer" (LeanConjecturer) are pioneering this open-ended exploration. These systems typically use a dual-agent architecture: a "Conjecturer" and a "Prover." The Conjecturer, powered by an LLM, is trained on a massive mathematical library like Lean’s Mathlib. It learns the conceptual style, object relationships, and aesthetic qualities of 'interesting' theorems. It then proposes a new theorem statement, often by synthesizing concepts from disparate mathematical domains or analogizing existing properties to new mathematical structures.

This statement is then immediately fed to the symbolic Prover (Lean). Lean first checks for syntactic validity. If valid, Lean attempts to prove it using its automated tools and the LLM assistant. If Lean succeeds, a new, verified theorem is discovered. If Lean fails, the system analyzes why it failed—whether the statement was provably false, potentially true but too difficult, or mathematically coherent but fundamentally trivial. This feedback loop is used to iteratively refine the LLM Conjecturer, training it to avoid obvious trivialities and move toward "useful novelty."

Early iterations of such systems have successfully rediscovered important theorems in topology and set theory and are beginning to generate genuinely non-trivial conjectures. For instance, a Lean Conjecturer might observe a pattern in algebraic geometry and hypothesize a similar, verified pattern in representation theory, forging an automated connection that no human had previously considered. This shifts the role of the human mathematician from the sole source of creativity to a strategic guide, organizing the search space and selecting the most interesting automated breakthroughs for further human development.

From Mathematics to Verifiable Scientific Hypotheses

The potential of neuro-symbolic reasoning extends far beyond abstract mathematics. In broader science, the discovery of novel hypotheses is often bottlenecked by the impossibility of synthesis. A modern materials scientist or biologist cannot possibly read the millions of papers relevant to their field. LLMs have already proven useful for synthesizing literature and identifying surprising correlations—such as linking a known drug with an obscure disease pathway. However, like LLMs in mathematics, LLMs in science suffer from the hallucination problem: they can generate plausible-sounding but physically or chemically impossible hypotheses.

By connecting scientific LLMs to domain-specific symbolic verifiers (essentially, 'Leans for science'), we can create trustworthy discovery engines. For instance, in drug discovery, an LLM might propose a novel molecular structure designed to bind to a specific protein. This proposal would not be accepted on face value but would be fed into a chain of symbolic tools: first, a chemical grammar checker to verify stability; second, a physics-based simulation (like molecular dynamics) to verify binding affinity; and third, an automated toxicology model to verify safety. The system would only output a hypothesis if it passed these verification guardrails, providing an "auditable reasoning trace" for the proposed discovery.

In materials science, this approach could link an LLM (literature synthesis, novelty) with specialized symbolic engines (crystallographic checkers, density functional theory simulators). A system might propose a novel battery cathode material by combining elements from disparate studies and verifying, via symbolic simulation, that the proposed crystal structure is both synthesizable and possesses the desired electrochemical properties. This transformation moves scientific discovery from probabilistic prediction (a "guess") toward verifiable hypothesis generation (a trusted suggestion).

A New Epistemology of Discovery

The combination of Large Language Models and formal verification tools like Lean marks a fundamental shift in the epistemology of mathematical and scientific discovery. It challenges the traditional view that discovery must originate solely within human intuition. Instead, we are seeing the rise of a hybrid intelligence model. In this model, neural networks provide the scale, speed, and creative synthesis required to explore a vast space of possibilities, while symbolic systems provide the unyielding rigor required to validate truth.

This does not render human insight obsolete. On the contrary, it amplifies it. Mathematicians and scientists are now liberated from the drudgery of low-level formalization and routine verification. Their role is evolving into that of editors, architects, and curators of discovery. They define the "interesting" problems, set the axioms, guide the automated search space, and refine the AI's "aesthetic" understanding of novelty. The output of these systems—whether a formal proof of an established theorem or a novel, verified scientific hypothesis—carries a stamp of certainty that was previously unattainable. This trusted collaboration promises to unlock a future where automated reasoning, anchored in symbolic rigor, accelerates the pace of understanding and expands the very definition of discoverable truth.