How to Build an AI Mathematician

(Toward a Multimodal Theory of Insight)

In the last decade, we learned how to make models that perform mathematics — but not yet models that invent mathematics.

Large Language Models now pass Olympiad problems and solve many exercises from graduate textbooks. They can manipulate formal objects and navigate complex chains of reasoning. Yet mathematicians do not feel replaced. They do not even feel imitated.

Because mathematics is not only about deduction. It is about insight — the emergence of a new representation, a new concept, a new shortcut, a new question.

To move from competence to creativity, an AI system must develop what human mathematicians rely on every day: the ability to jump across representations, sense which ideas matter, balance risk with rigor, and transform a search problem into a structural one.

In short: a mathematician, human or artificial, is a system that knows when to stay local and when to leap.

This article proposes a framework for building such a system.

It draws from neuroscience, topology, cognitive psychology, graph theory, and the architecture of world models and LLMs — but grounds everything in a simple idea:

Insight is not one mechanism. Insight emerges at the intersection of multiple modes of intelligence.

Let’s explore the pillars of this multimodal model.

---

⭐ 1. Exploration: The Dopaminic Principle

Insight begins where expectation breaks

In neuroscience, dopamine is not the molecule of pleasure; it is the molecule of prediction error. Its function is to signal: “This pathway is surprising — explore it.”

Mathematical discovery operates exactly this way. A mathematician encounters a contradiction, a strange coincidence, a pattern slightly too regular to be accidental — and the anomaly becomes a compass.

Modern AI systems lack this internal asymmetry. LLMs do not prefer surprising directions; they model what is statistically likely. World models predict what should happen next, not what might be intriguingly wrong.

A mathematical AI needs a mechanism whose role is:

• amplify novelty,
• shift the search region when a potential insight appears,
• destabilize the local gradient,
• break symmetry in exploration.

In practice, this could be implemented as:

• surprise-weighted sampling in generation,
• novelty gradients that modify internal temperature parameters,
• meta-learning loops that trigger global reconfiguration when the model encounters contradictions.

Mathematics begins where the gradient points nowhere obvious.

---

⭐ 2. Stability: The Serotonergic Principle

Creativity without collapse

In the brain, serotonin modulates confidence, mood, and learning rate. It determines whether a mind becomes rigid or chaotic.

Mathematical work requires the same balance:

• Too little structure → endless wandering.
• Too much structure → no innovation.

An AI mathematician must dynamically regulate its own:

• global learning rate,
• breadth of search,
• willingness to commit,
• threshold for “this is interesting enough to pursue.”

This stabilizing module could take the form of:

• adaptive regularization,
• learning-rate schedules conditioned on internal uncertainty,
• verification feedback loops that tighten or loosen exploration.

Humans do this unconsciously: they know when to doubt, when to trust a lemma, when to backtrack. Machines must learn it explicitly.

---

⭐ 3. Attention: The Acetylcholinic Principle

Insight requires knowing which structures matter

Acetylcholine in the brain is a pointer system: it tells neural circuits where to focus, what to process sharply, and what to ignore.

Mathematics is not a uniform landscape. A mathematician does not scan all directions equally. They develop a sense for:

• which lemmas are promising,
• which objects are central,
• which dimensions are irrelevant,
• which symmetries are exploitable.

In AI, this requires:

• structural sparsification of attention,
• graph-based relevance weighting,
• task-conditioned focus vectors,
• representations that emphasize mathematical objects over surface text.

Transformers began this work, but mathematics demands something sharper: attention must become semantic, not just lexical.

---

⭐ 4. Topology of Insight: Where Leaps Happen

Insight is a topological transition — a change of representation

Every mathematician knows the experience:

A problem seems intractable in its current form… until you translate it into another space, and suddenly it becomes trivial.

This is more than a trick. It is the architecture of mathematical creativity.

Examples include:

• converting analysis into topology via compactness (Heine-Borel / Borel–Lebesgue),
• converting algebra into geometry via varieties,
• converting combinatorics into spectral graph theory,
• “compactifying” spaces to apply powerful theorems,
• mapping problems into their duals,
• using Lyapunov functions to tame dynamical systems.

Mathematicians perform these transitions instinctively.

An AI mathematician must search across representation spaces, not just within them.

How?

Two mechanisms:

A. CK-Theory Style Concept Expansion

C-space: conceptual variations, hypothetical embeddings, analogies. K-space: established theorems, verified structures.

Insight emerges from alternating between the two.

B. Teleportation on Mathematical Graphs

Mathematical knowledge can be modeled as a:

• graph of theorems,
• graph of concepts,
• graph of transformations,
• graph of embeddings.

Insight is a graph jump — not random, but along dimensions like:

• topological similarity,
• centrality,
• spectral closeness,
• small-world distances,
• latent manifold proximity.

This is where embedding spaces are not enough. We need mathematical latent spaces with geometry aligned to the underlying theory.

Compactness becomes one such latent space: you temporarily embed a problem into a compact world, use global theorems, then project back.

This is a blueprint for machine insight.

---

⭐ 5. The Multimodal Architecture of a Mathematical Mind

A real mathematician is not one algorithm. They are:

• a world model (intuition for structure),
• a language model (manipulation of symbols),
• a graph reasoner (navigation of theory),
• an explorer (surprise-seeking),
• a verifier (proof discipline),
• a concept generator (C-space search).

Modern AI systems separate these capacities. A mathematical AI must integrate them.

Such a system would combine:

▸ An LLM for symbolic manipulation and narrative structure of proofs

▸ A world model for continuous intuition (limits, curvature, flows)

▸ A graph-based reasoning engine for theorem retrieval and concept connections

▸ A latent manifold navigator for topological leaps

▸ A neuromodulation-inspired controller for exploration, focus, stability

▸ A proof verifier for grounding new insights

This is not one model. This is an ecosystem of complementary intelligences working in concert.

The Grothendieck dream is alive here: the intelligence is not in the object — it is in the language and world you build around the object.

---

⭐ 6. Why This Matters Now

We may be entering a phase where:

• LLMs produce enormous mathematical volume,
• diffusion-like models explore concept space,
• graph reasoners navigate deep theoretical landscapes,
• world models give intuition for flows and continuous phenomena,
• neuromodulation-inspired controllers keep everything balanced, young, exploratory.

Mathematics may become the first domain where AI exhibits non-human creativity — not because it imitates us, but because it uses representational leverage we cannot access directly.

A system that can jump across mathematical worlds can discover what no single world contains.

---

⭐ 7. The Open Research Agenda

To build an AI mathematician, we need to answer questions that are both theoretical and engineering-driven:

• What is a “mathematical latent space”?
• How do we formalize surprise as a computational objective?
• How should CK-style concept expansion be operationalized?
• Can teleportation on concept graphs be optimized?
• What is the computational analog of compactness?
• How do we blend symbolic and continuous reasoning seamlessly?
• Can we create a model that invents intermediate concepts the way Grothendieck did?

None of these requires a paper to begin. They require a framework.

And frameworks, when solid, become research programs.

---

⭐ 8. Conclusion: Beyond Competence

An AI that can solve math problems is impressive. An AI that can discover new mathematics will redefine what we call intelligence.

It will not come from a single architecture. It will come from a multimodal mind:

• exploratory,
• stable,
• attentive,
• representationally fluid,
• structurally insightful.

Mathematics is a universe. To explore it, AI must learn what mathematicians do instinctively:

Change the map, not just the coordinates. Change the language, not just the symbols. Change the space, and insight follows.