Automated Conjecture Generation

158 concepts102 papers2 proposals

Most artificial intelligence research in mathematics focuses on deductive reasoning—specifically, automated theorem proving. Systems like AlphaProof (Hubert et al.) or MIZAR (Trybulec et al.) are designed to find a valid path from a known set of premises to a predefined conclusion.

However, this paradigm assumes that the destination is already known. As Thwaites et al. point out, our ability to evaluate mathematical models far outpaces our ability to propose valid theoretical frameworks in the first place. In pure mathematics, the initial experimental phase—recognizing patterns in raw data, defining new concepts, and formulating conjectures—is often left entirely to human intuition (Chau).

Automated conjecture generation flips the AI-for-math workflow. Instead of asking a machine to prove a given statement, researchers design systems to autonomously discover what is worth proving.

The Combinatorial Explosion of Trivial Truths

The fundamental challenge of automated discovery is not generating valid statements, but generating interesting ones. The space of syntactically correct mathematical expressions is infinite, but the vast majority of them are trivialities (e.g., $x + 0 = x$) or useless tautologies. To navigate this symbolic search space, systems must employ mechanisms that filter for mathematical significance.

Early symbolic systems relied on structured heuristic search. Lenat demonstrated with the AM system that applying general-purpose judgmental criteria to basic set-theoretic concepts could autonomously rediscover arithmetic, prime numbers, and complex relationships like Goldbach’s conjecture. Similarly, the HR program developed by Bundy et al. invented novel integer sequences by identifying recurring patterns across generated data points, evaluating them against criteria for "interestingness."

Modern approaches scale this search by replacing hard-coded human heuristics with optimization objectives. For example, Tsoukalas et al. utilize reinforcement learning to evolve programmatic measures of mathematical significance. By rewarding the system for finding definitions that are non-trivial yet rich in logical consequences, the agent autonomously navigates vast symbolic spaces to discover fundamental concepts in number theory.

Machine Learning as a Statistical Scout

How can neural networks, which are fundamentally statistical pattern recognizers, contribute to rigorous pure mathematics? They act as high-dimensional "scouts" that detect low-entropy structures within complex mathematical objects.

When a deep neural network achieves high predictive accuracy on a mathematical classification task, it serves as a heuristic signal that a hidden, structured law exists. Coates et al. and Kasprzyk et al. demonstrated this by training machine learning models to predict geometric properties (such as terminal singularities or dimensions) of complex algebraic varieties from raw, opaque data. The model's high accuracy acts as a flare gun: it proves that the relationship is not random, prompting human mathematicians to look closer, formulate a precise conjecture, and derive a formal proof.

[Raw Mathematical Data] ──> [Neural Network] ──> [High Predictive Accuracy]
                                                          │
                                                          ▼
[Rigorous Formal Proof] <── [Human Analysis] <── [Hidden Pattern Signaled]

This strategy allows researchers to bypass manual intuition when connecting disparate fields. Gukov et al. formalize this by using machine learning as a "conjecture engine" to map hidden dualities and correlations between entirely different mathematical invariants, such as the relationship between a knot's geometric volume and its topological signature.

Closing the Loop with Neuro-Symbolic Pipelines

To achieve true autonomy, modern conjecture generation relies on a closed neuro-symbolic loop: generative models propose broad, creative leaps, while symbolic engines prune or verify them.

Lin et al. outline this computer-aided discovery cycle as a multi-step process:

  1. Restricting the domain to a specific propositional language.
  2. Exhaustively generating rule combinations to observe empirical patterns.
  3. Formulating syntactic conjectures from those patterns.
  4. Applying meta-theorems or automated provers for universal verification.

In the era of large language models, this loop operates at a much higher level of abstraction. Large language models can explore vast combinatorial spaces more effectively than traditional manual methods, discovering novel extremal structures that elude human experts (Liu et al.).

In the FAR (Find, Attempt, and Recommend) framework introduced in The Problem Is the Problem, high-reasoning models scan existing literature to identify, solve, and prioritize open mathematical conjectures, generating publishable proofs and counterexamples.

This "conjecture-then-prove" workflow is highly practical. Zimmer et al. demonstrate that allowing autonomous agents to run numerical simulations to inform their theoretical conjectures enables them to successfully establish new lower bounds for complex optimization algorithms.

Furthermore, these generated conjectures do not just sit in isolation; Rabe et al. showed that machine-generated conjectures produced through structural completion tasks can achieve high provability rates when tested by automated theorem provers, subsequently serving as valid premises that make it easier to prove other independent theorems.

Key Foundational Work

Automated Theory Formation in Mathematics (1977) — contributed 4 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 102 papers
  2. Version 1 · 2026-10-04 · 71 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_automated_conjecture_generation_b6c4f567,
  title = {automated conjecture generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-conjecture-generation/rdr_b6c4f567aa7c41159c0797b86255fa20},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_automated_conjecture_generation_b6c4f567_v2,
  title = {automated conjecture generation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-conjecture-generation/rdr_b6c4f567aa7c41159c0797b86255fa20/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake