AI Framework Aims to Automate Discovery of Major Mathematical Conjectures

A recent submission to arXiv (2607.28632) introduces a three-phase process designed for the systematic creation and verification of significant mathematical conjectures, with the goal of minimizing dependence on expert intuition. This framework, crafted by an undisclosed team, combines regional searches from explicit local evidence modules, reflective validation assessing foundationality, novelty, and significance, along with formal validation utilizing Lean 4 and Mathlib. The aim is to uncover high 'problem taste' conjectures that could reshape research fields and offer lasting support to mathematicians. In tests involving twenty candidates, the pipeline successfully transitioned from natural language to formal verification: all candidates passed Lean parsing and type checking, but none were directly accepted by the 'exact?' tactic or automatically proved. While the paper suggests AI might assist in discovering conjectures similar to the Riemann Hypothesis, it does not assert that such a conjecture has been identified. This research aligns with a growing trend of employing AI in mathematics, specifically targeting conjecture discovery rather than theorem proving. The methodology and initial findings are summarized in the abstract of the paper. No specific authors or institutions are mentioned, and the source URL is https://arxiv.org/abs/2607.28632.