LLM Framework for Discovering Major Mathematical Conjectures: AI’s Quest for the Next Riemann Hypothesis

The search for the next great mathematical conjecture often feels like waiting for a spark of intuition that can reshape an entire field. A new paper proposes a systematic way to generate and test such ideas, using language models alongside formal proof assistants to move from informal hunches to candidates that could deserve the same attention as the Riemann Hypothesis.

What You Need to Know

The authors introduce a three‑stage pipeline designed to uncover conjectures with high “problem taste.” In the first stage, a set of local evidence modules scans existing literature, datasets, and known theorems to surface promising patterns or gaps. These modules produce natural‑language statements that capture a specific mathematical observation, such as a relationship between special values of L‑functions and distributions of prime knots.

In the second stage, each candidate undergoes reflective validation. The model evaluates the statement for foundationality (does it rest on well‑established definitions?), novelty (has it appeared before?), and potential significance (could a proof reorganize concepts in its area?). Scores from these checks filter out low‑impact ideas, leaving a smaller set worth deeper scrutiny.

The final stage translates surviving conjectures into Lean 4 code, leveraging the Mathlib library to attempt a formal proof sketch or to detect immediate contradictions. If the statement resists disproof and remains unproven after a bounded automated search, it is promoted as a conjecture worthy of human investigation. The paper reports that, out of twenty initial candidates generated from a corpus of number‑theory papers, twelve passed the reflective stage and eight reached the formal validation step without being refuted.

Why It Matters

Mathematical progress still relies heavily on the intuition of individual experts, which can be uneven and hard to scale. By providing a repeatable process that blends informal idea generation with rigorous formal checking, the pipeline offers a way to surface problems that might otherwise be overlooked, especially those that lie at the intersection of subfields where cross‑talk is infrequent.

Moreover, the emphasis on “problem taste” aligns the automated search with the criteria that human mathematicians use when they judge whether a conjecture is worth pursuing. If the method consistently yields statements whose eventual proofs would restructure a research area—much like the way the Riemann Hypothesis ties together analytic number theory, random matrix theory, and quantum chaos—then it could become a valuable supplement to traditional conjecture‑forming practices.

Key Details

• Local evidence modules extract patterns from theorem statements, proof sketches, and data sets.
• Reflective validation uses prompting to score foundationality, novelty, and projected impact on a 0‑5 scale.
• Formal validation attempts to encode each conjecture in Lean 4; failure to find a counter‑example within a set timeout advances the candidate.
• Experiments were run on a corpus of twenty number‑theory conjectures drawn from recent arXiv submissions.
• Eight candidates survived all three stages without being disproved, suggesting non‑trivial mathematical potential.
• The pipeline is modular; swapping in different evidence sources or proof assistants (e.g., Coq, Isabelle) is straightforward.

What’s Next

The authors plan to expand the evidence modules to include syntactic features from proof networks and to integrate interactive feedback from mathematicians who can label candidates as “interesting” or “not.” They also aim to increase the rigor of the reflective stage by training a reward model on historic conjecture outcomes, hoping to improve the pipeline’s ability to predict which statements will eventually inspire deep theoretical development.

📌 Source: Arxiv Ai