Moonshine: An Autonomous Mathematical Research Agent Centered on Conjecture Generation

Problem
This work addresses the gap in autonomous mathematical research agents capable of generating and proving conjectures, a largely unexplored area in AI. The authors present Moonshine, a system designed to autonomously generate mathematical conjectures and explore their implications. This paper is a preprint and has not undergone peer review, indicating that the findings are preliminary and subject to further validation.

Method
Moonshine employs a novel approach to conjecture generation by extracting structural insights from classical mathematical problems. The architecture integrates a one-hidden-layer affine-ridge sigmoid network to explore the Jacobian conjecture, leading to the formulation of the Neural Jacobian Conjecture (NJC). The NJC posits that if a network maintains a strictly positive Jacobian determinant across its domain, it must exhibit global injectivity. The authors utilized advanced models, specifically GPT-5.5-pro and DeepSeek-V4-pro, to derive independent proofs for the case where (N=n+1). Additionally, they leveraged ChatGPT for interactive proof development, resulting in a geometric-topological proof. The training compute specifics are not disclosed, but the use of state-of-the-art models suggests significant computational resources were employed.

Results
Moonshine successfully generated independent proofs for the NJC in the case of (N=n+1), demonstrating its capability to make rigorous mathematical progress. The paper highlights that while the conjecture holds for this specific case, the general case for (N \ge n+2) remains unresolved, indicating a need for further exploration. The results provide preliminary evidence supporting the conjecture’s plausibility, although specific effect sizes or quantitative comparisons to established baselines are not detailed.

Limitations
The authors acknowledge that the general case (N \ge n+2) is unresolved, which limits the applicability of their findings. Additionally, the reliance on advanced AI models raises questions about the reproducibility and verification of the proofs generated. The lack of peer review means that the methodologies and results should be interpreted with caution until validated by the broader mathematical community.

Why it matters
The development of Moonshine represents a significant step toward automating mathematical research, potentially transforming how conjectures are generated and explored. By demonstrating the ability to autonomously derive and prove conjectures, this work opens avenues for further research in AI-driven mathematics, particularly in conjecture generation and proof verification. The implications of this research extend to various fields, including theoretical computer science and mathematical logic, as it suggests that AI can contribute meaningfully to foundational mathematical problems. This work is available on arXiv and may inspire future investigations into the capabilities of AI in formal mathematics.