Semi-Autonomous Mathematics Discovery with Gemini

Analysis

Why This Paper Matters

This paper presents a compelling case study in semi-autonomous mathematics discovery, leveraging Gemini to systematically evaluate 700 conjectures from Bloom's Erdős Problems database. The significance lies in its demonstration that AI can effectively narrow the search space for open problems, addressing 13 of them through either novel solutions or literature identification. The finding that the 'Open' status was due to obscurity rather than difficulty challenges assumptions about the nature of unsolved problems and suggests that AI-assisted literature review could accelerate mathematical discovery.

The paper also critically examines the pitfalls of applying AI to mathematical conjectures at scale, particularly the difficulty of literature identification and the risk of 'subconscious plagiarism'—where AI may generate solutions that inadvertently replicate existing work without proper attribution. This raises important ethical and methodological considerations for the field.

Technical Contributions

• Hybrid methodology: Combines AI-driven natural language verification to filter conjectures with human expert evaluation for correctness and novelty.
• Scalable evaluation: Applied to 700 open conjectures, demonstrating the feasibility of AI-assisted discovery at scale.
• Identification of obscurity vs. difficulty: Provides evidence that many open problems are solvable but overlooked, shifting focus from problem difficulty to literature accessibility.
• Discussion of AI risks: Highlights 'subconscious plagiarism' and literature identification challenges, contributing to best practices for AI in mathematics.

Results

The paper reports that out of 700 open conjectures, 13 were addressed: 5 through seemingly novel autonomous solutions generated by Gemini, and 8 through identification of existing solutions in the literature. No specific metrics (e.g., accuracy, precision) are provided beyond these counts. The results underscore the potential of AI to uncover overlooked solutions but also the need for careful human oversight.

Significance

This work has broader implications for AI-assisted scientific discovery. It suggests that semi-autonomous systems can help solve open problems by improving literature search and generating candidate solutions. However, it also cautions against over-reliance on AI without robust verification mechanisms. For AI practitioners, the paper offers a template for integrating large language models into mathematical research workflows while addressing integrity concerns. The findings may influence how researchers approach problem-solving in mathematics and other fields with extensive prior literature.