The Problem Is the Problem: Towards Scalable Mathematical Discovery
The paper’s core claim is that choosing the right math problems has become a bottleneck for AI-assisted discovery.
The authors propose FAR, a “Find, Attempt, and Recommend” pipeline that starts from a research direction rather than one expert-picked problem. In a combinatorics pilot, it mined 5,245 papers, extracted 6,453 candidate conjectures or open problems, and filtered them to 4,717 apparently well-posed, still-open conjectures. Further reasoning and triage surfaced 598 potential resolutions, with 77 sent for author-team review. The paper says this produced several interesting discoveries, including results tied to conjectures and questions from Davies-Jenssen-Perkins-Roberts, Erdős-Straus, Ikenmeyer-Pak-Panova, and Lund-Saraf-Wolf. HF Daily Papers' note
The authors propose FAR, a “Find, Attempt, and Recommend” pipeline that starts from a research direction rather than one expert-picked problem. In a combinatorics pilot, it mined 5,245 papers, extracted 6,453 candidate conjectures or open problems, and filtered them to 4,717 apparently well-posed, still-open conjectures. Further reasoning and triage surfaced 598 potential resolutions, with 77 sent for author-team review. The paper says this produced several interesting discoveries, including results tied to conjectures and questions from Davies-Jenssen-Perkins-Roberts, Erdős-Straus, Ikenmeyer-Pak-Panova, and Lund-Saraf-Wolf. HF Daily Papers' note
score 5