AI System Automates Mathematical Problem Discovery and Resolution

AI systems are increasingly contributing to mathematical research, but the efficiency of AI-assisted discovery is often limited by the scarcity of expert mathematical review and the initial selection of suitable problems. Current workflows typically concentrate human effort at the beginning and end, creating bottlenecks. This research proposes a new human-AI collaboration paradigm to address these limitations. Instead of a single, pre-selected problem, human experts define a broad research direction. The system then searches a vast literature corpus for candidate problems within that direction, automating the initial problem discovery phase. The Find, Attempt, and Recommend (FAR) pipeline implements this by filtering thousands of combinatorics papers to identify well-posed and open conjectures. Subsequent AI reasoning and triage stages surface potential resolutions, ultimately selecting a manageable number of items for expert review. This approach led to significant discoveries, including results on conjectures by prominent mathematicians, demonstrating a more scalable and efficient mode of mathematical discovery.