Subscribe to Events

Mathematics has been shaped as much by its open questions as its proofs, famously some of Hilbert's 23 problems are still influencing research more than a century after they were conjectured. Despite the recent advancements in AI-assisted math, there is limited progress in machines generating consequential conjectures. We present ongoing work on a reinforcement learning framework for automated hypothesis generation, centered on discovering interestingness as a learnable reward signal. 

Our metric draws on the graph structure embedded in ProofWiki, a repository of theorems, definitions, and their dependencies. Conjectures that closely mirror densely connected, well-trodden regions of the graph score low, while those that forge links between distant or sparsely connected nodes score higher, capturing the intuition that interesting mathematics bridges unexpected concepts. Together with measures of a conjecture's structural non-triviality, these signals operationalize the middle ground between the obvious and the intractable. We discuss open challenges in quantifying interestingness, and the gap between formal metrics and mathematical taste.