Guiding AI Agents to Generate Conjectures by Encoding How Theorems Build on One Another
Summary
- Researchers proposed a method for automated mathematical conjecture generation using large language model agents.
- The approach introduces the Theorem Evolution Graph (TEG), which encodes how mathematical theorems build on one another.
- Agents construct the TEG by identifying relevant theorems through citation traversal and recording structural relationships.
- Evaluations by LLM judges and an expert mathematician showed that TEG-generated conjectures were preferred for significance, ambition, and insight over unconstrained baselines.
Why it matters
Automating mathematical conjecture generation addresses a core bottleneck in AI-driven scientific discovery, moving beyond theorem proving to discover what problems are worth pursuing.