Generating mathematically sound future claims requires grounding in both scientific context (citations) and formal logical structure—neither alone is sufficient for plausible theorem prediction.
This paper tackles predicting future mathematical theorems by combining two types of information: citation networks (what mathematicians cite) and formal theorem dependencies (what logically follows from existing proofs). The authors build COMPOSE, a system that uses both graphs together to generate plausible new theorem statements, validated against real papers from 2024-2025.