COMPOSE: Composing Future Theorems from Citations and Formal Structure

Researchers introduced COMPOSE, a dual-graph system for predicting plausible future theorem-like claims from an anchor paper.

It conditions a language model on both the scientific citation graph and aligned formal theorem dependencies, making predictions more grounded.

The team also built a 108K paired dataset from arXiv and Mathlib, plus a 47K-paper benchmark covering future papers from 2024 to 2025.

COMPOSE outperformed strong baselines in retrieval and LLM-judge evaluations, suggesting value for literature discovery and theorem forecasting.