AViD Journal: Automated Novelty Verification for Mathematics with Lean 4
The AViD Journal, a newly developed pipeline, seeks to streamline the process of verifying mathematical novelty, tackling the shortcoming of AI systems that can confirm correctness but fail to assess novelty. This system takes a LaTeX article, translates its statements into Lean 4, and delivers a novelty judgment through a decision tree that evaluates prior occurrences in both formal and informal sources, assesses non-triviality using automatic tactics, and measures the structural disparity between proofs. An evaluation of retracted papers revealed three challenges that hinder the effectiveness of this approach.
Key facts
- AViD Journal is a pipeline for automated novelty verification in mathematics.
- It formalizes statements from LaTeX articles in Lean 4.
- The decision tree uses three dimensions: prior existence, non-triviality, and structural distance.
- Prior existence is checked in Mathlib (formal corpus) and TheoremSearch/Matlas (informal corpus) with temporal filter and LLM judge.
- Non-triviality is assessed via automatic tactics.
- Structural distance between proofs is measured as Jaccard distance over premise sets.
- Evaluation was conducted on papers withdrawn from arXiv due to declared duplication.
- Three obstacles were identified that limit the approach regardless of implementation.
Entities
Institutions
- arXiv
- Mathlib
- TheoremSearch
- Matlas