Reformalization of the Jordan Curve Theorem

What Happened

Researchers have submitted a preprint to arXiv (2607.01734v1) documenting a case study in what they term "reformalization" — the process of translating a formal mathematical proof from one proof assistant system into another. Specifically, the work focuses on the Jordan Curve Theorem, a foundational result in topology stating that a simple closed curve divides the plane into an interior and exterior region. The authors distinguish reformalization from autoformalization by noting that the input is not natural language text but an existing formal development in a different proof assistant.

Why It Matters

This research addresses a practical bottleneck in formal mathematics: proof assistants like Coq, Lean, Isabelle, and HOL Light each have their own syntax, libraries, and foundational assumptions. Porting proofs between them is currently labor-intensive and error-prone, often requiring manual rewriting. The Jordan Curve Theorem is a particularly instructive test case because its proof is long, relies on multiple intermediate lemmas, and has been formalized in several systems over the years.

The significance lies in what reformalization reveals about the structure of formal proofs. Unlike natural language mathematics, which tolerates ambiguity and implicit reasoning, formal proofs are explicit and machine-checkable. Translating between them forces researchers to confront differences in type theories, axiom choices, and library dependencies. This makes reformalization a rigorous way to compare proof assistants and to identify which aspects of a proof are system-specific versus mathematically essential.

For the broader field of AI-assisted mathematics, this work provides a concrete benchmark. If AI systems can learn to automate reformalization, they would dramatically reduce the cost of maintaining formal libraries across platforms. The Jordan Curve Theorem case study offers a well-defined, non-trivial input-output pair that could train or evaluate models on formal-to-formal translation.

Implications for AI Practitioners

First, this research highlights a gap in current AI capabilities. While large language models have shown promise in autoformalization (natural language to formal code), reformalization (formal to formal) is a distinct task with different challenges — preserving logical structure while adapting to different type systems and library conventions. Practitioners working on neural theorem provers should consider adding reformalization benchmarks to their evaluation suites.

Second, the work underscores the importance of structured representations. Formal proofs are tree-like or graph-like objects, not just linear text. AI models that can reason about proof structure — perhaps using graph neural networks or transformers with positional encodings that respect dependency order — may be better suited for reformalization than standard language models.

Third, the Jordan Curve Theorem case study provides a realistic test of generalization. The proof involves algebraic topology, analysis, and combinatorial arguments. A successful reformalization system would need to handle diverse reasoning styles, not just simple algebraic manipulations. This makes it a more challenging and informative benchmark than toy problems.

Key Takeaways

• Reformalization is a distinct AI task from autoformalization, requiring translation between formal proof systems rather than from natural language
• The Jordan Curve Theorem serves as a rigorous, non-trivial benchmark for evaluating formal-to-formal translation systems
• AI practitioners should develop models that can handle structural dependencies and diverse reasoning styles across different proof assistant ecosystems
• Successful automation of reformalization could significantly reduce the labor cost of maintaining interoperable formal mathematics libraries