Theory-Level Autoformalization Moves Beyond One-Off Theorem Translation

Introduction

Autoformalization has become a key research direction at the intersection of language models, formal methods, and automated theorem proving. Much of the existing work focuses on a seemingly clean problem: given an informal mathematical statement, translate it into a formal language that systems such as Lean, Coq, or Isabelle can check.

The arXiv paper “Theory-Level Autoformalization: From Isolated Statements to Unified Formal Knowledge Bases” argues that this framing is too narrow. In real formalization projects, a theorem is rarely an isolated sentence. It depends on a surrounding web of definitions, assumptions, lemmas, notational choices, and previously established results. Without that structure, the target theorem may not even be expressible in a meaningful way.

Key points

• Formalization is not just translation. A theorem usually sits inside a theory. To formalize it, one must also formalize the concepts and intermediate facts that make the theorem well-defined.
• Structured libraries are the real output. A single checked statement has limited value. A coherent formal library can be reused by future proofs, verification tools, educational systems, and automated reasoning pipelines.
• The task definition changes. Theory-level autoformalization requires models to handle module organization, naming conventions, dependency graphs, concept consistency, and reuse of existing formal material.
• Evaluation must become richer. Passing a type checker for one statement is not enough to judge whether a generated theory is well structured, maintainable, comprehensive, or extensible.
• The paper sets a research agenda. As a position paper, it surveys the importance of this shift, discusses alternative views, identifies open challenges, and proposes promising paths forward.

Why it matters

The proposed shift turns autoformalization from a language conversion problem into a knowledge engineering problem. For AI for Science, this is a significant reframing. If future systems can help transform textbooks, papers, or domain notes into verified formal libraries, they could support mathematical knowledge curation, scientific theory checking, and high-assurance software development.

The paper also highlights a deeper limitation of current AI systems. Success in formalization is not only about generating syntactically valid code or matching a target statement. It also depends on long-range context, abstraction management, dependency tracking, and integration with existing libraries. These are closer to the workflow of expert human formalizers than to one-shot translation.

Importantly, the authors do not claim that theory-level autoformalization has been solved. Instead, they argue that the community should treat it as the more realistic and more useful target. If the direction matures, autoformalization could become a foundation for unified, verifiable, and reusable formal knowledge bases.

Source: arXiv