On 16 July 2026, researchers posted a position paper on arXiv arguing that AI-driven autoformalization has been working at the wrong scale [S1]. Most efforts translate one theorem at a time. The paper says that approach cannot capture how real mathematics works: an entire web of axioms, definitions, and lemmas must exist before you can even state a target theorem [S1]. The question is whether the field can shift to formalizing whole theories at once, and what breaks if it cannot.

The gap between statements and theories

Autoformalization is the process of turning informal natural language, the way mathematicians actually write, into formal languages that computers can check [S1]. Think of it as translating a handwritten proof into code a machine can verify line by line.

Most research so far treats this as a sentence-by-sentence problem. You give the model a theorem in plain English, it produces a formal version, and a verifier checks it. The new paper, by Marcus J. Min and colleagues including Sharad Malik, Aarti Gupta, Xujie Si and Osbert Bastani [P2], says this misses the structure of real mathematics. A theorem like the fundamental theorem of calculus does not stand alone. It depends on definitions of limits, continuity, integrals, and a chain of lemmas, each of which depends on earlier axioms. Formalize the theorem without that scaffolding and you have a floating claim with no foundation [S1].

The authors call for theory-level autoformalization: formalizing complete theories, including all their inter-dependencies, as structured libraries [S1]. The difference is between translating a single sentence and translating an entire textbook with its cross-references intact.

Why the timing matters

This argument lands as the field is already scaling up. In May 2026, Meta released AutoformBot, a multi-agent system that orchestrates thousands of LLM agents to build an autoformalized textbook library at scale in Lean 4 [P4]. That effort treats formalization as a coordinated, multi-step pipeline rather than a one-shot translation, which is a step toward what the position paper describes.

Separately, tools like ReForm, a reflective autoformalization system created in October 2025, have been working on the quality of individual formalizations through iterative refinement [P3]. The position paper's argument is that even with better single-statement tools, the field needs to think at the level of entire theories.

What it means

The paper identifies a structural problem. If you formalize statements one at a time, you get a pile of verified theorems with no shared foundation. Two theorems might use slightly different definitions of the same concept. One might depend on an axiom the other contradicts. Without a unified library structure, these conflicts stay invisible until someone tries to combine the results.

Theory-level autoformalization would treat a mathematical theory as a single unit. All definitions, axioms, lemmas, and theorems would be formalized together, with their dependencies tracked and checked as a system. This is closer to how mathematicians actually work: you build up a theory, and each new result sits on top of what came before.

For a reader with no background, the stakes are about trust. If AI can formalize entire theories and a verifier confirms them, you get mathematical knowledge that is machine-checked end to end. If the field stays at the statement level, you get isolated claims that might be individually correct but collectively incoherent.

What it means for business

The direct commercial applications are narrow but growing. Companies building AI for scientific and engineering reasoning, from drug discovery to chip design, need verified mathematical foundations. A two-person AI startup building tools for formal verification of hardware designs currently faces a choice: hand-formalize the relevant theories, which takes months, or rely on AI to do it statement by statement, which produces fragments rather than a usable library.

If theory-level autoformalization becomes practical, that startup could point an AI system at a textbook and get back a structured, verified library in Lean 4 or another proof assistant. The cost of building verified knowledge bases would drop from months of expert time to hours of compute.

The paper also proposes three paths forward, though the abstract does not detail them [S1]. For teams already working with proof assistants, the practical question is whether existing multi-agent pipelines like AutoformBot can be extended to handle theory-level dependencies, or whether a fundamentally different architecture is needed.

What we don't know yet

This is a position paper, not a system or a benchmark. It advances an argument and identifies challenges, but it does not present a new algorithm or empirical results [S1]. The three proposed paths forward are outlined but not tested.

The paper has not been peer-reviewed; it is an arXiv preprint [S1]. The authors' framing of the field's direction reflects their perspective and should not be treated as settled consensus.

The authors maintain a survey of autoformalization research on GitHub [S1], but its scope and curation standards are not independently verified.

What would answer the open questions: a working system that formalizes a complete mathematical theory end to end, with a benchmark measuring its accuracy against human formalization. Meta's AutoformBot, which already builds textbook-scale libraries in Lean 4 [P4], is the closest existing test case. Whether it can handle the full dependency web the position paper describes, or whether it still operates at the statement level with loose coordination, will tell us how far the field has to go.

If you want to keep up with where AI verification is heading, subscribe for the next dispatch.

Sources

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.