PaperFormalizer: a software agent that closes gaps in research papers with formal proofs
Mathematical proofs in scientific papers remain vulnerable to subtle gaps in reasoning chains or undisclosed assumptions, even after undergoing careful and repeated reviews. This challenge led researchers Jiwon Tan, Vasileis Sergkanis, and José Blanchette to develop the PaperFormalizer system, a workflow built on the interactive proof environment Lean to help authors detect and close those gaps within their manuscript drafts, and to convert each study into a formally verified and programmatically documented form with strict computational precision.
The process begins by converting the source file written in LaTeX into a graph that depicts the dependency network and logical connections among theories and mathematical statements. Through this graph structure, each natural-language statement is examined to assess wording fidelity, measure ambiguity likelihood, and ensure logical correctness, preparing it for formulation and proof in Lean using the Mathlib library, together with specialized libraries and the researcher’s personal mathematical toolset.
The author's presence at the core of the decision loop serves as the project's technical safety valve.When we examined the system’s operational architecture, we found that it rests on a clear principle: only the researcher possesses the precise understanding of what a statement aims to prove, and is therefore the sole authority to determine how to correct any flaw in the proof. Accordingly, the software agent pauses and consults the author whenever the text may be read more than once, presenting correction options in linguistic formulations and LaTeX symbols for the researcher to accept or reject, preventing isolated automated alteration of the paper’s scientific claims.
The system functions as an internal plugin for intelligent coding assistants such as Claude Code or CodeX, ultimately producing a refined version that includes explanatory notes, a report of implemented improvements, and Lean code linking each program function to the corresponding mathematical statement it implements. The tool’s practical value grows as the researcher’s personal library expands with each newly vetted paper, and the verified versions can be published to enable other researchers to leverage the proven code and build upon it without having to re-audit from scratch.
Automated proof auditing reshapes the skill set required within regional research centers.In universities and scientific institutes across the Gulf, Egypt, and the Levant, teams developing theoretical AI models and cryptographic applications face a shortage of specialists capable of manually engineering formal programmatic proofs. This innovation reduces the cost of auditing mathematical theories and allows academic institutions to examine their research before submission to international journals, making the ability to manage coding agents and review mathematical documentation a practical benchmark that raises the region’s scientific publishing quality and shields it from invisible errors.
Restricting core permissions to the human researcher prevents common generative drifts and demonstrates that computational mathematical proofing is now within research teams’ reach without consuming months of arduous manual work.