PaperFormalizer converts LaTeX papers to Lean 4 declarations with author-in-loop audits
PaperFormalizer provides an author-controlled Lean formalization pipeline starting from LaTeX. It surfaces ambiguities for human resolution and produces traceable verified code. The approach pairs local PDF tooling with agentic proof search.
The pipeline transcribes each paper into a graph of natural-language statements, audits them for faithfulness and ambiguity, then attempts formalization inside a coding agent. When multiple readings exist, it surfaces repairs in both natural language and TeX notation for explicit author choice. Output includes a revised paper, change report, and Lean declarations traceable back to source statements.
ArXiv v1 (22 Oct 2026) reports the system runs inside Claude Code or Codex and maintains a growing author toolkit of verified lemmas. Shared libraries allow downstream reuse. The experimental PDF-to-HTML viewer on the same arXiv page demonstrates local WebAssembly document handling, reducing cloud dependency for the initial LaTeX ingestion step.
Operationally this shifts verification cost from post-publication reviewers to authors during drafting. Each accepted repair updates the personal toolkit, creating compounding reuse for follow-on work. The design explicitly avoids autonomous claims that would alter paper semantics.
Next steps include domain-library integration and shared corpus growth. Threshold: 100 papers with linked Lean code hosted by Q4 2027.
Claude Code: at least 30 papers reach public Lean-linked status within 12 months of release
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.31810)
- [2]Supporting Source(https://leanprover-community.github.io/mathlib_docs/)