Visored: A Controlled-Natural-Language Prover for LLM-Generated Mathematics

A new proof assistant called Visored, submitted on 16 June 2026, is designed to work with mathematics generated by large language models, complementing established systems such as Lean and Rocq [1][2]. Visored is a dependent-type-based prover whose surface imitates mathematical natural language and includes a rule-driven automation layer that closes routine steps a textbook would omit [1][2]. An accepted proof can be re-emitted as a checked Lean file, with output excerpts available on GitHub [1][2]. Early experiments suggest that, even without any prover-specific training data, LLMs can learn to use it effectively on the miniF2F benchmark [1][2]. Proof assistants are interactive software tools that support human–machine collaboration in developing formal proofs, with recent efforts incorporating artificial intelligence to automate the formalization of ordinary mathematics [4]. Lean, one of the systems Visored complements, is a proof assistant and functional programming language based on the calculus of constructions with inductive types; it is free and open-source, with development supported by the nonprofit Lean Focused Research Organization [5]. A separate project, Informath, explores symbolic informalization, a method for reliably converting formal mathematics into natural language without loss of precision [3]. The Informath architecture uses Dedukti as a hub between different proof systems—including Agda, Lean, and Rocq—and Grammatical Framework to handle linguistic correctness and variation across natural languages [3]. This approach aims to make machine-checked content human-readable and to explain precisely what has been constructed when proofs are generated by artificial intelligence [3]. Visored enters a landscape where proof assistants are increasingly expected to bridge formal verification and natural mathematical expression. The system’s ability to re-emit proofs as checked Lean files positions it as a bridge between LLM-generated reasoning and the rigorous standards of formal verification [1][2].