Users praise the BRIDGE approach of decomposing into code/spec/proof domains as the right call because direct natural language to Lean translation was always going to hit a wall.
Based on 1 visible X reactions from 2 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@AnimaAnandkumar @icmlconf BRIDGE decomposing into code/spec/proof domains is the right call. Direct NL-to-Lean was always going to hit a wall with Lean's total/immutable recursion style. Functional intermediate reps bridge that gap naturally.
ITPEval appears at the @icmlconf 2026 AI4Math Workshop and studies formal translation across Lean 4, Rocq/Coq, Isabelle/HOL, and HOL Light. The goal is not just to translate syntax, but to test whether models preserve mathematical meaning across different foundations, libraries, theorem names, tactics, and proof styles. We also release the infrastructure to evaluate these systems across multiple theorem provers, making it easier to benchmark progress on prover interoperability. Paper: https://openreview.net/forum?id=4PShGfBOMG Repo: https://github.com/lean-dojo/ITPEval
Both open-source tools are hosted under GitHub's Lean Dojo
Users praise the BRIDGE approach of decomposing into code/spec/proof domains as the right call because direct natural language to Lean translation was always going to hit a wall.
Based on 1 visible X reactions from 2 accounts; directional sample.
Ask a question below.
Published answers will appear here.