Evaluating the Robustness of Proof Autoformalization in Lean 4

In three linesStudy of robustness in LLM-based proof autoformalization models translating natural language proofs to Lean 4. Authors evaluate 7 recent models against global perturbations (paraphrases) and local perturbations (value/step modifications). Finding: all models are sensitive to global perturbations and fail to faithfully reflect local perturbations.

Summary generated by Claude — human-verified