Preprint2026
Beyond Solver Verdicts: Generative Reward Models for Autoformalization
Vikash Singh, Debargha Ganguly, Aman Goel, et al.
Introduces Verdict-Preserving-Unfaithfulness (VPU) in autoformalization and proposes Generative Verification (GenV), a reference-free continuous equivalence score achieving 0.961 AUROC.
0Sep 10, 2026Reasoning
arXiv