Semantic Correctness Evaluation on OpenMath
54FaithfulnessNL vs. Lean
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| NL vs. LeanEvaluation Protocol=direct, Model=DeepSeek2026.06 | 54 | 4 | 42 | 72.6 | |
| NL vs. NLbEvaluation Protocol=round-trip, Model=GPT-oss-120b2026.06 | 46 | 14 | 40 | 72.6 |