Numeral Evaluation on The benchmark held-out slice
100AccuracyPython Solver Oracle
Evaluation Results
| Method | Links | |
|---|---|---|
| Python Solver Oracletype=reverse-engineered Python program2026.06 | 100 | |
| Solver-Grounded Synthetic CoT (Model)adapter=fine-tuned rank-≤32, decoding=greedy CoT2026.06 | 100 | |
| Base Model2026.06 | 95.9 |