Many large language models exhibit significant overfitting on established mathematical reasoning benchmarks, suggesting their performance may not reflect true reasoning abilities.
