Most benchmarks struggle to assess whether the model is truly “reasoning” or merely recognizing patterns from its training ...