Duplicating 3 layers in a 24B LLM claims a jump in a logical deduction metric from 0.22 to 0.76 with no training.