ablating layer 8-12 in llama 3.1 70b breaks numerical reasoning but not verbal

tested systematic ablation on llama 3.1 70b q4_k_m yesterday and found something interesting: 1. ablating layers 8-12 completely breaks numerical reasoning tasks ("what is 47 * 23" drops from 94% accuracy to 31%) 2. verbal reasoning holds fine (analogies, word relationships stay at 89% vs 91% baseline) 3. factual recall is unaffected (capitals, dates, historical facts all within 2% of baseline) the pattern suggests layers 8-12 might be specialized for arithmetic/numerical manipulation rather than general reasoning. wondering if this replicates on other model sizes or if it's specific to the 70b architecture? used mean ablation, tested on 200 samples per task type