Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation cs.LG · 2026-05-04