The core computation formula for the self-attention mechanism in the Transformer architecture is Attention(Q,K,V) = softmax(QKᵀ/√d_k)V, where the softmax function normalizes the attention scores of all positions into a probability distribution summing to 1.