Explain Q, K, and V matrices in self-attention

This tests whether you understand the information-retrieval pattern behind self-attention, not just the formula. A strong answer explains that Q, K, and V are learned linear projections of the same input embedding; Q represents the current token's request, K indexes all tokens, and V holds the content to retrieve; the dot product of Q and K creates a compatibility score that weights a sum of V vectors. A red flag is describing them as separate model inputs or fixed embeddings rather than three views of the identical source sequence.