Large language models represent each word as a vector in space, with meaning coming from contextual embedding in relation to surrounding words, enabling generalization from training data to novel contexts without explicit rule encoding.