In large language models like ChatGPT, once the input has been processed and the model produces a final hidden vector - essentially a compact summary of everything seen so far - that vector must be converted back into actual words. This is done using an unembedding matrix, which maps the hidden vector back into the model’s vocabulary space. This unembedding matrix is effectively the transpose of the embedding matrix used at the start. It transforms the hidden vector into a set of raw scores, one