In large language models (LLMs) like ChatGPT, after the model processes the input, it produces a final hidden vector - a compact representation that summarizes the context up to that point. To convert this vector back into actual words, the model uses an unembedding matrix, effectively reversing the embedding step. This unembedding matrix is essentially the transpose of the original embedding matrix. It projects the hidden vector into the vocabulary space, producing a numerical score for every p