Improving Next Tokens via Second-Last Predictions with Generate and Refine

원문 초록

Autoregressive language models like GPT aim at predicting next tokens, while autoencoding models such as BERT are trained on tasks such as predicting masked tokens. We train a decoder only architecture for predicting the second last token for a sequence of tokens. Our approach yields higher computational training efficiency than BERT-style models by employing a structured deterministic approach towards masking tokens. We use our model to improve the next token predictions of a standard GPT by combining both predictions in a ``generate-then-refine'' approach. We show on different variants of GPT-2 and different datasets that (not unexpectedly) second last token predictions are much more accurate, i.e., more than 15\% higher accuracy than ordinary next token predictors. The ``generate-then-refine'' approach also demonstrates notable improvements in next-token predictions, yielding smaller yet consistent and significant gains.