On the Sample Complexity of Next-Token Prediction
2025
A loss that scores a predicted probability distribution by the negative log-probability assigned to the observed outcome.
For conditional language modeling, it is applied token by token to the distribution predicted from the preceding context.