“Increasing a model's context length improves its accuracy in predicting the token loss at a given position.”