Fig. 2.

CDRNN-estimated functional form of surprisal ($x$-axis) effects on reading times ($y$-axis) across language model types (n-gram, PCFG, GPT-2, GPT-J, GPT-3, and human cloze) with no delay (i.e. at the surprising word). Plots cover the interdecile range of values in each training dataset (for plots covering the full empirical range, see SI Appendix, Fig. S9). Kernel density plots show the distribution of surprisal values in the training data over the plotted range.