When a language model settles on a prediction for a given word, it does so at some point along its depth. A new preprint argues that where along that depth the decision happens predicts how long people take to read that word, and it does so beyond anything the final-layer surprisal already tells you.
Donato Riccio and Fabrizio Maturo of Universitas Mercatorum in Rome ran the experiment (https://arxiv.org/abs/2611.25970) across eight English eye-tracking and self-paced reading corpora and ten models from three families (GPT-2, OPT, Pythia). The method is a measurement of when, not how much: read every word's next-token probability at every layer with the tuned lens, take the trajectory of surprisal across depth, and summarize it by its resolution depth D, the centroid of where the absolute change sits. D runs from 0 to 1 and equals 0.5 when the change is spread evenly; across the ten models its median sits between 0.38 and 0.50, so the settling point lives in the middle of the network. The paper's alternative threshold landmark, the depth at which 90% of the cumulative change is complete, falls in the top fifth of depth for 60% to 89% of words, which is why the centroid is the measure that carries information.
Later resolution, slower reading
Pooled over the eight corpora and ten models on the primary measure, adding D to a control set that already contains word frequency, length, final-layer surprisal, entropy and the best single readout improves held-out log-likelihood by a further 0.26 millinats per word (95% CI 0.10 to 0.40; p = 0.001). Each model family gains on its own: GPT-2 0.36 (p = 0.004), OPT 0.25 (p = 0.002) and Pythia 0.18 (p = 0.021).
In units a reader can feel, on the OneStop corpus a one-standard-deviation rise in D lengthens go-past time by 0.6% to 1.3% on the word itself and by 1.1% to 1.6% on the word that follows. Most of the effect lands on the successor: D of the preceding word alone gains 0.17 (p = 0.002) against 0.09 (p = 0.015) for D of the current word.
Which words resolve late
Comparing the fifth of words with the highest D against the fifth with the lowest D within each decile of final surprisal, the late-resolved words are shorter in 74 of the 80 corpus-model pairs (median 0.44 letters), more frequent in 65 and more often function words in 60 (median share 55% against 50%). On OneStop under GPT-2 XL, the commonest late-resolved words of middling surprisal are and, in, a, the, to, for, more and of; the commonest early-resolved ones are the, have, has, is and was.
The words that slow readers down are often the connectives and function words that look information-free. Readers spend more time on and than on a rare noun.
Three caveats the authors raise themselves
First, the effect is small. Measured against the gain from final-layer surprisal, D contributes 2.1% (0.26 against 12.05 millinats per word). Treating the eight corpora as a sample, the random-effects mean of the net gain is 0.07 with p = 0.37 and 75% of the variation between corpora unexplained, giving a prediction interval of -0.93 to 1.07 for a new corpus. What does hold up is the comparison against a layer-shuffled D: positive in seven of the eight corpora, random-effects mean 0.23 (0.03 to 0.44; p = 0.016), with between-corpus variation down to 27%. Shuffled D on its own manages -0.016 (p = 0.79).
Second, the corpora disagree sharply. GECO gives gains near zero for every model except OPT 6.7B (0.31; p = 0.024). On the UCL self-paced reading corpus eight of the ten models gain (0.36 pooled; p = 0.19) while UCL eye tracking loses for nine (-1.30). For GPT-2 and Pythia the effect rests on one corpus each; drop OneStop and GPT-2 keeps just 0.02 (p = 0.41).
Third, the result depends on the readout. The gain shows up only under the tuned lens: untrained networks with random weights gain -0.03 (p = 0.75) while trained models read the same way gain 0.22 (p = 0.025). The tuned-lens paper itself reports that the released lenses are severely undertrained, and under the logit lens with logit-lens controls and without the embedding readout the gain is 0.01 (p = 0.39). There is also a training-data wrinkle: ZuCo, MECO and two Natural Stories texts derive from Wikipedia, which the training data of OPT and Pythia include and GPT-2's excludes. The paper calls its analyses exploratory and covers decoder-only transformers up to 6.9B parameters.
Three independent lines of work
Three separate groups have been converging on the same idea. Hu, Lepori and Franke tested 20 open-source models across 6 domains and found that dynamic measures improve prediction of human processing measures relative to static final-layer measures, and that larger models do not always show more human-like processing patterns (https://arxiv.org/abs/2504.14107). Barenholtz introduced trajectory extrapolation error, fitting a linear path to the preceding hidden states and measuring deviation from it; the measure is nearly orthogonal to surprisal (r = .044) and independently predicts self-paced reading times on Natural Stories, replicating across GPT-2 and Pythia/RoPE positional encodings (https://arxiv.org/abs/2606.05346). Hayakawa and Saggion tied the distribution of predictive depth to reader proficiency, finding across 12 tested LLMs that readers with lower vocabulary proficiency show deeper predictive depth for first-pass gaze duration (https://arxiv.org/abs/2609.37688).
The metrics differ, but the direction agrees: the middle layers of a model are not a byproduct of residual computation, they carry information about human cognition.
So what
If the layer at which a model resolves a word really does predict how long people read it, then one thing worth recomputing is this: treating large models as black boxes when explaining human reading habits throws away an entire depth axis. The measure itself is cheap - one forward pass and a lens, no particular model depth required, comparable from GPT-2 up to 6.9B.
That said, do not treat it as a usable method yet. Right now it is a measure that holds across most corpora, collapses to zero on GECO, and depends on lens quality. Of the 80 corpus-model pairs, only 9 are significant on their own (4.0 expected by chance) and 30 of the 260 trained cells gain significantly (13.0 expected). The paper's own next step is to run experiments that vary the computation a prediction needs at a fixed probability, and see whether that computation itself costs human readers time.
Until then, D is better treated as a diagnostic lens than as a product.
Note: the acknowledgments state that the authors used a generative AI system (Claude, Anthropic), under their supervision throughout, in writing the analysis code and drafting the manuscript text.