Why is it at the word-to-word level? Reading seems to go smoothly when it’s using patterns that make sense. While suddenly bogging down when a phrase has unclear meaning or is constructed in a surprising way. Now, according to a new study by a research team from New York University and the University of Massachusetts Amherst, the answer may be found in the “moment-to-moment fluctuations” we see in a variety of reading behaviors, including eye movements, the researchers say, and this work may show to what extent even a system built just to predict the next word may account for differences in human reading.
The work matters, say the authors, not least because the LLMs, which learn by trying to match, to the tune of tens of thousands to a zillion words’ worth of text, the probabilities of one next, and then another next, and another next, are already being deployed to model language processes, though they are not obviously the process.
Was the statistical predictiveness of a word, how hard to predict it was (surprise), a useful guide to the reasons that readers might pause in reading and have trouble understanding sentences? The results implied that prediction describes a valuable aspect early in reading. But doesn’t entirely account for the effort of overturning and modifying once committed misreadings.
Read More: AI Examining Sleep Data as Predictors of Cognitive Health
What happens during reading
Reading is not simply decoding. We must, with speed, be able to extract from the printed character the identity of a letter, retrieve the meaning of a word, link it to preceding words, construct a parse for a sentence, and reparse it if new information undermines the original parse. Reading is usually fast, and it becomes apparent that it is even under when a reader constructs a sentence toward the inappropriate interpretation.
Another classic case involves the garden-path sentence, “The old man the boat.” “Man” in this sentence many first interpret to refer to an elderly gentleman, but in actuality it refers to the action of populating a boat with men or anything for that matter; the “the boat,” then, would be the object, man, then in effect, being the primary means of achieving what the writer of this sentence wants the sentence to mean, and thereby, what we’ve then had, as our prior reading (eye-words or back reads), have been of some value.
LLMs, too, make use of context but simply have these autoregressive models output probabilities of next words. It thus seems that these models will be appropriate to use when studying reading based on prediction of future words. The research team, moreover, differentiates between the normal routine predictions involved when reading forwards, as compared to processing the more systematic information to detect an ambiguity and resolving that.
Read More: Can Human Cognition Adapt to an AI-Assisted Future?
Experimental design
This experiment was conducted by William Timkey with Kuan-Jung Huang, Byung-Doh Oh, Grusha Prasad, Suhas Arehalli, Tal Linzen, and Brian Dillon. Eye movement data was analyzed from 368 adult readers as they read specially constructed sentences (including syntactically difficult garden-path items). They published their experiment in the Proceedings of the National Academy of Sciences.
They used the data to clock how much time participants spent staring at any particular word and whether they saccaded backwards, rereading previous bits of the sentence. The information is a more precise report of individual readings than a total reading time measurement. And this information then became the humans’ side of the story in a dataset also containing 409 predictions about the surprisal, or unusualness, of individual words in a list (or a sentence) that were calculated by over 400 different setups of various language models on different datasets.
However, rather than asking whether LLMs could recognize tough words, they put to the test a question in psycholinguistics: When our reading falters, are we mostly being challenged by unlikely words, or is it related to further challenges in forming, verifying, and repairing grammatical information about a sentence?
Slip between early and later stages of comprehension
An important difference emerged between earlier and later aspects of comprehension. When reading smoothly, readers’ left-to-right scanning could partially explain the time they took to identify words, with estimates of surprisal computed from an LLM. That is, when a word was more surprising (less predictable) from the context provided up to that word, this time would on average be longer for a human reader. This fits well with the notion that prediction constitutes a genuine part of language processing in human adults.
However, they did not fully account for the total reading difficulty when readers were confronted by the text they needed to reanalyze syntactically. With the garden-path construction, the point of syntactic resolution had an extra computational cost that could not be reduced to “predictability of upcoming words.” Readers not only slowed down at this difficult word. They frequently jumped back to previous text, which shows they reconsidered their initial grammatical interpretation.
There’s psychologically a good deal in this difference. This suggests that, rather than human readers merely collecting word-based probability estimates passively, human readers build mental-size structural hypotheses of a sentence’s meaning and detect evidence of violation, revise a hypothesis, and then reconcile a fixed interpretation. What standard LLM measures call the probability-based measure will have been doing some (perhaps) capture of this routine prediction, but much less capture of this repair.
Regressions are also important in the study. Roughly 20% of human fixations read leftward; backward eye movements are most useful for interpreting ambiguity or failure to integrate. Traditional autoregressive LLMs read the next words in a left-to-right direction. Thus have no equivalent of human backtracking and reinterpretation of material seen previously.
Author’s interpretation
As they explain, they take the results as suggesting a split between what they identify as difficulty in prediction and that in processing a structure. They say, “It is important to state we do not deny prediction accounts of what it takes for us to process what others say/write; what we aim to do is just show to what extent they make predictions fall short.” The evidence would be consistent with a theory that prediction does aid word recognition and the comprehension of what is typically processed within structure. However, it wouldn’t go all the way to say prediction would carry all the cognitive strain of realizing that processing something has been incorrect and repairing that.
Under this viewpoint, then LLMs are still scientifically important instruments: In showing that it’s possible to successfully reproduce an element of human learning behavior in the context of early reading ability by learning statistical regularities in language. But it has an important warning that there’s more than just successfully predicting the next word for a total theory about the meaning for humans. Language has elements of monitorability, checking for errors, reading, and changes of the structure that it might need computations for, unlike normal language models.
The broader implication concerns both cognitive science and AI design. For cognitive science, the paper provides evidence that eye movement data can dissociate certain processes that are confounded in reading time analyses. And for AI, it suggests that one should design systems that can represent uncertainty about a sentence’s meaning. Look back in the discourse to resolve this uncertainty, and explicitly revise their representation of the sentence’s meaning.
Conclusion
Researchers suggest that both measures may share one explanation: how surprising a word is within its context. The researchers found mixed evidence for this explanation. But the overall trend was that the surprisal-based LLM calculations were related to difficulty in processing word meaning. However, it was not enough to explain difficulty in syntactic ambiguity resolution. The resolution required extra processing steps not captured by these calculations.
The main point is that fluent language prediction is not equivalent to reading by humans. Understanding what causes people to reread challenged sentences could help researchers discover more accurate theories of reading and advance other areas of investigation, such as the construction of computational models of language comprehension and the development of more flexible, cognitively plausible artificial systems.
References +
- Neuroscience News. (2026m, August 10). Eye tracking reveals where human reading and AI processing diverge. Neuroscience News. https://neurosciencenews.com/human-reading-llm-neuroscience-ai-31203/
- Chandrasekharan, S., & Jacob, J. E. (2025). Bridging neuroscience and AI: A survey on large language models for neurological signal interpretation. Frontiers in Neuroinformatics, 19. https://doi.org/10.3389/fninf.2025.1561401
- Neuroscience News. (2026a, June 27). Reading is the ultimate cognitive enhancer. Neuroscience News. https://neurosciencenews.com/literacy-reshapes-cognition-30955/
- Luo, X., Rechardt, A., Sun, G., Nejad, K. K., Yáñez, F., Yilmaz, B., Lee, K., Cohen, A. O., Borghesani, V., Pashkov, A., Marinazzo, D., Nicholas, J., Salatiello, A., Sucholutsky, I., Minervini, P., Razavi, S., Rocca, R., Yusifov, E., Okalova, T., … Love, B. C. (2024). Large language models surpass human experts in predicting neuroscience results. Nature Human Behaviour, 1–11. https://doi.org/10.1038/s41562-024-02046-9
- Jiang, T., Wu, J., & Leung, S. C. H. (2025). The cognitive impacts of large language model interactions on problem solving and decision-making using EEG analysis. Frontiers in Computational Neuroscience, 19. https://doi.org/10.3389/fncom.2025.1556483
- Zewe, A. (2025). Like human brains, large language models reason about diverse data in a general way. In MIT News | Massachusetts Institute of Technology. https://news.mit.edu/2025/large-language-models-reason-about-diverse-data-general-way-021 9
- Schain, M., & Goldstein, A. (2025). Deciphering language processing in the human brain through LLM representations. In Research.google. https://research.google/blog/deciphering-language-processing-in-the-human-brain-through-ll m-representations/
- Roos, C. A., Postmes, T., & Koudenburg, N. (2023). Feeling heard: Operationalizing a key concept for social relations. PLOS ONE, 18(11), e0292865–e0292865. https://doi.org/10.1371/journal.pone.0292865
