Elamur Research⌕
☰
home / papers / 2026 / Towards A Scanpath-Conditioned Surprisal Theory: Modeling Reader Information States

Towards A Scanpath-Conditioned Surprisal Theory: Modeling Reader Information States

Paper · 2026 · Michael Mooney and Edmond S. L. Ho

Summary

Surprisal theory usually measures how predictable a word is from the full written prefix. That choice assumes a reader who sees every word in order and remembers all of it. Real readers skip words, go back, and forget. Mooney and Ho keep the outcome space and change the conditioning context instead: predictability is computed from what a particular reader has actually fixated, according to their scanpath, and from what memory still makes accessible.

They define two families of variants:

  • Deterministic: a dense context up to the furthest word reached, or a sparse context containing only fixated words.
  • Stochastic: a random accessibility mask over past fixations models forgetting, and surprisal is estimated by Monte Carlo averaging of language-model probabilities.

On the OneStop and CELER eye-tracking corpora, scanpath-conditioned surprisal achieves better held-out log-likelihood on duration-based reading measures than standard surprisal. The authors take this as evidence that surprisal's weak account of non-linear reading stems partly from how the context is defined, not from the theory itself.

Key points

  • Model suite:
    • M0;dense and M0;sparse assume perfect memory.
    • M1 uses kernelized, recency-based forgetting.
    • M2 uses retrieval-based accessibility that is sensitive to how often a word was fixated.
    • Belief-update (BU) and information-gain (IG) extensions are added on top.
  • Base models and data: ModernBERT (masked) and Pythia-70m (autoregressive) were tested on OneStop and CELER across four reading regimes, including ordinary and repeated reading. Parameters were tuned on OneStop and held fixed for all evaluations.
  • Main ranking: M1 and M2 consistently beat M0, and M2 beats M1. BU/IG add further gains, with IG helping proportionally more under M2. Gains are much larger with ModernBERT, although M0;dense does best under the autoregressive model in ordinary reading.
  • Where it helps and where it does not: gains appear in measures such as GD, FRT and RGD. FFD and RGo-Past barely improve, which suggests that the cost of reintegration after a regression lies outside expectation-based accounts.
  • Limitations: results on CELER are volatile, which the authors attribute to its sentence-level stimuli producing denser scanpaths. The BU/IG extension is expensive, needing tens of GB of GPU memory per batch.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of GlasgowGlasgow · United Kingdom11 works · 635 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksPsychology · 1,552 worksDevelopmental and Educational Psychology · 499 worksHuman-Computer Interaction · 141 works

🧭 Topics 1Research subjects and languages the work is about.Psycholinguistics & language acquisition · 102 works

Cite

Cite as Michael Mooney and Edmond S. L. Ho. 2026. Towards A Scanpath-Conditioned Surprisal Theory: Modeling Reader Information States. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1919

@inproceedings{mooney2026scanpath,
  title = {{Towards A Scanpath-Conditioned Surprisal Theory: Modeling Reader Information States}},
  author = {Michael Mooney and Edmond S. L. Ho},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1919},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1919}
}

References

53 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ahmed M. Ahmed, A. Feder Cooper, Oluwasanmi Koyejo, and Percy Liang. 2026. Extracting books from production language models. ArXiv, abs/2601.02671. find
  2. Brian Bartek, Richard L. Lewis, Shravan Vasishth, and Mason R. Smith. 2011. In search of on-line locality effects in sentence comprehension. Journal of experimental psychology. Learning, memory, and cognition, 37 5:1178–98. find
  3. Yoav Benjamini and Yosef Hochberg. 1995. Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society. Series B (Methodological), 57(1):289–300. find
  4. Yevgeni Berzak and Roger Levy. 2023. Eye movement traces of linguistic knowledge in native and non-native reading. Open Mind, 7:179–196. find
  5. Yevgeni Berzak, Jonathan Malmaud, Omer Shubi, Yoav Meiri, Ella Lion, and Roger Levy. 2025. Onestop: A 360-participant english eye tracking dataset with different reading regimes. Scientific Data, 12. find
  6. Yevgeni Berzak, Chie Nakamura, Amelia Smith, Emily Weng, Boris Katz, Suzanne Flynn, and Roger Levy. 2022. Celer: A 365-participant corpus of eye movements in l1 and l2 english reading. Open Mind, 6:41– 50. find
  7. Klinton Bicknell and Roger Levy. 2009. A model of local coherence effects in human sentence processing as consequences of updates from bottom-up prior to posterior beliefs. In North American Chapter of the Association for Computational Linguistics. find
  8. Thomas M. Cover and Joy A. Thomas. 2005. Elements of information theory. find
  9. Ralf Engbert, Antje Nuthmann, Eike Richter, and Reinhold Kliegl. 2005. Swift: a dynamical model of saccade generation during reading. Psychological review, 112 4:777–813. find
  10. Janet D. Fodor and Fernanda Ferreira. 1998. Reanalysis in Sentence Processing. Springer Netherlands. find
  11. Nicolas Ford, Daniel Duckworth, Mohammad Norouzi, and George Dahl. 2018. The importance of generation order in language modeling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2942–2946, Brussels, Belgium. Association for Computational Linguistics. find
  12. Richard Futrell, Edward Gibson, and Roger P. Levy. 2020. Lossy-context surprisal: An information-theoretic model of memory effects in sentence processing. Cognitive Science, 44(3):e12814. find
  13. Mario Giulianelli, Sarenne Wallbridge, and Raquel Fern’andez. 2023. Information value: Measuring utterance predictability as distance from plausible alternatives. ArXiv, abs/2310.13676. find
  14. Adam Goodkind and K. Bicknell. 2018. Predictive power of word surprisal for reading times is a linear function of language model quality. In Workshop on Cognitive Modeling and Computational Linguistics.
  15. Michael Hahn and Frank Keller. 2023. Modeling task effects in human reading with neural network-based attention. Cognition, 230:105289. find
  16. John Hale. 2001. A probabilistic earley parser as a psycholinguistic model. In North American Chapter of the Association for Computational Linguistics.
  17. John Hale. 2006. Uncertainty about the rest of the sentence. Cognitive Science, 30(4):643–672. find
  18. Trevor Hastie and Robert Tibshirani. 1986. Generalized additive models. Statistical Science, 1(3):297–310. find
  19. Kuan-Jung Huang, Suhas Arehalli, Mari Kugemoto, Christian Muxica, Grusha Prasad, Brian Dillon, and Tal Linzen. 2024. Large-scale benchmark yields no evidence that language model surprisal explains syntactic disambiguation difficulty. Journal of Memory and Language, 137:104510. find
  20. Jukka Hyönä and Pekka Niemi. 1990. Eye movements during repeated reading of a text. Acta psychologica, 73 3:259–80. find
  21. Albrecht W. Inhoff, Andrew Kim, and Ralph Radach. 2019. Regressions during reading. Vision, 3(3). find
  22. Lena A. Jäger, Felix Engelmann, and Shravan Vasishth. 2017. Similarity-based interference in sentence comprehension: Literature review and bayesian meta-analysis. Journal of Memory and Language, 94:316– 339. find
  23. Carina Kauf and Anna Ivanova. 2023. A better way to do masked language model scoring. Preprint, arXiv:2305.10588.
  24. Keren Gruteke Klein, Yoav Meiri, Omer Shubi, and Yevgeni Berzak. 2024. The effect of surprisal on reading times in information seeking and repeated reading. ArXiv, abs/2410.08162. find
  25. Roger Levy. 2008. Expectation-based syntactic comprehension. Cognition, 106 3:1126–77. find
  26. Roger Levy, Klinton Bicknell, Tim Slattery, and Keith Rayner. 2009. Eye movement evidence that readers maintain and act on uncertainty about past linguistic input. Proceedings of the National Academy of Sciences, 106(50):21086–21090. find
  27. Richard L. Lewis and Shravan Vasishth. 2005. An activation-based model of sentence processing as skilled memory retrieval. Cognitive science, 29 3:375–419. find
  28. Richard L. Lewis, Shravan Vasishth, and Julie A. Van Dyke. 2006. Computational principles of working memory in sentence comprehension. Trends in cognitive sciences, 10 10:447–54. find
  29. Brian McElree. 2000. Sentence comprehension is mediated by content-addressable memory structures. Journal of Psycholinguistic Research, 29(2):111–123. find
  30. Clara Meister, Mario Giulianelli, and Tiago Pimentel. 2024. Towards a similarity-adjusted surprisal theory. In Conference on Empirical Methods in Natural Language Processing. find
  31. Akira Miyake, Marcel A. Just, and Patricia A. Carpenter. 1994. Working memory constraints on the resolution of lexical ambiguity: Maintaining multiple interpretations in neutral contexts. Journal of Memory and Language, 33(2):175–202. find
  32. Byung-Doh Oh and Tal Linzen. 2025. To model human linguistic prediction, make llms less superhuman. Preprint, arXiv:2510.05141. — [arXiv:2510.05141](https://arxiv.org/abs/2510.05141) find
  33. Byung-Doh Oh and William Schuler. 2023. Why does surprisal from larger transformer-based language models provide a poorer fit to human reading times? Transactions of the Association for Computational Linguistics, 11:336–350.
  34. Byung-Doh Oh and William Schuler. 2025. The impact of token granularity on the predictive power of language model surprisal. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4150–4162, Vienna, Austria. Association for Computati…
  35. Tiago Pimentel and Clara Meister. 2024. How to compute the probability of a word. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 18358–18375, Miami, Florida, USA. Association for Computational Linguistics.
  36. Maximilian M. Rabe, Dario Paape, Daniela Mertzen, Shravan Vasishth, and Ralf Engbert. 2024. Seam: An integrated activation-coupled model of sentence processing and eye movements in reading. Journal of Memory and Language, 135:104496. find
  37. Gary E. Raney and Keith Rayner. 1995. Word frequency effects and eye movements during two readings of a text. Canadian journal of experimental psychology = Revue canadienne de psychologie experimentale, 49 2:151–72. find
  38. Keith Rayner. 1998. Eye movements in reading and information processing: 20 years of research. Psychological bulletin, 124 3:372–422. find
  39. Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2020. Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699–2712, Online. Association for Computational Linguistics.
  40. Cory Shain, Clara Meister, Tiago Pimentel, Ryan Cotterell, and Roger Levy. 2024. Large-scale evidence for logarithmic effects of word predictability on reading time. Proceedings of the National Academy of Sciences, 121(10):e2307876121. find
  41. Omer Shubi and Yevgeni Berzak. 2023. Eye movements in information-seeking reading. In Annual Meeting of the Cognitive Science Society. find
  42. Omer Shubi, Cfir Avraham Hadar, and Yevgeni Berzak. 2025. Decoding reading goals from eye movements. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5616–5637, Vienna, Austria. Association for Computational Linguistics. find
  43. Koustuv Sinha, Robin Jia, Dieuwke Hupkes, Joelle Pineau, Adina Williams, and Douwe Kiela. 2021. Masked language modeling and the distributional hypothesis: Order word matters pre-training for little. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 288…
  44. Nathaniel J. Smith and Roger Levy. 2008. Optimal processing times in reading: A formal model and empirical investigation. volume 30, pages 595–600. find
  45. Nathaniel J. Smith and Roger Levy. 2013. The effect of word predictability on reading time is logarithmic. Cognition, 128(3):302–319. find
  46. Robyn Speer, Joshua Chin, Andrew Lin, Sara Jewett, and Lance Nathan. 2018. Luminosoinsight/wordfreq: v2.2. find
  47. Adrian Staub. 2025. Predictability in language comprehension: Prospects and problems for surprisal. Annual Review of Linguistics, 11(Volume 11, 2025):17– 34. find
  48. William Timkey, Kuan-Jung Huang, Byung-Doh Oh, Grusha Prasad, Suhas Arehalli, Tal Linzen, and Brian Dillon. 2025. Eye movements reveal a dissociation between prediction and structural processing in language comprehension. Preprint. find
  49. Aditya R. Vaidya, Javier Turek, and Alexander G. Huth. 2023. Humans and language models diverge when predicting repeating text. ArXiv, abs/2310.06408. find
  50. Julie A Van Dyke and Richard L Lewis. 2003. Distinguishing effects of structure and decay on attachment and repair: A cue-based parsing account of recovery from misanalyzed ambiguities. Journal of Memory and Language, 49(3):285–316. find
  51. Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Nathan Cooper, Griffin Adams, Jeremy Howard, and Iacopo Poli. 2024. Smarter, better, faster, longer: A modern bidirectional encoder for fast,…
  52. Ethan G. Wilcox, Tiago Pimentel, Clara Meister, Ryan Cotterell, and Roger P. Levy. 2023. Testing the predictions of surprisal theory in 11 languages. Transactions of the Association for Computational Linguistics, 11:1451–1470.
  53. Ethan Gotlieb Wilcox, Jon Gauthier, Jennifer Hu, Peng Qian, and Roger Philip Levy. 2020. On the predictive power of neural language models for human real-time comprehension behavior. ArXiv, abs/2006.01912. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mooney-2026-scanpath-conditioned-surprisal-theory-modeling/