Pith. sign in

REVIEW 5 major objections 7 minor 52 references

Multimodal Behavioral Patterns Analysis with Eye-Tracking and LLM-Based Reasoning

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A human-AI pipeline that splits eye-tracking tables into rows and columns raises LLM consistency scores to 1.00 and difficulty-prediction accuracy from 0.317 to 0.50.

desk verdict A practical pipeline idea for LLM-based eye-tracking analysis, but the consistency evaluation is circular and the trust scores should not be taken at face value. read the letter →

arxiv 2507.18252 v1 pith:CAAXYVQQ submitted 2025-07-24 cs.HC cs.AIcs.CLcs.LG

Dongyang Guo , Yasmeen Abdrabou , Enkeleda Thaqi , Enkelejda Kasneci This is my paper
classification cs.HCcs.AIcs.CLcs.LG
keywords eyetrackinglargelanguagemodelsbehavioralpatternmininghuman-AIcollaborationcognitivestateanalysisanomalydetectionquestiondifficultypredictioneducationaltechnology
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that large language models, which normally stumble on rows and columns of numbers, can extract trustworthy cognitive patterns from eye-tracking data if the data is first split into horizontal (row-wise) and vertical (column-wise) segments and fed to the model with structured prompts. The authors argue that adding a human-in-the-loop co-scoring step, where a domain expert and a literature-based check both rate each extracted pattern, makes the LLM's interpretations consistent and interpretable enough to be useful in educational assessment and adaptive interfaces. The concrete payoff claimed is measurable: with both segmentation modules and detailed prompts, two of the three tested LLMs reach a consistency trust score of 1.00, and difficulty-prediction accuracy rises from 0.317 in the baseline to 0.50. A sympathetic reader would care because the framework offers a scalable path from raw physiological signals to semantic, actionable descriptions of cognitive states without abandoning interpretability.

What carries the argument

The load-bearing mechanism is the two-dimensional segmentation pipeline: the horizontal module splits the table into rows, converts each row to JSON, and asks the LLM for within-instance associations, while the vertical module pairs ID columns with numerical columns to trace feature evolution over time and reanalyzes them together with the horizontal patterns. The second mechanism is the Expert-Model Co-Scoring Module, which scores every extracted pattern with a weighted formula $F=\sum_{i=1}^{5}(C_i+R_i)\cdot S_i$, where $C$ is the citation count of one of five recommended papers, $R$ its journal ranking, and $S$ the paper's stance toward the pattern (support, oppose, or neutral), and then measures expert-to-literature agreement with Cohen's Kappa. A third module trains an LSTM on expert eye-tracking data and feeds the anomaly distribution across areas of interest to the LLM for semantic interpretation. Together these modules convert raw numerical gaze data into language-level behavioral patterns that can be checked against expert judgment.

What would settle it

Take a held-out behavioral pattern set, have human experts manually select the five related papers used for validation instead of letting the LLM recommend them, and recompute the Cohen's Kappa between expert ratings and literature-based validity; if agreement falls below the 0.6 threshold while the LLM-recommended version stays near 1.00, the reported trust scores are inflated by circularity.

Watch

Extended reading notes

Core claim

The central claim, stated in the discussion, is that combining human expertise with LLM reasoning through structured analysis modules provides measurable improvements in both consistency and accuracy. Concretely, the horizontal-plus-vertical (h+v) condition with fully detailed prompts raises consistency trust scores to 1.00 for ChatGPT-o1 and Deepseek-R1, up from direct-prompt baselines of 0.647 and 0.715 respectively, and lifts Deepseek-R1's question-difficulty prediction accuracy from 0.317 to 0.50. The same combination outperforms either module alone and remains partially effective even when the prompt carries no background information. In the authors' account, the framework reveals latent gaze behaviors, such as role-specific shifts toward error regions in pair programming, that a direct read of the tabular data misses, and it can flag both student-level anomalies and flaws in question design.

Load-bearing premise

The consistency scores assume that the literature check performed by the LLM is an independent referee, even though the same kind of model produced the patterns being checked.

Editorial extensions

If this is right

  • If the pipeline works as claimed, LLMs become a practical front end for exploring tabular physiological data, not just text.
  • The h+v combination's robustness under brief prompts suggests the structural segmentation itself, not prompt detail, carries much of the benefit.
  • Difficulty-prediction improvements imply the framework could assist educators in calibrating task difficulty without manual cognitive-load analysis.
  • The anomaly module's dual output, personalized student feedback plus question-design critique, points toward closed-loop adaptive learning systems.
  • The co-scoring trust scores could serve as a quality-control layer for filtering LLM-generated behavioral claims before they influence decisions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the 1.00 trust scores may be partly circular, because the same model family both generates the patterns and supplies the literature used as ground truth; a human-verified citation set could lower them.
  • Editorial inference: the horizontal/vertical segmentation recipe is likely transferable to any low-semantic-density tabular signal, such as EEG, ECG, or sensor logs, where the same 'make rows and columns legible to an LLM' move should apply.
  • Editorial inference: testing the framework on a second dataset with known expert labels would separate the contribution of segmentation from the contribution of the validation scheme, since the paper only runs one educational dataset.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

Summary. The paper proposes a human–AI collaborative framework for extracting behavioral patterns from tabular eye-tracking data. The framework consists of a horizontal/vertical segmentation pipeline that feeds structured data to LLMs, an expert–model co-scoring module that computes trust scores via Cohen's Kappa between an expert and an LLM-assisted literature check, and an LSTM–LLM hybrid anomaly detection module. The method is evaluated on a public pair-programming eye-tracking dataset using ChatGPT-4o, ChatGPT-o1, and Deepseek-R1 under three prompt-detail levels. The authors report improved consistency trust scores (up to 1.00) and improved question-difficulty prediction accuracy (up to 0.50) relative to a direct-prompt baseline, and conclude that combining human expertise with LLM reasoning yields measurable gains in consistency and accuracy.

Significance. If the validation were independent, the framework would be a useful contribution to HCI and educational analytics: it applies LLMs to structured physiological data in a modular way, uses a real public eye-tracking dataset, systematically varies prompt detail and segmentation mode, and repeats stochastic LLM runs. The paper does not ship code or machine-checked proofs, and the cited appendix with prompt templates is absent from the submitted text. However, the central evaluation is currently not convincing because the consistency validation is self-referential, the human-expert contribution is never ablated, and the accuracy comparisons lack statistical support. The result is a promising but unvalidated pipeline rather than an established claim.

major comments (5)
  1. [Section 4.2, Eqs. (1)-(3)] The consistency evaluation is circular in a load-bearing way. Behavioral patterns are generated by LLMs (ChatGPT-4o, o1, Deepseek-R1 in Section 4.1), and the 'literature-based verification' in Section 4.2 asks a large language model to recommend five related papers and to rate their content; no retrieval from an actual literature database or human verification of the recommended references is reported. The Kappa coefficient in Eq. (3) therefore measures agreement between the expert and an LLM-generated proxy of the literature, not between the expert and independently established literature. The text does not specify whether the validating LLM is the same model used for generation; if it is from the same model family, the perfect trust scores of 1.00 for o1 and r1 in Table 2 are exactly what one would expect from a validator that shares priors with the generator. Even if a different LLM is used, the recommended papers and stance scores are unverified model outputs rather than an external ground truth. Section 6.5's concession that the weighting can inflate validity scores does not address this more basic self-referentiality, which invalidates the consistency pillar of the central claim in Section 6.1.
  2. [Section 6.1 vs. Sections 4.2 and 5.2] The headline claim that 'combining human expertise with LLM reasoning' provides measurable improvements is not tested by the reported experiments. In every condition, including the Directly baseline, the same expert evaluation and LLM-assisted literature validation are applied; the only manipulated variables are the segmentation modules (h, v, h+v) and the prompt detail. There is no expert-only, model-only, or no-expert condition, so the observed gains in Tables 2 and 3 are attributable to the structural modules and prompts, not to the presence of human expertise. A factorial design or ablation that separates the human contribution is needed to support the causal wording of Section 6.1.
  3. [Section 4.3 and Table 3] No statistical evidence is provided for the accuracy claims. The difficulty-prediction experiment uses only 12 questions (4 per level), each setting is repeated 5 times, and Table 3 reports only point averages, with no variance, confidence intervals, or significance tests. With n=12, a one-question difference is 8.3 percentage points, so the reported improvement from 0.317 to 0.50 for Deepseek-R1 corresponds to roughly two questions and may be within run-to-run noise. LLM sampling parameters (e.g., temperature) are not reported. The table should include per-run results, standard deviations, and a test of whether accuracies exceed the 0.333 chance level and differ across conditions.
  4. [Sections 3.4, 4.4, and 5.1] The anomaly-detection component is not evaluated quantitatively. The LSTM's architecture, training/validation split, reconstruction-error threshold, and other free parameters are not specified, and no ground-truth labels, detection metrics (precision, recall, F1), or baseline comparisons are reported for the detected anomalies. Section 5.1 offers narrative interpretations for two students only. As a result, the third claimed contribution and the 'performance' part of the central claim are unsupported.
  5. [Abstract and Sections 5.2 and 6.1] The claimed improvement in interpretability is never measured. The abstract lists interpretability alongside consistency and performance as an improved dimension, but the only human input collected is a binary validity rating from one expert in Section 4.2; there is no user study, rating of explanation quality, or quantitative interpretability metric. In addition, relying on a single expert without reporting inter-rater reliability or blinding details makes the expert component itself difficult to assess. The interpretability claim should either be evaluated or narrowed.
minor comments (7)
  1. [Sections 4.1, 4.3, 4.4, 5.1, and 5.2] The manuscript repeatedly refers to Appendix sections for prompt templates and behavioral pattern examples, but no appendix is present in the submitted text; without the prompt templates the pipeline cannot be reproduced.
  2. [Section 5.2 and Table 2] Table 2 is labeled 'Consistency comparison' but the text calls the values 'trust scores'; please clarify whether the table reports Cohen's Kappa from Eq. (3) or a derived trust score, and report the number of patterns used for each Kappa calculation.
  3. [Figures 1-3] Figures 1-3 are referenced with captions but do not appear in the manuscript text; ensure the figures are embedded in the submitted version.
  4. [Section 4.1] The thresholds of 30% for high-frequency and 10% for low-frequency pattern selection are introduced without justification; a sensitivity analysis would help establish that the reported trust scores are not artifacts of these arbitrary cutoffs.
  5. [Section 6.5 and Table 1] The limitation paragraph states that there is an 'absence of weighting factors for citation count and relevance score,' but Table 1 already defines weights for C, R, and S; please reconcile this apparent contradiction.
  6. [References] Reference [34] appears to be misattributed: the 'Language models are few-shot learners' entry should list Brown et al. as first author, and reference [13] contains the typo 'cofficient.'
  7. [Abstract and Conclusion] The phrase 'up to 50% accuracy improvement' is misleading; the results show an accuracy of up to 50%, not an improvement of 50 percentage points, and the conclusion repeats this wording.

Circularity Check

2 steps flagged · score 6.0 of 10

Consistency evaluation is self-referential: the same LLM family generates patterns and supplies literature-based validity scores, so trust scores do not not measure independent agreement; difficulty prediction remains independent.

  1. other [Section 4.2 (Consistency of Behavioral Patterns), Eqs. (1)-(3); Section 4.1 (Pattern Mining); Section 5.2 Table 2]
    "For each behavior pattern, we input it into the large language model and ask it to recommend 5 related papers. We then performed a comparative analysis of the five related research papers to determine its effectiveness and rationality in the context of existing research. To save time and costs associated with manual literature review, we introduce a large language model to assist in the literature reading and content rating tasks..."

    The behavioral patterns are generated by LLMs (ChatGPT-4o, ChatGPT-o1, Deepseek-R1), and Section 4.1 additionally uses Deepseek-R1 to summarize and select the pattern set. The literature-based validation in Section 4.2 then asks an LLM to recommend the five papers and to rate their content, so the F score and the resulting Kappa/trust score are computed from evidence supplied by the same kind of model being validated. This makes the 'literature support' side of the Kappa a self-consistency check rather than an independent external ground truth, which explains how perfect trust scores of 1.00 can arise for o1 and r1 in Table 2.

  2. other [Section 6.1 (Human-AI Collaborative Framework); Section 6.5 (Limitations)]
    "Our findings confirm that combining human expertise with LLM reasoning via structured analysis modules provides measurable improvements in both consistency and accuracy. ... the weighting lacks finer granularity and doesn't fully account for contextual domain alignment. This can lead to inflated validity scores for studies with high citation popularity but lower relevance."

    The headline claim bundles 'consistency and accuracy', but only the accuracy component is supported by an independent difficulty-prediction benchmark. The consistency component is the Kappa/trust score derived from the self-referential validation loop described in Section 4.2, so the central claim relies on a circularly obtained result for one of its two pillars. The paper itself concedes that the scoring can inflate validity scores, yet the consistency conclusion is not revised in light of that concession.

full rationale

The paper's strongest independent result is question difficulty prediction, where accuracy is measured against pre-existing difficulty labels and is not circular. The anomaly detection and pattern mining outputs are presented descriptively and are not used as quantitative evidence of external validity. The main circularity is confined to the consistency evaluation: the LLM generates the behavioral patterns, the LLM recommends the supporting papers, and the LLM rates that literature to compute F, which is then compared with expert judgment via Cohen's Kappa. No human verification of the recommended papers or their relevance is reported, so the trust scores in Table 2 can reflect the model agreeing with itself rather than being anchored to an external literature standard. The paper's Section 6.5 acknowledges that the weighting can produce 'inflated validity scores', which is consistent with this concern. Because the consistency claim is one of the two pillars of the central contribution, the circularity is load-bearing; however, the difficulty prediction results remain independent, so the overall score is 6 rather than higher.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The framework's results depend on three hand-chosen evaluation parameters (selection thresholds, scoring weights, LSTM threshold) and four assumptions, the most important being that LLM-generated literature support is a valid independent check. There are no invented physical entities. The difficulty prediction task is the most self-contained part, but it relies on the same LLM systems without error bars.

free parameters (3)
  • High/low frequency selection thresholds = 30% high, 10% low
    Chosen by hand in Section 4.1 to reduce evaluation cost; no principled basis, affects trust scores.
  • Scoring weights C, R, S in Eq. 1-2 = C: 5..1, R: 4..1, S: +1/0/-1
    Hand-assigned weights for citation count, journal ranking, and stance; no sensitivity analysis and the paper admits weighting lacks granularity.
  • LSTM reconstruction error threshold = Not reported
    Anomaly detection uses reconstruction error, but no threshold is specified; the definition of 'anomaly' is therefore under-specified.
assumptions (4)
  • ad hoc to paper Literature support, as generated by an LLM and scored by citation count and journal rank, is a valid proxy for the correctness of behavioral patterns.
    Invoked in Section 4.2 to compute trust scores and Cohen's Kappa; this is the main validation of pattern consistency.
  • domain assumption Expert gaze patterns are normative; deviations from them are anomalies.
    The LSTM is trained only on expert data (Sections 3.4, 4.4), so any student deviation is treated as anomalous, though the authors acknowledge students may sometimes outperform experts.
  • domain assumption AOI annotations based on experimental design are correct and complete.
    All horizontal and vertical analyses and anomaly distributions rely on predefined Areas of Interest (Section 3.2).
  • standard math Cohen's Kappa is an appropriate measure of expert-model consistency.
    Used in Section 4.2; standard, but requires independent raters, which is questionable here because one rater is LLM-based.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Behavioral Patterns Analysis with Eye-Tracking and LLM-Based Reasoning." pith.science (2026). https://pith.science/paper/CAAXYVQQ

@misc{pith2026250718252,
  author       = {Pith},
  title        = {Pith review of: Multimodal Behavioral Patterns Analysis with Eye-Tracking and LLM-Based Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CAAXYVQQ}},
  note         = {Machine review of arXiv:2507.18252}
}
read the original abstract

Eye-tracking data reveals valuable insights into users' cognitive states but is difficult to analyze due to its structured, non-linguistic nature. While large language models (LLMs) excel at reasoning over text, they struggle with temporal and numerical data. This paper presents a multimodal human-AI collaborative framework designed to enhance cognitive pattern extraction from eye-tracking signals. The framework includes: (1) a multi-stage pipeline using horizontal and vertical segmentation alongside LLM reasoning to uncover latent gaze patterns; (2) an Expert-Model Co-Scoring Module that integrates expert judgment with LLM output to generate trust scores for behavioral interpretations; and (3) a hybrid anomaly detection module combining LSTM-based temporal modeling with LLM-driven semantic analysis. Our results across several LLMs and prompt strategies show improvements in consistency, interpretability, and performance, with up to 50% accuracy in difficulty prediction tasks. This approach offers a scalable, interpretable solution for cognitive modeling and has broad potential in adaptive learning, human-computer interaction, and educational analytics.

Figures

Figures reproduced from arXiv: 2507.18252 by the authors.

Figure 1
Figure 1. Multi-module integrated process design for attention shift and cognitive load detection. a) represents data processing. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. LLM-based 2D Data Analysis Flowchart: (A) Blue [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Outlier distribution: A1–D3 (questions), S1–S2 (stu [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 39 canonical work pages

  1. [1]

    Şeniz Harputlu Aksu, Erman Çakıt, and Metin Dağdeviren. 2024. Mental workload assessment using machine learning techniques based on eeg and eye tracking data. Applied Sciences 14, 6 (2024), 2282

    work page 2024

  2. [2]

    Tobias Appel, Peter Gerjets, Stefan Hoffmann, Korbinian Moeller, Manuel Ninaus, Christian Scharinger, Natalia Sevcenko, Franz Wortha, and Enkelejda Kasneci

    work page

  3. [3]

    Tobias Appel, Natalia Sevcenko, Franz Wortha, Katerina Tsarava, Korbinian Moeller, Manuel Ninaus, Enkelejda Kasneci, and Peter Gerjets. 2019. Predicting cognitive load in an emergency simulation based on behavioral and physiological measures. In 2019 International Conference on Multimodal Interaction . 154–163

    work page 2019

  4. [4]

    Gilbert Badaro, Mohammed Saeed, and Paolo Papotti. 2023. Transformers for tabular data representation: A survey of models and applications. Transactions of the Association for Computational Linguistics 11 (2023), 227–249

    work page 2023

  5. [5]

    Eirik Bakke, David Karger, and Rob Miller. 2011. A spreadsheet-based user interface for managing plural relationships in structured data. In Proceedings of the SIGCHI conference on human factors in computing systems . 2541–2550

    work page 2011

  6. [6]

    Gabriela Ben Melech Stan, Estelle Aflalo, Raanan Yehezkel Rohekar, Anahita Bhiwandiwalla, Shao-Yen Tseng, Matthew Lyle Olson, Yaniv Gurwicz, Chenfei Wu, Nan Duan, and Vasudev Lal. 2024. LVLM-Intrepret: An Interpretability Tool for Large Vision-Language Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 8182–8187

    work page 2024

  7. [7]

    Hannah Brown, Katherine Lee, Fatemehsadat Mireshghallah, Reza Shokri, and Florian Tramèr. 2022. What does it mean for a language model to preserve privacy?. In Proceedings of the 2022 ACM conference on fairness, accountability, and transparency. 2280–2292

    work page 2022

  8. [8]

    Babette Bühler, Efe Bozkir, Hannah Deininger, Peter Gerjets, Ulrich Trautwein, and Enkelejda Kasneci. 2024. On task and in sync: Examining the relationship between gaze synchrony and self-reported attention during video lecture learning. Proceedings of the ACM on Human-Computer Interaction 8, ETRA (2024), 1–18

    work page 2024

Show all 52 references
  1. [9]

    Simona Caldani, Christophe-Loïc Gerard, Hugo Peyre, and Maria Pia Bucci. 2020. Visual attentional training improves reading capabilities in children with dyslexia: An eye tracker study during a reading task. Brain sciences 10, 8 (2020), 558

    2020

  2. [10]

    Nora Castner, Enkelejda Kasneci, Thomas Kübler, Katharina Scheiter, Juliane Richter, Thérése Eder, Fabian Hüttig, and Constanze Keutel. 2018. Scanpath comparison in medical image reading skills of dental students: distinguishing stages of expertise development. In Proceedings ...

    2018

  3. [11]

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al . 2024. A survey on evaluation of large language models. ACM transactions on intelligent systems and technology 15, 3 (2024), 1–45

    2024

  4. [12]

    Yuxing Chen, Weijie Wang, Sylvain Lobry, and Camille Kurtz. 2024. An llm agent for automatic geospatial data analysis. arXiv preprint arXiv:2410.18792 (2024)

    2024 arXiv

  5. [13]

    Jacob Cohen. 1960. A cofficient of agreement for nominal scales. Educational and psychological measurement 20, 1 (1960), 37–46

    1960

  6. [14]

    Bosheng Ding, Chengwei Qin, Ruochen Zhao, Tianze Luo, Xinze Li, Guizhen Chen, Wenhan Xia, Junjie Hu, Luu Anh Tuan, and Shafiq Joty. 2024. Data augmentation using llms: Data perspectives, learning paradigms and challenges. In Findings of the Association for Computational Lingui...

    2024

  7. [15]

    X Fang, W Xu, FA Tan, J Zhang, Z Hu, Y Qi, S Nickleach, D Socolinsky, S Sen- gamedu, and C Faloutsos. 2024. Large Language Models (LLMs) on Tabular Data: Prediction, Generation, and Understanding—A Survey. arXiv 2024. arXiv preprint arXiv:2402.17944 (2024)

    2024 arXiv

  8. [16]

    Sven Festag and Cord Spreckelsen. 2021. Semantic anomaly detection in medical time series. In German Medical Data Sciences: Bringing Data to Life . IOS Press, 118–125

    2021

  9. [17]

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Pal: Program-aided language models. In International Conference on Machine Learning . PMLR, 10764–10799

    2023

  10. [18]

    Lisa A Giacumo and Jeroen Bremen. 2016. Emerging evidence on the use of big data and analytics in workplace learning: A systematic literature review. Quarterly Review of Distance Education 17, 4 (2016), 21

    2016

  11. [19]

    Xusen Guo, Qiming Zhang, Junyue Jiang, Mingxing Peng, Meixin Zhu, and Hao Frank Yang. 2024. Towards explainable traffic flow prediction with large language models. Communications in Transportation Research 4 (2024), 100150

    2024

  12. [20]

    Coen Hacking, Hilde Verbeek, Jan PH Hamers, and Sil Aarts. 2023. Comparing text mining and manual coding methods: Analysing interview data on quality of care in long-term care for older adults. Plos one 18, 11 (2023), e0292578

    2023

  13. [21]

    Jun He, Lixia Li, Wenhao Yao, and Hongying Gao. 2024. Exploring Future Ed- ucation: The Innovative Integration and Practice of Multimodal Learning and ChatGPT. In 2024 5th International Conference on Computer Science, Engineering, and Education (CSEE). IEEE, 18–23

    2024

  14. [22]

    Jingxuan He and Martin Vechev. 2023. Large language models for code: Secu- rity hardening and adversarial testing. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security . 1865–1879

    2023

  15. [23]

    Roy S Hessels and Ignace TC Hooge. 2019. Eye tracking in developmental cognitive neuroscience–The good, the bad and the ugly. Developmental cognitive neuroscience 40 (2019), 100710

    2019

  16. [24]

    Hanyao Huang, Ou Zheng, Dongdong Wang, Jiayi Yin, Zijin Wang, Shengxuan Ding, Heng Yin, Chuan Xu, Renjie Yang, Qian Zheng, et al . 2023. ChatGPT for shaping the future of dentistry: the potential of multi-modal large language model. International Journal of Oral Science 15, 1 ...

    2023

  17. [25]

    Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022. Large language models can self-improve.arXiv preprint arXiv:2210.11610 (2022)

    2022 arXiv

  18. [26]

    Wunmin Jang, Hong Gao, Tilman Michaeli, and Enkelejda Kasneci. 2024. Ex- ploring communication dynamics: Eye-tracking analysis in pair programming of computer science education. In Proceedings of the 2024 Symposium on Eye Tracking Research and Applications. 1–7

    2024

  19. [27]

    Enkelejda Kasneci, Hong Gao, Suleyman Ozdel, Virmarie Maquiling, Enkeleda Thaqi, Carrie Lau, Yao Rong, Gjergji Kasneci, and Efe Bozkir. 2024. Intro- duction to Eye Tracking: A Hands-On Tutorial for Students and Practitioners. arXiv:2404.15435 [cs.HC] https://arxiv.org/abs/2404.15435

    2024 arXiv

  20. [28]

    Enkelejda Kasneci, Gjergji Kasneci, Thomas C Kübler, and Wolfgang Rosenstiel

  21. [29]

    Enkelejda Kasneci, Gjergji Kasneci, Ulrich Trautwein, Tobias Appel, Maike Tibus, Susanne M Jaeggi, and Peter Gerjets. 2022. Do your eye movements reveal your performance on an IQ test? A study linking eye movements and socio- demographic information to fluid intelligence. Plos...

    2022

  22. [30]

    Enkelejda Kasneci, Kathrin Sessler, Stefan Küchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Günnemann, Eyke Hüllermeier, Stephan Krusche, Gitta Kutyniok, Tilman Michaeli, Claudia Nerdel, Jürgen Pfeffer, Oleksandra Poquet, Michael Sail...

    2023

  23. [31]

    Fengfeng Ke, Ruohan Liu, Zlatko Sokolikj, Ibrahim Dahlstrom-Hakki, and Maya Israel. 2024. Using eye-tracking in education: review of empirical research and technology. Educational technology research and development 72, 3 (2024), 1383– 1418

    2024

  24. [32]

    Dong-Ho Lee, Jay Pujara, Mohit Sewak, Ryen W White, and Sujay Kumar Jauhar. 2023. Making large language models better data creators. arXiv preprint arXiv:2310.20111 (2023)

    2023 arXiv

  25. [33]

    Chen Li, Weiqi Wang, Jingcheng Hu, Yixuan Wei, Nanning Zheng, Han Hu, Zheng Zhang, and Houwen Peng. 2024. Common 7b language models already possess strong math capabilities. arXiv preprint arXiv:2403.04706 (2024)

    2024 arXiv

  26. [34]

    Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165 1 (2020), 3

    2020 arXiv

  27. [35]

    Tom M Mitchell. 1999. Machine learning and data mining. Commun. ACM 42, 11 (1999), 30–36

    1999

  28. [36]

    Waddah Saeed and Christian Omlin. 2023. Explainable AI (XAI): A systematic meta-survey of current challenges and future opportunities. Knowledge-based systems 263 (2023), 110273

    2023

  29. [37]

    María Consuelo Sáiz Manzanares, Juan José Rodríguez Diez, Raúl Marti- corena Sánchez, Maria Jose Zaparain Yanez, and Rebeca Cerezo Menéndez. 2020. Lifelong learning from sustainable education: An analysis with eye tracking and data mining techniques. Sustainability 12, 5 (2020), 1970

    2020

  30. [38]

    Currier Sarah, Barton Jane, O’Beirne Rónán, and Ryan Ben. 2004. Quality as- surance for digital learning object repositories: issues for the metadata creation process. ALT-J 12, 1 (2004), 5–20. ICMI ’25, October 13–17, 2025, Canberra, ACT, Australia Guo et al

    2004

  31. [39]

    Chandan Singh, Armin Askari, Rich Caruana, and Jianfeng Gao. 2023. Augment- ing interpretable models with large language models during training. Nature Communications 14, 1 (2023), 7913

    2023

  32. [40]

    Vasileios Skaramagkas, Giorgos Giannakakis, Emmanouil Ktistakis, Dimitris Manousos, Ioannis Karatzanis, Nikolaos S Tachos, Evanthia Tripoliti, Kostas Marias, Dimitrios I Fotiadis, and Manolis Tsiknakis. 2021. Review of eye tracking metrics involved in emotional and cognitive p...

    2021

  33. [41]

    Hedda Martina Šola, Fayyaz Hussain Qureshi, and Sarwar Khawaja. 2024. AI Eye-Tracking Technology: A New Era in Managing Cognitive Loads for Online Learners. Education Sciences 14, 9 (2024), 933

    2024

  34. [42]

    Enkelejda Tafaj, Thomas C Kübler, Gjergji Kasneci, Wolfgang Rosenstiel, and Martin Bogdan. 2013. Online classification of eye tracking data for automated analysis of traffic hazard perception. In International Conference on Artificial Neural Networks. Springer, 442–450

    2013

  35. [43]

    Jiahui Wang, Pavlo Antonenko, and Kara Dawson. 2020. Does visual attention to the instructor in online video affect learning and learner perceptions? An eye-tracking analysis. Computers & Education 146 (2020), 103779

    2020

  36. [44]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  37. [45]

    Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun. 2022. Transformers in time series: A survey. arXiv preprint arXiv:2202.07125 (2022)

    2022 arXiv

  38. [46]

    Adam Wright, Justine Pang, Joshua C Feblowitz, Francine L Maloney, Allison R Wilcox, Harley Z Ramelson, Louise I Schneider, and David W Bates. 2011. A method and knowledge base for automated inference of patient problems from structured data in an electronic medical record. Jo...

    2011

  39. [47]

    Kailai Yang, Shaoxiong Ji, Tianlin Zhang, Qianqian Xie, Ziyan Kuang, and Sophia Ananiadou. 2023. Towards interpretable mental health analysis with large lan- guage models. arXiv preprint arXiv:2304.03347 (2023)

    2023 arXiv

  40. [48]

    Yifan Yao, Jinhao Duan, Kaidi Xu, Yuanfang Cai, Zhibo Sun, and Yue Zhang. 2024. A survey on large language model (llm) security and privacy: The good, the bad, and the ugly. High-Confidence Computing (2024), 100211

    2024

  41. [49]

    Xiao Yu, Zexian Zhang, Feifei Niu, Xing Hu, Xin Xia, and John Grundy. 2024. What Makes a High-Quality Training Dataset for Large Language Models: A Practitioners’ Perspective. In Proceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering . 656–668

    2024

  42. [50]

    Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. 2024. Explainability for large lan- guage models: A survey. ACM Transactions on Intelligent Systems and Technology 15, 2 (2024), 1–38

    2024

  43. [2015]

    InArtificial Neural Networks: Methods and Applications in Bio-/Neuroinformatics

    Online recognition of fixations, saccades, and smooth pursuits for auto- mated analysis of traffic hazard perception. InArtificial Neural Networks: Methods and Applications in Bio-/Neuroinformatics. Springer, 411–434

  44. [2021]

    IEEE Transactions on Affective Computing 14, 2 (2021), 1558–1571

    Cross-task and cross-participant classification of cognitive load in an emergency simulation game. IEEE Transactions on Affective Computing 14, 2 (2021), 1558–1571

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.