Pith. sign in

REVIEW 4 major objections 4 minor 87 references

Predicting Human Visual Attention on Words in Source Code

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning a language model's internal attention with eye-tracking data predicts which words in source code a programmer will look at, outperforming prior software-engineering and computer-vision baselines on three datasets.

desk verdict Solid method paper, but the headline claim is over-sold: the evaluation only covers fixated tokens, so attended-vs-unattended discrimination is never tested. read the letter →

arxiv 2607.14535 v2 pith:AVOVUJGQ submitted 2026-07-16 cs.SE

Chia-Yi Su , Collin McMillan This is my paper
classification cs.SE
keywords humanvisualattentioneyetrackingsourcecodeTransformerlanguagemodelpercenttotalgazetimescanpathpredictionreadoutsoftwareengineering
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

The paper claims that a Transformer language model can predict aggregate human visual attention over words in source code by fine-tuning its internal self-attention mechanism with eye-tracking data. The authors introduce a custom loss that aligns the model's attention scores with the percent total gaze time (ptgt) measured from programmers, using both mean squared error and Pearson correlation. Across three eye-tracking datasets (two Java, one C), the model outperforms the graph neural network baseline and a computer vision baseline, with 64%, 16%, and 467% improvements in Pearson correlation. The model also improves scanpath prediction over the previous baseline on a reading task, suggesting it captures order as well as distribution of attention. A sympathetic reader would care because predicting visual attention has practical value for interface design, tooling, and understanding what information programmers need while reading code.

What carries the argument

The key mechanism is the attention readout: the model extracts the average of self-attention probability distributions from attention heads 10 through 16 in the final Transformer layer. The custom loss function then computes a weighted combination of the standard cross-entropy loss, a mean squared error term, and a Pearson correlation term between this readout and the human gaze data (ptgt). This readout is what converts the model's internal attention into a prediction of where a human will look, and the loss is what shapes that readout during fine-tuning.

What would settle it

Run an ablation that replaces the fixed heads-10–16 readout with a learned linear readout trained to predict ptgt from the same final-layer attention; if the learned readout performs as well or better than the fixed slice, then the specific head choice is not load-bearing, and the reported gains cannot be attributed to that design decision.

Watch

Extended reading notes

Core claim

The central claim is that human visual attention over source code can be predicted by averaging the self-attention scores from a specific subset of heads (heads 10 through 16) in the final Transformer layer of a fine-tuned language model. The fine-tuning loss combines standard cross-entropy with mean squared error and Pearson correlation between this attention readout and the human gaze metric ptgt, reshaping the model's internal attention to be more human-like. The paper reports that the resulting model predicts which words programmers fixate on more accurately than the compared baselines on all three datasets, and that its predicted scanpaths on a reading task are closer to human reading o

Load-bearing premise

The whole method depends on the unverified premise that the average of attention heads 10–16 in the final Transformer layer is the right readout for human attention; if that choice is wrong, the custom loss may distort the language model without producing trustworthy attention predictions.

Editorial extensions

If this is right

  • Code editors and development environments could rank or highlight tokens by predicted human attention to improve tool layouts and documentation generation.
  • The approach can be extended to other programming languages and tasks, provided eye-tracking data exists, to build task-specific attentional priors.
  • Because the fine-tuned model remains a standard autoregressive language model, it could serve as a drop-in backbone for code summarization or navigation that implicitly uses human-like attention.
  • The success of the attention readout suggests that certain internal layers of Transformer models encode behaviorally plausible attention, which may inform interpretability research.
  • The improvement in scanpath prediction indicates the model captures fixation order, not just aggregate distribution, opening the door to predicting reading order in code.

Reading between the lines

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

  • A direct next step would be to ablate the choice of heads 10–16 and the final-layer readout; the paper does not test whether other head subsets or layers would yield similar or better predictions, so the specificity of the readout remains an open question.
  • The custom loss may implicitly regularize the language model to attend in a human-like way, which could improve downstream tasks that benefit from human-like focus, such as code summarization; the paper hints at but does not test this transfer.
  • The strikingly large improvement on the bug-localization dataset (467%) suggests the method may be especially valuable for tasks where human attention is scattered and hard to model, but this needs replication across other bug-localization datasets.
  • One could test whether the attentional prior transfers across tasks: fine-tune on one task (e.g., Java code summarization) and evaluate on another (e.g., C bug localization) without further tuning, which would indicate a task-independent human attentional prior.
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

4 major / 4 minor

Summary. The paper proposes a Transformer-based model for predicting human visual attention over source-code words, operationalized as percent total gaze time (ptgt). The model is built by fine-tuning a GPT-2-style code language model (jamm) with a custom loss that combines standard cross-entropy with MSE and Pearson-correlation terms, where the prediction is read out by averaging self-attention heads 10-16 of the final layer. The evaluation compares this model against a software-engineering GNN baseline, a computer-vision baseline, and an ablation without the correlation loss, on three eye-tracking datasets (Wallace, Smith, Rodeghero) using four correlation metrics. A second experiment applies the model to scanpath prediction on the Bansal dataset, comparing against the Bansal scanpath baseline, GPT-5, and Claude. The paper claims consistent improvements over baselines, with the largest gains on the Smith dataset, and a statistically significant improvement in normalized Levenshtein distance on the reading scanpath task.

Significance. If the claims hold, the work would provide a practical way to predict where programmers look while reading code, with applications to code-summarization tools, IDE design, and cognitive studies of software engineering. The paper has several concrete strengths: it releases the source code, evaluates on three independent eye-tracking datasets spanning two languages and two task types, uses leave-one-participant-out evaluation, and includes a transfer task (scanpath prediction) that provides some independent evidence beyond direct ptgt fitting. The novel loss formulation is an interesting and clearly stated idea, and the comparison against the most relevant prior models is appropriate. However, the headline claim is currently broader than what the evaluation actually measures, and several methodological gaps need to be addressed before the quantitative conclusions can be taken at face value.

major comments (4)
  1. [§IV-E and Table IV] The correlation vectors are built exclusively from W', the set of canonical fixation tokens; non-fixated words, whose ptgt is 0 by the definition in §III-A, are never included in a or a-hat. Thus the Pearson correlations in Table IV measure how well the model ranks words already known to have been fixated by at least one participant, not whether the model can distinguish attended from non-attended words. The Introduction defines the task as forecasting 'which words in a program’s source code will be read most often, reread most often, and/or to what degree and in what order,' and the abstract's unqualified claim is not supported by the reported evaluation. Because the same truncation is applied to all baselines, the relative margins are internally consistent, but the absolute claim is overstated. Please include non-fixated tokens with ptgt=0 in the vectors (or clearly reframe the claim t
  2. [§III-C, §IV-B, Table II] The attention readout is a hand-selected slice: heads 10-16 of the final Transformer layer, with layers 14-24 fine-tuned, adopted from Wang et al. No ablation of alternative head/layer choices is provided, and the training hyperparameters in Table II are described as 'based on our experiments' on the same three datasets used for evaluation. Leave-one-participant-out protects against participant-level overfitting but does not protect against dataset-level selection of the readout and hyperparameters. Please add ablations over readout choices and/or a nested validation scheme, or explicitly state which design choices were fixed a priori before inspecting the evaluation datasets.
  3. [§V-A, Table IV] The main correlation results are reported without confidence intervals, bootstrap estimates, or significance tests. For example, in the Smith study the Bansal baseline has Pearson r=0.0438, ours has r=0.2484, and the human reference is 0.2134; with a small number of functions and participants, these differences could easily be within sampling noise. The scanpath experiment uses Wilcoxon tests for Table V, but Table IV has no such quantification. Please add per-participant or per-function confidence intervals (or equivalent uncertainty measures) for each metric before claiming consistent improvement over baselines.
  4. [§IV-E6] The human-reference formula as written appears self-referential. For a held-out participant p*, the text defines a-hat_{f,t}=a_{p*,f,t} and a_{s,f,t}=a_{s,f,t}-a_{p*,f,t}, where a_s ranges over other participants' observations. This computes the correlation between one participant's attention and a residualized version of the other participants' attention, not the similarity of that participant to the group. If implemented literally, the human-reference rows in Table IV are uninterpretable. If this is a typo, please correct the definition and recompute; the comparison to inter-human agreement is load-bearing for the claim that the model approaches human-level consistency.
minor comments (4)
  1. [§III-C, Eq. (5)] The attention-readout formula has an arithmetic inconsistency: with H=16, the sum over h=10 to H includes 7 heads, but the denominator H-10 = 6. The denominator should be H-9 (or the sum should be h=10 to H-1).
  2. [§III-C, Eq. (8)] The scaling factors α_MSE and α_corr are never defined. Please state how they are computed from the ratio between cross-entropy loss and attention loss; without this, the final loss is not fully reproducible.
  3. [§IV-B] The statement that the Rodeghero model is initialized from a model already fine-tuned on Wallace data should be described more carefully; as written, Rodeghero is not evaluated from a fresh training run, which should be acknowledged in the dataset-description section.
  4. [Throughout] Several typos and formatting issues: 'stat-of-the-art' in the Conclusion, 'Disounted' in §VI-C, 'jamasM pre' in §IV-B, and the formula in §IV-E6 (as noted above). Please proofread carefully.

Circularity Check

1 steps flagged · score 6.0 of 10

Attention 'prediction' is evaluated only on fixated tokens, and those same fixation tokens are given as model input; the which-words-attended question is supplied, not predicted.

  1. self definitional [§III-B (input S), §III-C2 (readout K′), §IV-E1 (W′/virtual fixation vectors)]
    "“The input sequence is composed of words from a section of source code, fixation tokens, and tokens next to the fixation tokens code” (§III-B); “Q′ is a set of tokens next to the fixation tokens and K′ is a set of the fixation tokens in the input” (§III-C2); “We define W′ as the subset of W, which represents a set of canonical fixation tokens in a function f” and “We extract a single attention score for each canonical fixation word token from the model” (§IV-E1)."

    The model's readout â is averaged over query tokens Q′ (tokens next to fixations) and key tokens K′ (fixation tokens), so it only produces scores for tokens that the input already labels as fixated. The evaluation builds a and â only over the same fixation-token subset W′_s; non-fixated words (ptgt=0) are never in the correlation vectors. Thus the reported Pearson margins do not test the introduction's core forecast of “which words … will be read most often.” The attended set is an input, and the evaluation is confined to that set, so the claimed word-attention prediction reduces by construction to ranking gaze time among known-fixated words, not to predicting which words are attended. The comparison to baselines is also asymmetric: baselines receive source-code words only, while “ours” re

full rationale

The core supervised fitting is not circular: the custom loss supervises the attention readout with ptgt on training participants and is evaluated on held-out participants, which is ordinary supervised generalization. The scanpath-transfer experiment (Bansal et al. dataset, NLD/NDCG gains) provides independent, non-circular support. The main circularity is the fixation-token leakage: because the model input includes the words that were fixated, and the evaluation is restricted to exactly those words, the “which words are attended” component of the paper's stated prediction target is satisfied by construction, not learned. This affects the interpretation of the headline 64%/16%/467% gains. The self-citations (jammodel [25,63], Bansal et al. baselines [11,23], Rodeghero/Wallace datasets) are ordinary baselines and model choices, not load-bearing circular justifications; the readout-choice citation to Wang et al. [61] is external, though the specific head/layer slice is an unablated hyperparameter choice. Overall: partial circularity in the central claim, with independent content remaining in the held-out regression and scanpath results.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The model is trained end-to-end on human attention, and the primary evaluation is on held-out participants from the same datasets. The main pre-theoretical commitments are the ptgt operationalization, the choice of attention readout, and the assumption that tuning hyperparameters on the benchmark datasets still yields a fair estimate.

free parameters (3)
  • Loss weights λ_MSE, λ_corr = 5 (all datasets)
    Fixed at 5 'based on our experiments'; no principled tuning or separate validation search is described (Table II).
  • Per-dataset training hyperparameters (epochs, learning rate, batch size) = Wallace: 9, 7e-6, 8; Smith: 5, 3e-6, 8; Rodeghero: 7, 1e-5, 4
    Chosen experimentally on the same datasets used for evaluation; leave-one-out does not appear to separate hyperparameter selection from test evaluation.
  • Attention head subset (heads 10-16) and fine-tuned layer range (layers 14-24) = heads 10-16; layers 14-24
    Selected based on Wang et al. [61] and practical batch-size considerations; alternative readouts are not ablated, so the choice is a hand-selected component of the prediction.
assumptions (5)
  • domain assumption ptgt is a valid and sufficient operationalization of human visual attention on code tokens
    Both the training loss and all evaluation metrics are functions of ptgt; if gaze time diverges from cognitive attention, all claims inherit that divergence (Section III-A).
  • domain assumption Self-attention averaged over heads 10-16 of the final layer is the right readout for human-like attention
    Adopted from Wang et al. [61] with no internal validation or ablation of head/layer choices (Section III-C1).
  • domain assumption Leave-one-participant-out correlation on the same datasets used for hyperparameter tuning estimates real predictive performance
    No separate validation set is described; hyperparameters in Table II were chosen with knowledge of these datasets, so the estimate is likely optimistic (Sections III-B, V).
  • domain assumption The re-implemented baselines are faithful to Bansal et al. [23] and Tafasca et al. [27]
    Reimplementation code is not provided; the paper states they are 'faithful' but gives no verification (Section IV-D).
  • standard math Standard statistical and machine-learning background (CCE, MSE, Pearson correlation, Levenshtein distance)
    Used throughout; the paper's formulas match standard definitions and no re-derivation is needed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Human Visual Attention on Words in Source Code." pith.science (2026). https://pith.science/paper/AVOVUJGQ

@misc{pith2026260714535,
  author       = {Pith},
  title        = {Pith review of: Predicting Human Visual Attention on Words in Source Code},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AVOVUJGQ}},
  note         = {Machine review of arXiv:2607.14535}
}
read the original abstract

This paper presents a computational model to predict human visual attention over words in software source code. The visual attention of software engineers when reading source code has long been studied as a means to understand human cognitive processes during software engineering tasks. Predicting this visual attention is important for perfecting user interface design and understanding what information human programmers need. We propose a model of programmer visual attention in which we design a novel loss function that computes similarity between human attention measured during eye tracking experiments and the internal attention of the artificial neural network. We evaluate our model by comparing its outputs to actual eye tracking data from three separate datasets. Two are in the Java programming language and one is in the C programming language. Our model outperforms the baseline in software engineering by 64%, 16%, and 467% in each of these studies according to Pearson correlation. We used scanpath prediction as an example to demonstrate that our model is more capable of the task that requires the understanding of human thought process. Our model achieves a statistically significant improvement over the close baseline in the reading task according to normalized Levenshtein distance and outperforms both Claude and GPT-5 on both reading and writing tasks.

Figures

Figures reproduced from arXiv: 2607.14535 by the authors.

Figure 1
Figure 1. Example Java method compareIds() from the Wallace study followed by heatmaps of aggregate reference human attention and predicted attention from the main and nocorr versions of our model. Darker shades of red indicate higher levels of actual or predicted ptgt. We demonstrate a model that is optimized against an ag￾gregate attention signal (percent total gaze time). In effect the model learns a population-level atten… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

87 extracted references · 2 linked inside Pith

  1. [1]

    Visual attention: The past 25 years,

    M. Carrasco, “Visual attention: The past 25 years,”Vision research, vol. 51, no. 13, pp. 1484–1525, 2011

    2011

  2. [2]

    Eye movements in code reading: Relaxing the linear order,

    T. Busjahn, R. Bednarik, A. Begel, M. Crosby, J. H. Paterson, C. Schulte, B. Sharif, and S. Tamm, “Eye movements in code reading: Relaxing the linear order,” in2015 IEEE 23rd International Conference on Program Comprehension, 2015, pp. 255–265

    2015

  3. [3]

    Programmer visual attention during context-aware code summarization,

    R. Wallace, A. Bansal, Z. Karas, N. Tang, Y . Huang, T. Jia-Jun Li, and C. McMillan, “Programmer visual attention during context-aware code summarization,”IEEE Trans. Softw. Eng., vol. 51, no. 5, p. 1524–1537, May 2025. [Online]. Available: https://doi.org/10.1109/ TSE.2025.3554990

    arXiv 2025

  4. [4]

    Developers’ visuo-spatial mental model and program comprehension,

    A. Bouraffa, G.-L. Fuhrmann, and W. Maalej, “Developers’ visuo-spatial mental model and program comprehension,” in2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE). IEEE, 2023, pp. 1920–1932

    2023

  5. [5]

    Cognitive processes in program comprehension,

    S. Letovsky, “Cognitive processes in program comprehension,”Journal of Systems and software, vol. 7, no. 4, pp. 325–339, 1987

    1987

  6. [6]

    Mental models and software maintenance,

    D. C. Littman, J. Pinto, S. Letovsky, and E. Soloway, “Mental models and software maintenance,”Journal of Systems and Software, vol. 7, no. 4, pp. 341–355, 1987

    1987

  7. [7]

    Program comprehension during software maintenance and evolution,

    A. V on Mayrhauser and A. M. Vans, “Program comprehension during software maintenance and evolution,”Computer, vol. 28, no. 8, pp. 44– 55, 2002

    2002

  8. [8]

    Finding causes of program output with the java whyline,

    A. J. Ko and B. A. Myers, “Finding causes of program output with the java whyline,” inProceedings of the SIGCHI Conference on Human Factors in Computing Systems, ser. CHI ’09. New York, NY , USA: Association for Computing Machinery, 2009, p. 1569–1578. [Online]. Available: https://doi.org/10.1145/1518701.1518942

    arXiv 2009

Show all 87 references
  1. [9]

    Automated recording and semantics-aware replaying of high-speed eye tracking and interaction data to support cognitive studies of software engineering tasks,

    V . Zyrianov, D. T. Guarnera, C. S. Peterson, B. Sharif, and J. I. Maletic, “Automated recording and semantics-aware replaying of high-speed eye tracking and interaction data to support cognitive studies of software engineering tasks,” in2020 IEEE International Conference on S...

    2020

  2. [10]

    Eye-tracking based adaptive user interface: implicit human-computer interaction for preference indication,

    S. Cheng and Y . Liu, “Eye-tracking based adaptive user interface: implicit human-computer interaction for preference indication,”Journal on Multimodal User Interfaces, vol. 5, no. 1, pp. 77–84, 2012

    2012

  3. [11]

    Modeling programmer attention as scanpath prediction,

    A. Bansal, C.-Y . Su, Z. Karas, Y . Zhang, Y . Huang, T. J.-J. Li, and C. McMillan, “Modeling programmer attention as scanpath prediction,” in2023 38th IEEE/ACM International Conference on Automated Soft- ware Engineering (ASE). IEEE, 2023, pp. 1732–1736

    2023

  4. [12]

    Eyetrans: Merging human and machine attention for neural code summarization,

    Y . Zhang, J. Li, Z. Karas, A. Bansal, T. J.-J. Li, C. McMillan, K. Leach, and Y . Huang, “Eyetrans: Merging human and machine attention for neural code summarization,”Proc. ACM Softw. Eng., vol. 1, no. FSE, Jul. 2024. [Online]. Available: https://doi.org/10.1145/3643732

    2024 doi

  5. [13]

    Using developer eye movements to externalize the mental model used in code summarization tasks,

    N. J. Abid, J. I. Maletic, and B. Sharif, “Using developer eye movements to externalize the mental model used in code summarization tasks,” inProceedings of the 11th ACM Symposium on Eye Tracking Research & Applications, ser. ETRA ’19. New York, NY , USA: Association for Compu...

    2019

  6. [14]

    Expertise-dependent visual attention strategies develop over time during debugging with multiple code representations,

    R. Bednarik, “Expertise-dependent visual attention strategies develop over time during debugging with multiple code representations,”Inter- national Journal of Human-Computer Studies, vol. 70, no. 2, pp. 143– 155, 2012

    2012

  7. [15]

    A tale of two comprehensions? analyzing student programmer attention during code summarization,

    Z. Karas, A. Bansal, Y . Zhang, T. Li, C. McMillan, and Y . Huang, “A tale of two comprehensions? analyzing student programmer attention during code summarization,”ACM Transactions on Software Engineering and Methodology, vol. 33, no. 7, pp. 1–37, 2024

    2024

  8. [16]

    A systematic literature review on the usage of eye-tracking in software engineering,

    Z. Sharafi, Z. Soh, and Y .-G. Gu ´eh´eneuc, “A systematic literature review on the usage of eye-tracking in software engineering,”Information and Software Technology, vol. 67, pp. 79–107, 2015

    2015

  9. [17]

    Dgaze: Cnn-based gaze prediction in dynamic scenes,

    Z. Hu, S. Li, C. Zhang, K. Yi, G. Wang, and D. Manocha, “Dgaze: Cnn-based gaze prediction in dynamic scenes,”IEEE transactions on visualization and computer graphics, vol. 26, no. 5, pp. 1902–1911, 2020

    1902

  10. [18]

    Deep learning based eye gaze estimation and prediction,

    P. L. Mazzeo, D. D’Amico, P. Spagnolo, and C. Distante, “Deep learning based eye gaze estimation and prediction,” in2021 6th International Conference on Smart and Sustainable Technologies (SpliTech). IEEE, 2021, pp. 1–6

    2021

  11. [19]

    Gaze prediction for recommender systems,

    Q. Zhao, S. Chang, F. M. Harper, and J. A. Konstan, “Gaze prediction for recommender systems,” inProceedings of the 10th ACM Conference on Recommender Systems, 2016, pp. 131–138

    2016

  12. [20]

    Critical review of the models of reading comprehension with a focus on situation models,

    M. Davoudi and H. R. H. Moghadam, “Critical review of the models of reading comprehension with a focus on situation models,”International Journal of Linguistics, vol. 7, no. 5, pp. 172–187, 2015

    2015

  13. [21]

    Models of the reading process,

    K. Rayner and E. D. Reichle, “Models of the reading process,”Wiley Interdisciplinary Reviews: Cognitive Science, vol. 1, no. 6, pp. 787–799, 2010

    2010

  14. [22]

    Do transformer models show similar attention patterns to task-specific human gaze?

    O. Eberle, S. Brandl, J. Pilot, and A. Søgaard, “Do transformer models show similar attention patterns to task-specific human gaze?” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2022, pp. 4295– 4309

    2022

  15. [23]

    Towards modeling human attention from eye movements for neural source code summarization,

    A. Bansal, B. Sharif, and C. McMillan, “Towards modeling human attention from eye movements for neural source code summarization,” Proc. ACM Hum.-Comput. Interact., vol. 7, no. ETRA, May 2023. [Online]. Available: https://doi.org/10.1145/3591136

    2023 doi

  16. [24]

    Improving automated source code summarization via an eye-tracking study of programmers,

    P. Rodeghero, C. McMillan, P. W. McBurney, N. Bosch, and S. D’Mello, “Improving automated source code summarization via an eye-tracking study of programmers,” inProceedings of the 36th International Conference on Software Engineering, ser. ICSE 2014. New York, NY , USA: Associ...

    2014

  17. [25]

    Distilled gpt for source code summariza- tion,

    C.-Y . Su and C. McMillan, “Distilled gpt for source code summariza- tion,”Automated Software Engineering, vol. 31, no. 1, p. 22, 2024

    2024

  18. [26]

    Human attention during localization of memory bugs in c programs,

    E. Smith, R. Wallace, M. Robison, Y . Huang, and C. McMillan, “Human attention during localization of memory bugs in c programs,”arXiv preprint arXiv:2506.00693, 2025

    2025 arXiv

  19. [27]

    Toward semantic gaze target detection,

    S. Tafasca, A. Gupta, V . Bros, and J.-M. Odobez, “Toward semantic gaze target detection,”Advances in neural information processing systems, vol. 37, pp. 121 422–121 448, 2024

    2024

  20. [28]

    Combining eye tracking with navigation paths for iden- tification of cross-language code dependencies,

    M. Konopka, “Combining eye tracking with navigation paths for iden- tification of cross-language code dependencies,” inProceedings of the 2015 10th Joint Meeting on F oundations of Software Engineering, 2015, pp. 1057–1059

    2015

  21. [29]

    Leveraging biometric data to boost software developer productivity,

    T. Fritz and S. C. M ¨uller, “Leveraging biometric data to boost software developer productivity,” in2016 IEEE 23rd International Conference on Software Analysis, Evolution, and Reengineering (SANER), vol. 5, 2016, pp. 66–77

    2016

  22. [30]

    Using (bio)metrics to predict code quality online,

    S. C. M ¨uller and T. Fritz, “Using (bio)metrics to predict code quality online,” inProceedings of the 38th International Conference on Software Engineering, ser. ICSE ’16. New York, NY , USA: Association for Computing Machinery, 2016, p. 452–463. [Online]. Available: https://...

    2016

  23. [31]

    Eye gaze and interaction contexts for change tasks– observations and potential,

    K. Kevic, B. M. Walters, T. R. Shaffer, B. Sharif, D. C. Shepherd, and T. Fritz, “Eye gaze and interaction contexts for change tasks– observations and potential,”Journal of Systems and Software, vol. 128, pp. 252–266, 2017

    2017

  24. [32]

    Eye movements in software traceability link recovery,

    B. Sharif, J. Meinken, T. Shaffer, and H. Kagdi, “Eye movements in software traceability link recovery,”Empirical Software Engineering, vol. 22, no. 3, pp. 1063–1102, 2017

    2017

  25. [33]

    Toward a definition of cognitive-driven development,

    A. L. O. T. de Souza and V . H. S. C. Pinto, “Toward a definition of cognitive-driven development,” in2020 IEEE International Conference on Software Maintenance and Evolution (ICSME). IEEE, 2020, pp. 776–778

    2020

  26. [34]

    Program comprehension and code complexity metrics: An fmri study,

    N. Peitek, S. Apel, C. Parnin, A. Brechmann, and J. Siegmund, “Program comprehension and code complexity metrics: An fmri study,” in2021 IEEE/ACM 43rd International Conference on Software Engineering (ICSE), 2021, pp. 524–536

    2021

  27. [35]

    Eyes on code: A study on developers’ code navigation strategies,

    Z. Sharafi, I. Bertram, M. Flanagan, and W. Weimer, “Eyes on code: A study on developers’ code navigation strategies,”IEEE Transactions on Software Engineering, vol. 48, no. 5, pp. 1692–1704, 2022

    2022

  28. [36]

    Estimating developers’ cognitive load at a fine-grained level using eye-tracking measures,

    A. Abbad-Andaloussi, T. Sorg, and B. Weber, “Estimating developers’ cognitive load at a fine-grained level using eye-tracking measures,” in Proceedings of the 30th IEEE/ACM international conference on program comprehension, 2022, pp. 111–121

    2022

  29. [37]

    Applying machine learning to gaze data in software development: a mapping study,

    P. Kuang, E. S ¨oderberg, D. C. Niehorster, and M. H ¨ost, “Applying machine learning to gaze data in software development: a mapping study,” inProceedings of the 2023 Symposium on Eye Tracking Research and Applications, 2023, pp. 1–7

    2023

  30. [38]

    Program code navigation model for individuals based on lstm with co-clustering,

    M. Sun and M. Nakayama, “Program code navigation model for individuals based on lstm with co-clustering,” inProceedings of the 2023 Symposium on Eye Tracking Research and Applications, 2023, pp. 1–6

    2023

  31. [39]

    eye2vec: Learning distributed representations of eye movement for program comprehension analysis,

    H. Yoshioka, K. Shimari, H. Uwano, and K. Matsumoto, “eye2vec: Learning distributed representations of eye movement for program comprehension analysis,” inProceedings of the 2025 Symposium on Eye Tracking Research and Applications, 2025, pp. 1–3

    2025

  32. [40]

    Modeling programmer attention as scanpath prediction,

    A. Bansal, C.-Y . Su, Z. Karas, Y . Zhang, Y . Huang, T. J.-J. Li, and C. McMillan, “Modeling programmer attention as scanpath prediction,” inProceedings of the 38th IEEE/ACM International Conference on Automated Software Engineering, ser. ASE ’23. IEEE Press, 2024, JOURNAL OF...

    2024

  33. [41]

    Follow-up attention: An empirical study of developer and neural model code exploration,

    M. Paltenghi, R. Pandita, A. Z. Henley, and A. Ziegler, “Follow-up attention: An empirical study of developer and neural model code exploration,”IEEE Transactions on Software Engineering, 2024

    2024

  34. [42]

    Comcat: Expertise-guided context generation to enhance code comprehension,

    S. Grandel, S. T. Andersen, Y . Huang, and K. Leach, “Comcat: Expertise-guided context generation to enhance code comprehension,” ACM Transactions on Software Engineering and Methodology, 2025

    2025

  35. [43]

    Eyemulator: Improving code language models by mimicking human visual attention,

    Y . Zhang, C. Huang, Y . Zhang, J. Zhang, T. J.-J. Li, C. McMillan, K. Leach, and Y . Huang, “Eyemulator: Improving code language models by mimicking human visual attention,”arXiv preprint arXiv:2508.16771, 2025

    2025 arXiv

  36. [44]

    Affect detection: An interdisciplinary review of models, methods, and their applications,

    R. A. Calvo and S. D’Mello, “Affect detection: An interdisciplinary review of models, methods, and their applications,”IEEE Transactions on affective computing, vol. 1, no. 1, pp. 18–37, 2010

    2010

  37. [45]

    Automated detection of engagement using video-based estimation of facial expres- sions and heart rate,

    H. Monkaresi, N. Bosch, R. A. Calvo, and S. K. D’Mello, “Automated detection of engagement using video-based estimation of facial expres- sions and heart rate,”IEEE Transactions on Affective Computing, vol. 8, no. 1, pp. 15–28, 2016

    2016

  38. [46]

    Best practices in eye tracking research,

    B. T. Carter and S. G. Luke, “Best practices in eye tracking research,” International Journal of Psychophysiology, vol. 155, pp. 49–62, 2020

    2020

  39. [47]

    Eye tracking studies in visualization: Phases, guidelines, and checklist,

    M. Burch, K. Kurzhals, and D. Weiskopf, “Eye tracking studies in visualization: Phases, guidelines, and checklist,” inProceedings of the 2025 Symposium on Eye Tracking Research and Applications, 2025, pp. 1–7

    2025

  40. [48]

    A practical guide on conducting eye tracking studies in software engineering,

    Z. Sharafi, B. Sharif, Y .-G. Gu ´eh´eneuc, A. Begel, R. Bednarik, and M. Crosby, “A practical guide on conducting eye tracking studies in software engineering,”Empirical Software Engineering, vol. 25, no. 5, pp. 3128–3174, 2020

    2020

  41. [49]

    On eye tracking in software engineering,

    L. Grabinger, F. Hauser, C. Wolff, and J. Mottok, “On eye tracking in software engineering,”SN Computer Science, vol. 5, no. 6, p. 729, 2024

    2024

  42. [50]

    A survey on the usage of eye-tracking in computer programming,

    U. Obaidellah, M. Al Haek, and P. C.-H. Cheng, “A survey on the usage of eye-tracking in computer programming,”ACM Computing Surveys (CSUR), vol. 51, no. 1, pp. 1–58, 2018

    2018

  43. [51]

    Cognitive insights into document comprehension: The role of reading order and visual attention in human and large language models,

    Q. Wang, H. Wang, H. Zhang, C. Chu, R. Wang, and P. Zhu, “Cognitive insights into document comprehension: The role of reading order and visual attention in human and large language models,” inProceedings of the Annual Meeting of the Cognitive Science Society, vol. 47, 2025

    2025

  44. [52]

    Modeling the human visual sys- tem: Comparative insights from response-optimized and task-optimized vision models, language models, and different readout mechanisms,

    S. Saha, I. Chadha, and M. Khosla, “Modeling the human visual sys- tem: Comparative insights from response-optimized and task-optimized vision models, language models, and different readout mechanisms,” arXiv preprint arXiv:2410.14031, 2024

    2024

  45. [53]

    Predicting human gaze beyond pixels,

    J. Xu, M. Jiang, S. Wang, M. S. Kankanhalli, and Q. Zhao, “Predicting human gaze beyond pixels,”Journal of vision, vol. 14, no. 1, pp. 28–28, 2014

    2014

  46. [54]

    Programmers’ visual attention on function call graphs during code summarization,

    S. McLoughlin, Z. Karas, R. Wallace, A. Bansal, C. McMillan, and Y . Huang, “Programmers’ visual attention on function call graphs during code summarization,” inProceedings of the 40th IEEE/ACM International Conference on Automated Software Engineering (ASE ’25). Nov. 16–20: I...

    2025

  47. [55]

    Do machines and humans focus on similar code? exploring explainability of large language models in code summarization,

    J. Li, Y . Zhang, Z. Karas, C. McMillan, K. Leach, and Y . Huang, “Do machines and humans focus on similar code? exploring explainability of large language models in code summarization,” inProceedings of the 32nd IEEE/ACM International Conference on Program Comprehension, 2024...

    2024

  48. [56]

    Eyettention: An attention-based dual-sequence model for predicting human scanpaths during reading,

    S. Deng, D. R. Reich, P. Prasse, P. Haller, T. Scheffer, and L. A. J ¨ager, “Eyettention: An attention-based dual-sequence model for predicting human scanpaths during reading,”Proceedings of the ACM on Human- Computer Interaction, vol. 7, no. ETRA, pp. 1–24, 2023

    2023

  49. [57]

    Toward a model of eye movement control in reading

    E. D. Reichle, A. Pollatsek, D. L. Fisher, and K. Rayner, “Toward a model of eye movement control in reading.”Psychological review, vol. 105, no. 1, p. 125, 1998

    1998

  50. [58]

    Swift: a dynamical model of saccade generation during reading

    R. Engbert, A. Nuthmann, E. M. Richter, and R. Kliegl, “Swift: a dynamical model of saccade generation during reading.”Psychological review, vol. 112, no. 4, p. 777, 2005

    2005

  51. [59]

    Learning where to look: Modeling eye movements in reading,

    M. Nilsson and J. Nivre, “Learning where to look: Modeling eye movements in reading,” inProceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL-2009), 2009, pp. 93–101

    2009

  52. [60]

    Entropy-driven evaluation of models of eye movement control in reading,

    ——, “Entropy-driven evaluation of models of eye movement control in reading,” inProceedings of the 8th International NLPCS Workshop, 2011, pp. 201–212

    2011

  53. [61]

    Probing large language models from a human behavioral perspective,

    X. Wang, X. Li, X. Li, and C. Biemann, “Probing large language models from a human behavioral perspective,” inProceedings of the Workshop: Bridging Neurons and Symbols for Natural Language Processing and Knowledge Graphs Reasoning (NeusymBridge)@ LREC-COLING-2024, 2024, pp. 1–7

    2024

  54. [62]

    A comprehensive overview of large language models,

    H. Naveed, A. U. Khan, S. Qiu, M. Saqib, S. Anwar, M. Usman, N. Akhtar, N. Barnes, and A. Mian, “A comprehensive overview of large language models,”ACM Transactions on Intelligent Systems and Technology, vol. 16, no. 5, pp. 1–72, 2025

    2025

  55. [63]

    A language model of java methods with train/test deduplication,

    C.-Y . Su, A. Bansal, V . Jain, S. Ghanavati, and C. McMillan, “A language model of java methods with train/test deduplication,” in Proceedings of the 31st ACM Joint European Software Engineering Conference and Symposium on the F oundations of Software Engineering, 2023, pp. 2152–2156

    2023

  56. [64]

    A systematic evaluation of large language models of code,

    F. F. Xu, U. Alon, G. Neubig, and V . J. Hellendoorn, “A systematic evaluation of large language models of code,” inProceedings of the 6th ACM SIGPLAN international symposium on machine programming, 2022, pp. 1–10

    2022

  57. [65]

    3d segmentation with fully trainable gabor kernels and pearson’s correlation coefficient,

    K. C. Wong and M. Moradi, “3d segmentation with fully trainable gabor kernels and pearson’s correlation coefficient,” inInternational Workshop on Machine Learning in Medical Imaging. Springer, 2022, pp. 53–61

    2022

  58. [66]

    Modeling the human visual system: Comparative insights from response-optimized and task-optimized vision models, language models, and different readout mechanisms,

    S. Saha, I. Chadha, and M. Khosla, “Modeling the human visual system: Comparative insights from response-optimized and task-optimized vision models, language models, and different readout mechanisms,” inProceedings of the 8th Annual Conference on Cognitive Computational Neuros...

    2025

  59. [67]

    Intersubject synchronization of cortical activity during natural vision,

    U. Hasson, Y . Nir, I. Levy, G. Fuhrmann, and R. Malach, “Intersubject synchronization of cortical activity during natural vision,”science, vol. 303, no. 5664, pp. 1634–1640, 2004

    2004

  60. [68]

    Reliability of cortical activity during natural stimulation,

    U. Hasson, R. Malach, and D. J. Heeger, “Reliability of cortical activity during natural stimulation,”Trends in cognitive sciences, vol. 14, no. 1, pp. 40–48, 2010

    2010

  61. [69]

    Measuring shared responses across subjects using intersubject correlation,

    S. A. Nastase, V . Gazzola, U. Hasson, and C. Keysers, “Measuring shared responses across subjects using intersubject correlation,” pp. 667– 685, 2019

    2019

  62. [70]

    Identifier names in computer programs: Literature review

    I. Herka, “Identifier names in computer programs: Literature review.” Advances in Cognitive Psychology, vol. 19, no. 3, 2023

    2023

  63. [71]

    Fixing faults in c and java source code: Abbreviated vs. full-word identifier names,

    G. Scanniello, M. Risi, P. Tramontana, and S. Romano, “Fixing faults in c and java source code: Abbreviated vs. full-word identifier names,” ACM Transactions on Software Engineering and Methodology (TOSEM), vol. 26, no. 2, pp. 1–43, 2017

    2017

  64. [72]

    Assessing the effect of programming language and task type on eye movements of computer science students,

    N. Mansoor, C. S. Peterson, M. D. Dodd, and B. Sharif, “Assessing the effect of programming language and task type on eye movements of computer science students,”ACM Transactions on Computing Educa- tion, vol. 24, no. 1, pp. 1–38, 2024

    2024

  65. [73]

    Searching and skimming: An ex- ploratory study,

    J. Starke, C. Luce, and J. Sillito, “Searching and skimming: An ex- ploratory study,” in2009 IEEE international conference on software maintenance. IEEE, 2009, pp. 157–166

    2009

  66. [74]

    A comparison of program comprehension strategies by blind and sighted programmers,

    A. Armaly, P. Rodeghero, and C. McMillan, “A comparison of program comprehension strategies by blind and sighted programmers,” inPro- ceedings of the 40th International Conference on Software Engineering, 2018, pp. 788–788

    2018

  67. [75]

    Thinking like a developer? comparing the attention of humans with neural models of code,

    M. Paltenghi and M. Pradel, “Thinking like a developer? comparing the attention of humans with neural models of code,” in2021 36th IEEE/ACM International Conference on Automated Software Engineer- ing (ASE), 2021, pp. 867–879

    2021

  68. [76]

    Do code llms do static analysis?

    C.-Y . Su and C. McMillan, “Do code llms do static analysis?”Empirical Software Engineering, vol. 31, no. 5, p. 116, 2026

    2026

  69. [77]

    Do code summarization models process too much information? function signature may be all that is needed,

    X. Ding, R. Peng, X. Chen, Y . Huang, J. Bian, and Z. Zheng, “Do code summarization models process too much information? function signature may be all that is needed,”ACM Transactions on Software Engineering and Methodology, vol. 33, no. 6, pp. 1–35, 2024

    2024

  70. [78]

    Wohlin, P

    C. Wohlin, P. Runeson, M. H ¨ost, M. C. Ohlsson, B. Regnell, and A. Wessl´en,Experimentation in software engineering. Springer Science & Business Media, 2012

    2012

  71. [79]

    Is attention explanation? an introduction to the debate,

    A. Bibal, R. Cardon, D. Alfter, R. Wilkens, X. Wang, T. Franc ¸ois, and P. Watrin, “Is attention explanation? an introduction to the debate,” inProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2022, pp. 3889– 3900

    2022

  72. [80]

    Attention is not explanation,

    S. Jain and B. C. Wallace, “Attention is not explanation,” inProceedings of the 2019 Conference of the North American Chapter of the Associ- ation for Computational Linguistics: Human Language Technologies, V olume 1 (Long and Short Papers), 2019, pp. 3543–3556

    2019

  73. [81]

    Attention is not not explanation,

    S. Wiegreffe and Y . Pinter, “Attention is not not explanation,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, pp. 11–20

    2019

  74. [82]

    Why attention is not explanation: Surgical intervention and causal reasoning about neural JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13 models,

    C. Grimsley, E. Mayfield, and J. R. Bursten, “Why attention is not explanation: Surgical intervention and causal reasoning about neural JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13 models,” inProceedings of the Twelfth Language Resources and Eval- uation Confer...

    2021

  75. [83]

    Empirical study of transformers for source code,

    N. Chirkova and S. Troshin, “Empirical study of transformers for source code,” inProceedings of the 29th ACM joint meeting on European software engineering conference and symposium on the foundations of software engineering, 2021, pp. 703–715

    2021

  76. [84]

    Code structure–guided transformer for source code summarization,

    S. Gao, C. Gao, Y . He, J. Zeng, L. Nie, X. Xia, and M. Lyu, “Code structure–guided transformer for source code summarization,”ACM Transactions on Software Engineering and Methodology, vol. 32, no. 1, pp. 1–32, 2023

    2023

  77. [85]

    Ex- plaining transformer-based code models: What do they learn? when they do not work?

    A. H. Mohammadkhani, C. Tantithamthavorn, and H. Hemmatif, “Ex- plaining transformer-based code models: What do they learn? when they do not work?” in2023 IEEE 23rd International Working Conference on Source Code Analysis and Manipulation (SCAM). IEEE, 2023, pp. 96–106

    2023

  78. [86]

    An exploratory study on code attention in bert,

    R. Sharma, F. Chen, F. Fard, and D. Lo, “An exploratory study on code attention in bert,” inProceedings of the 30th ieee/acm international conference on program comprehension, 2022, pp. 437–448

    2022

  79. [87]

    Developer reading behavior while summarizing java methods: Size and context matters,

    N. J. Abid, B. Sharif, N. Dragan, H. Alrasheed, and J. I. Maletic, “Developer reading behavior while summarizing java methods: Size and context matters,” in2019 IEEE/ACM 41st International Conference on Software Engineering (ICSE). IEEE, 2019, pp. 384–395

    2019

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.