LEXIC: Lightweight Eye-tracking eXtension via Injected Complexity

# LEXIC: Lightweight Eye-tracking eXtension via Injected Complexity
Source: [https://arxiv.org/html/2607.08152](https://arxiv.org/html/2607.08152)
Sumin LeeKyeonghun KimSubeen LeeJiwon YangSeoul National UniversityOUTTASeoul National UniversitySeoul National Universitycirtuare@snu\.ac\.krkyeonghun\.kim@outta\.ai5584sb@snu\.ac\.krjwyang29@snu\.ac\.kr Tien NguyenKen Ying\-Kai LiaoNam\-Joon Kim†Seoul National UniversityNVIDIASeoul National Universitytien\_noob@snu\.ac\.krkenyingkail@nvidia\.comknj01@snu\.ac\.kr

###### Abstract

On the recent EyeBench benchmark, predicting reading comprehension from eye movements exposes a stark gap: text\-aware models using pretrained language models reach5656–63%63\\%AUROC, while gaze\-only models operate at chance\. We ask how far a gaze\-only model can be pushed by lightweight, language\-model\-free conditioning\. Building on the EyeBench AhnCNN baseline \(LEXIC\-Base\), we propose two mechanisms to inject three precomputed word\-level difficulty signals \(GPT\-2 surprisal, word frequency, and word length\) into the per\-fixation input: direct concatenation \(LEXIC\-Concat\) and a residual mechanism \(LEXIC\-Res\) where a small head predicts typical\-reader gaze response and the encoder is conditioned on the deviation\. On the OneStop reading comprehension task, withK=5K\{=\}5seed\-ensemble training across ten folds, both mechanisms produce statistically consistent AUROC gains on Unseen Text \(\+1\.8\+1\.8to\+2\.2\+2\.2pp, Wilcoxonp≤0\.065p\\leq 0\.065\); LEXIC\-Concat additionally lifts Unseen Reader by\+2\.9\+2\.9pp \(p=0\.010p\{=\}0\.010\)\. We trace an architectural boundary in LEXIC\-Res on Unseen Reader \(\+1\.8\+1\.8pp,p=0\.19p\{=\}0\.19\) to the prediction head being calibrated to training readers, transferring imperfectly to out\-of\-distribution readers\.

††footnotetext:
† Corresponding author## IIntroduction

Eye movements collected during reading reflect cognitive processes ranging from lexical access to discourse comprehension\[[5](https://arxiv.org/html/2607.08152#bib.bib19)\]\. Predicting reader\-level cognitive states or reader–text interactions from these movements has both scientific value for refining cognitive theories and practical value for adaptive interfaces\. The EyeBench benchmark\[[14](https://arxiv.org/html/2607.08152#bib.bib1)\]standardizes these prediction tasks over six datasets, evaluating models across out\-of\-distribution generalization regimes on corpora such as OneStop\[[2](https://arxiv.org/html/2607.08152#bib.bib2)\]and MECO\[[15](https://arxiv.org/html/2607.08152#bib.bib12)\]\.

On the reader–text tasks of EyeBench—specifically binary comprehension classification—predicting cognitive processing has become an active interdisciplinary target\[[8](https://arxiv.org/html/2607.08152#bib.bib15),[7](https://arxiv.org/html/2607.08152#bib.bib18)\]\. The leaderboard exposes a sharp dichotomy: models that ingest text content via large pretrained language models \(PLMs\) reach AUROC in the5656–63%63\\%range\[[19](https://arxiv.org/html/2607.08152#bib.bib6),[13](https://arxiv.org/html/2607.08152#bib.bib5)\], whereas pure gaze\-only models operate at or near chance \(49\.649\.6–50\.0%50\.0\\%\)\[[1](https://arxiv.org/html/2607.08152#bib.bib3),[14](https://arxiv.org/html/2607.08152#bib.bib1)\]\. Models that augment gaze with linguistic features without a language model, like BEyeLSTM\[[12](https://arxiv.org/html/2607.08152#bib.bib4)\], sit between the two regimes \(52\.5%52\.5\\%\)\. Comprehension depends largely on the content being read; thus, gaze\-only models lack text\-side signal\.

We ask: can a gaze\-only model be pushed above chance by injecting*lightweight*text signals, rather than the full text? We use three precomputed word\-level difficulty signals: GPT\-2\[[10](https://arxiv.org/html/2607.08152#bib.bib7)\]surprisal, English word frequency\[[17](https://arxiv.org/html/2607.08152#bib.bib8)\], and word length\. All three are computed offline once per dataset and merged onto fixation rows; no language\-model forward pass is invoked at inference\. The architectural changes to the gaze\-only AhnCNN backbone are minimal: three additional input channels \(concatenation\), or a small auxiliary head predicting typical\-reader gaze, whose residual against the observed gaze augments the input \(residual injection\)\.

Our contributions are:

- •We propose two lightweight mechanisms for injecting precomputed word\-level difficulty features into the gaze\-only AhnCNN baseline of EyeBench, requiring no language\-model inference at deployment\.
- •Across ten cross\-validation folds and three regimes, with aK=5K\{=\}5seed\-ensemble, both mechanisms yield statistically significant AUROC improvements \(Wilcoxonp≤0\.065p\\leq 0\.065\) of 1\.8 to 2\.9 percentage points on Unseen Text and \(for concatenation\) Unseen Reader, reproducing the baseline within0\.30\.3pp\.
- •We characterize an architectural boundary condition: the residual mechanism’s typical\-reader prediction head, calibrated to training readers, attenuates its gain in the Unseen Reader regime — clarifying when the simpler concatenation mechanism is preferable\.

![Refer to caption](https://arxiv.org/html/2607.08152v1/x1.png)Figure 1:Proposed LEXIC variants: \(a\) LEXIC\-Base consumes a 4\-channel fixation tensor\. \(b\) LEXIC\-Concat appends precomputed difficulty signals \(surprisal, frequency, word length\), yielding a 7\-channel input\. \(c\) LEXIC\-Res uses a NormativeGazeHead \(MLP3→32→23\\to 32\\to 2\) to predict typical\-reader gazeg^\\hat\{g\}; the residualrt=gtobs−stopgrad​\(g^t\)r\_\{t\}=g\_\{t\}^\{\\mathrm\{obs\}\}\-\\mathrm\{stopgrad\}\(\\hat\{g\}\_\{t\}\)is concatenated to base channels\. All variants share the downstream backbone\.
## IIRelated Work

Predictive Modeling on Eye Movements:Eye movements have long been studied as indicators of lexical access, attention allocation, and comprehension during reading\[[11](https://arxiv.org/html/2607.08152#bib.bib10)\]\. Recent computational work has moved from handcrafted gaze features to neural models over fixation sequences and scanpaths\[[1](https://arxiv.org/html/2607.08152#bib.bib3)\]\. Ahnet al\.\[[1](https://arxiv.org/html/2607.08152#bib.bib3)\]introduced architectures adopted by EyeBench\[[14](https://arxiv.org/html/2607.08152#bib.bib1)\]as the AhnCNN and AhnRNN baselines; these remain strong gaze\-only baselines but operate near chance on OneStop reading comprehension\. Reichet al\.\[[12](https://arxiv.org/html/2607.08152#bib.bib4)\]introduced BEyeLSTM, ingesting fixation\-level and word\-level linguistic features, sitting between the gaze\-only and text\-aware regimes\. EyeBench v1\.0\[[14](https://arxiv.org/html/2607.08152#bib.bib1)\]also evaluates text\-aware models integrating PLMs\[[19](https://arxiv.org/html/2607.08152#bib.bib6),[13](https://arxiv.org/html/2607.08152#bib.bib5)\], which dominate the leaderboard at high inference costs\.

Lightweight Linguistic Features:Word\-level surprisal, frequency, and length are standard psycholinguistic predictors of fixation duration\. Surprisal traces back to Hale’s expectation\-based account\[[4](https://arxiv.org/html/2607.08152#bib.bib11)\]and its logarithmic relationship with reading time\[[16](https://arxiv.org/html/2607.08152#bib.bib16)\]; frequency and length have independent and well\-documented effects on fixation durations and skipping probabilities\[[6](https://arxiv.org/html/2607.08152#bib.bib21)\]\. We use these signals because they strongly correlate with fixation behavior while remaining computationally lightweight: all features are precomputed offline and reused at inference without any runtime language\-model execution\.

Position of This Work:Rather than maximizing leaderboard performance, we study how much performance gain can be obtained under a strict lightweight\-inference constraint\. Our constraint is no language\-model forward pass at inference; we examine the lift achievable from a chance\-level baseline under that constraint\.

## IIIMethod

### III\-ABackbone: LEXIC\-Base

We use the EyeBench AhnCNN baseline\[[1](https://arxiv.org/html/2607.08152#bib.bib3),[14](https://arxiv.org/html/2607.08152#bib.bib1)\]as our backbone\. LEXIC\-Base consumes a per\-fixation tensor of four channels—duration, pupil size, and position \(xx,yy\)—through three 1D convolutional layers \(kernel size 3, ReLU\)\. A max\-pooling layer, dropout, and flattening lead to a three\-Linear\-layer fully connected classifier head, producing logits\. We preserve the architecture and optimization protocol of the published baseline\[[14](https://arxiv.org/html/2607.08152#bib.bib1)\]\.

### III\-BDifficulty Features

For each word we use three difficulty signals precomputed by the EyeBench pipeline:

- •gpt2\_surprisal: surprisal under GPT\-2 \(base\)\[[10](https://arxiv.org/html/2607.08152#bib.bib7)\]\.
- •wordfreq\_frequency: English frequency fromwordfreq\[[17](https://arxiv.org/html/2607.08152#bib.bib8)\]\.
- •word\_length: character length of the word\.

These signals are defined at the word level and are never regenerated at inference\.

### III\-CConcatenation Mechanism \(LEXIC\-Concat\)

The simplest injection appends the three difficulty signals as additional channels, expanding the input width from 4 to 7 via the inheritedfixation\_dimmechanism\. We apply anan\_to\_numpass in the forward step, replacing any NaN entries with zero to prevent NaNs on unaligned punctuation\.

### III\-DResidual Mechanism \(LEXIC\-Res\)

This mechanism interprets gaze as the deviation from what a typical reader would produce given local difficulty\. A*NormativeGazeHead*\(MLP3→32→23\\to 32\\to 2\) maps difficulty𝐝t∈ℝ3\\mathbf\{d\}\_\{t\}\\in\\mathbb\{R\}^\{3\}to typical\-reader predictiong^t∈ℝ2\\hat\{g\}\_\{t\}\\in\\mathbb\{R\}^\{2\}\(duration and saccade amplitude\)\. The residual is:

rt=gtobs−stopgrad​\(g^t\),r\_\{t\}\\;=\\;g\_\{t\}^\{\\mathrm\{obs\}\}\\;\-\\;\\mathrm\{stopgrad\}\\bigl\(\\hat\{g\}\_\{t\}\\bigr\),\(1\)wheregtobs∈ℝ2g\_\{t\}^\{\\mathrm\{obs\}\}\\in\\mathbb\{R\}^\{2\}is the observed value\. The CNN input is the concatenation of the 4 base channels and 2 residual channels \(6 total\)\. The NormativeGazeHead is shaped through an auxiliary MSE loss:

ℒaux=1\|𝒯0\|​∑t∈𝒯0‖g^t−gtobs‖22\.\\mathcal\{L\}\_\{\\mathrm\{aux\}\}=\\frac\{1\}\{\|\\mathcal\{T\}\_\{0\}\|\}\\sum\_\{t\\in\\mathcal\{T\}\_\{0\}\}\\bigl\\lVert\\hat\{g\}\_\{t\}\-g\_\{t\}^\{\\mathrm\{obs\}\}\\bigr\\rVert\_\{2\}^\{2\}\.\(2\)This loss is restricted to Incorrect\-answer trials \(𝒯0\\mathcal\{T\}\_\{0\}\)\. The total loss is:

ℒ=ℒcls\+λ​ℒaux,λ=0\.3\.\\mathcal\{L\}\\;=\\;\\mathcal\{L\}\_\{\\mathrm\{cls\}\}\\;\+\\;\\lambda\\,\\mathcal\{L\}\_\{\\mathrm\{aux\}\},\\qquad\\lambda=0\.3\.\(3\)λ=0\.3\\lambda=0\.3was inherited from the original residual head specification and held fixed across all experiments\.

### III\-EInference Cost

Both mechanisms perform a single backbone forward pass\. LEXIC\-Res adds a forward pass through a tiny3→32→23\\to 32\\to 2MLP\. Because the difficulty signals are text\-side features computed once per stimulus and shared across readers, no language model is invoked at inference\.

## IVExperimental Setup

Dataset and Task:We use the OneStop reading comprehension task from EyeBench v1\.0\[[14](https://arxiv.org/html/2607.08152#bib.bib1)\], based on the OneStop Eye Movements corpus\[[2](https://arxiv.org/html/2607.08152#bib.bib2)\], focusing on binary classification of whether a reader correctly answered a comprehension question \(\[Incorrect, Correct\]\)\. As standardized by EyeBench\[[14](https://arxiv.org/html/2607.08152#bib.bib1)\], OneStop comprises 180 participants reading passages totaling 19,428 words \(1\.1M fixations\)\. The task contains 9,718 trial instances \(54 passages per participant\)\. We use the Ordinary Reading subset\.

Cross\-Validation Protocol:EyeBench defines ten cross\-validation folds evaluated under three out\-of\-distribution regimes:*Unseen Text*\(unseen texts, seen readers\),*Unseen Reader*\(seen texts, unseen readers\), and*Unseen Both*\. We evaluate all ten folds\. For each fold, we train aK=5K\{=\}5independent seed ensemble, averaging per\-trial probabilities\. Run\-level selection is based on validation cross\-entropy loss\.

Metrics and Threshold Policy:AUROCis our primary threshold\-free metric, directly comparable to the published AhnCNN leaderboard \(49\.7%49\.7\\%\)\.Balanced accuracy at 0\.5\(balacc0\.5\) enables comparison with the leaderboard’s balanced accuracy \(50\.0%50\.0\\%\)\.Balanced accuracy at a tuned threshold\(balaccτ\) serves as a secondary metric\. Per fold,τ\\tauis selected to maximize balanced accuracy on validation trials pooled across the three regimes, and applied unchanged to each regime’s test trials\. Pooling stabilizes threshold selection given limited per\-regime validation size\. We report paired deltas, sign consistency, and Wilcoxon signed\-rank testpp\-values\[[18](https://arxiv.org/html/2607.08152#bib.bib9),[3](https://arxiv.org/html/2607.08152#bib.bib20)\]\.

## VResults

Baseline Reproduction:LEXIC\-Base closely reproduces the published AhnCNN baseline on EyeBench\[[14](https://arxiv.org/html/2607.08152#bib.bib1)\]: AUROC0\.4890\.489–0\.5120\.512across regimes \(leaderboard:0\.4970\.497\) and balacc0\.5of 0\.500 \(leaderboard:0\.5000\.500\), validating our pipeline\.

Main Results:Table[I](https://arxiv.org/html/2607.08152#S5.T1)shows absolute performance\. Both mechanisms lift AUROC above the chance baseline: to0\.510\.51–0\.560\.56\. Under threshold tuning, balaccτincreases to0\.510\.51–0\.540\.54, reaching the same performance tier as BEyeLSTM \(published52\.5%52\.5\\%\[[14](https://arxiv.org/html/2607.08152#bib.bib1),[12](https://arxiv.org/html/2607.08152#bib.bib4)\]\), which consumes significantly richer global and linguistic features\.

TABLE I:Test AUROC and balanced accuracy at tuned threshold \(balaccτ\), reported as mean±\\pmstandard deviation across ten folds \(K=5K\{=\}5seed ensemble\)\. Best means are bolded\.Paired Deltas and Sign Consistency:Table[II](https://arxiv.org/html/2607.08152#S5.T2)details the paired within\-fold gains\. LEXIC\-Concat shows significant AUROC gains on Unseen Text \(\+1\.82\+1\.82pp, 9/10,p=0\.065p\{=\}0\.065\) and Unseen Reader \(\+2\.92\+2\.92pp, 8/10,p=0\.010p\{=\}0\.010\)\. LEXIC\-Res achieves a significant AUROC gain on Unseen Text \(\+2\.18\+2\.18pp, 9/10,p=0\.010p\{=\}0\.010\), but a weaker trend on Unseen Reader \(\+1\.84\+1\.84pp, 7/10,p=0\.19p\{=\}0\.19\)\. Under threshold tuning, balaccτdeltas are uniformly larger and statistically significant atp≤0\.05p\\leq 0\.05in five of six cells\.

TABLE II:Paired delta versus baseline within fold\. The sign indicates positive\-delta folds out of ten\.Boldvalues indicate statistically significant positive improvements \(Δ\>0\\Delta\>0andp≤0\.10p\\leq 0\.10\)\.MechanismRegimeΔ\\Delta\(pp\)signppAUROCLEXIC\-ConcatUnseen Text\+1\.82\\mathbf\{\+1\.82\}9/109/100\.0650\.065Unseen Reader\+2\.92\\mathbf\{\+2\.92\}8/108/100\.0100\.010Unseen Both\+0\.33\+0\.335/105/101\.0001\.000LEXIC\-ResUnseen Text\+2\.18\\mathbf\{\+2\.18\}9/109/100\.0100\.010Unseen Reader\+1\.84\+1\.847/107/100\.1930\.193Unseen Both\+4\.44\+4\.446/106/100\.2750\.275Balanced accuracy at tuned threshold \(balaccτ\)LEXIC\-ConcatUnseen Text\+1\.15\\mathbf\{\+1\.15\}9/109/100\.0840\.084Unseen Reader\+3\.65\\mathbf\{\+3\.65\}9/109/100\.0200\.020Unseen Both\+5\.51\\mathbf\{\+5\.51\}9/109/100\.0200\.020LEXIC\-ResUnseen Text\+1\.27\\mathbf\{\+1\.27\}9/109/100\.0840\.084Unseen Reader\+2\.28\\mathbf\{\+2\.28\}9/109/100\.0200\.020Unseen Both\+5\.96\\mathbf\{\+5\.96\}7/107/100\.0490\.049Boundary Condition: LEXIC\-Res on Unseen Reader:LEXIC\-Res exhibits a clear Unseen Text\-versus\-Unseen Reader asymmetry \(AUROC gain of\+1\.84\+1\.84pp vs\.\+2\.18\+2\.18pp\)\. The gain on Unseen Reader is roughly half the LEXIC\-Concat gain \(\+2\.92\+2\.92pp\)\. We trace this transfer bottleneck to the typical\-reader prediction head in LEXIC\-Res, which is calibrated to a population\-averaged gaze response on training readers\. For unseen readers, the residualrt=gtobs−g^tr\_\{t\}=g\_\{t\}^\{\\mathrm\{obs\}\}\-\\hat\{g\}\_\{t\}mixes typical\-reader gaze with an out\-of\-distribution reader’s observed gaze, collapsing toward noise\. The direct input injection in LEXIC\-Concat bypasses this bottleneck\.

## VIDiscussion

Closing Part of the Gap:Without invoking a language model at inference, our lightweight LEXIC variants reach performance tiers comparable to BEyeLSTM \(0\.510\.51–0\.560\.56AUROC\)\. This demonstrates how far gaze\-only models can be pushed without heavy PLM architectures\.

Ranking versus Calibration:While balacc0\.5stays near chance, balaccτimprovements are significant\. Lightweight injection shifts the ranking of trials by predicted probability without affecting default\-threshold calibration\. Deployments with a fixed threshold require calibration \(e\.g\., Platt scaling\[[9](https://arxiv.org/html/2607.08152#bib.bib17)\]\)\.

Practical Recommendation:LEXIC\-Concat is simpler and statistically more robust across generalization regimes, making it the recommended default when reader generalization is expected\. LEXIC\-Res is most effective when the reader population overlaps with training\.

Limitations:We only cover the OneStop reading comprehension task and evaluate on a single backbone \(LEXIC\-Base\)\. Extending our mechanisms to other gaze backbones \(AhnRNN, BEyeLSTM\) and alternative EyeBench reading comprehension tasks \(SB\-SAT, PoTeC\) remain future priorities\.

## VIIConclusion

We presented two lightweight mechanisms for injecting precomputed word\-level difficulty signals into a gaze\-only CNN baseline\. Without PLM inference at deployment, both variants produce statistically consistent gains on the OneStop reading comprehension task\. The direct concatenation mechanism \(LEXIC\-Concat\) is robust across regimes, while the residual mechanism \(LEXIC\-Res\) is constrained by a reader\-transfer bottleneck\. This quantifies the value of lightweight features in gaze\-based cognitive modeling\.

## References

- \[1\]\(2020\)Towards predicting reading comprehension from gaze behavior\.InACM Symposium on Eye Tracking Research and Applications,pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1145/3379156.3391335)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p2.5),[§II](https://arxiv.org/html/2607.08152#S2.p1.1),[§III\-A](https://arxiv.org/html/2607.08152#S3.SS1.p1.2)\.
- \[2\]Y\. Berzak, J\. Malmaud, O\. Shubi, Y\. Meiri, E\. Lion, and R\. Levy\(2025\)OneStop: a 360\-participant english eye tracking dataset with different reading regimes\.Scientific Data12\(1\)\.External Links:[Link](https://doi.org/10.1038/s41597-025-06272-2)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p1.1),[§IV](https://arxiv.org/html/2607.08152#S4.p1.1)\.
- \[3\]J\. Demšar\(2006\)Statistical comparisons of classifiers over multiple data sets\.Journal of Machine Learning Research7,pp\. 1–30\.Cited by:[§IV](https://arxiv.org/html/2607.08152#S4.p3.6)\.
- \[4\]J\. Hale\(2001\)A probabilistic earley parser as a psycholinguistic model\.InProceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics \(NAACL\),pp\. 1–8\.Cited by:[§II](https://arxiv.org/html/2607.08152#S2.p2.1)\.
- \[5\]M\. A\. Just and P\. A\. Carpenter\(1980\)A theory of reading: from eye fixations to comprehension\.Psychological Review87\(4\),pp\. 329–354\.External Links:[Document](https://dx.doi.org/10.1037/0033-295X.87.4.329)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p1.1)\.
- \[6\]R\. Kliegl, E\. Grabner, M\. Rolfs, and R\. Engbert\(2004\-01\)Length, frequency, and predictability effects of words on eye movements in reading\.European Journal of Cognitive Psychology16,pp\. 262–284\.External Links:[Document](https://dx.doi.org/10.1080/09541440340000213)Cited by:[§II](https://arxiv.org/html/2607.08152#S2.p2.1)\.
- \[7\]S\. Makowski, L\. A\. Jäger, A\. Abdelwahab, N\. Landwehr, and T\. Scheffer\(2018\)A discriminative model for identifying readers and assessing text comprehension from eye movements\.InMachine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2018, Proceedings, Part I,pp\. 209–225\.External Links:[Document](https://dx.doi.org/10.1007/978-3-030-10925-7%5F13),[Link](https://doi.org/10.1007/978-3-030-10925-7_13)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p2.5)\.
- \[8\]C\. Mézière, E\. D\. Relchle, T\. von der Malsburg, and G\. McArthur\(2023\)Using eye\-tracking measures to predict reading comprehension\.Reading Research Quarterly58\(3\),pp\. 425–449\.Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p2.5)\.
- \[9\]J\. Platt\(1999\)Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods\.InAdvances in Large Margin Classifiers,Vol\.10,pp\. 61–74\.Cited by:[§VI](https://arxiv.org/html/2607.08152#S6.p2.2)\.
- \[10\]A\. Radford, J\. Wu, R\. Child, D\. Luan, D\. Amodei, and I\. Sutskever\(2019\)Language models are unsupervised multitask learners\.Technical reportOpenAI\.External Links:[Link](https://api.semanticscholar.org/CorpusID:160025533)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p3.1),[1st item](https://arxiv.org/html/2607.08152#S3.I1.i1.p1.1)\.
- \[11\]K\. Rayner\(1998\)Eye movements in reading and information processing: 20 years of research\.Psychological Bulletin124\(3\),pp\. 372–422\.External Links:[Document](https://dx.doi.org/10.1037/0033-2909.124.3.372)Cited by:[§II](https://arxiv.org/html/2607.08152#S2.p1.1)\.
- \[12\]D\. R\. Reich, P\. Prasse, C\. Tschirner, P\. Haller, F\. Goldhammer, and L\. A\. Jäger\.\(2022\)Inferring native and non\-native human reading comprehension and subjective text difficulty from scanpaths in reading\.In2022 Symposium on Eye Tracking Research and Applications \(ETRA\),pp\. 1–8\.External Links:[Document](https://dx.doi.org/10.1145/3517031.3529639)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p2.5),[§II](https://arxiv.org/html/2607.08152#S2.p1.1),[§V](https://arxiv.org/html/2607.08152#S5.p2.6)\.
- \[13\]O\. Shubi, Y\. Meiri, C\. A\. Hadar, and Y\. Berzak\(2024\)Fine\-grained prediction of reading comprehension from eye movements\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),pp\. 3372–3391\.External Links:[Link](https://doi.org/10.18653/v1/2024.emnlp-main.198)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p2.5),[§II](https://arxiv.org/html/2607.08152#S2.p1.1)\.
- \[14\]Shubi,Omer, D\. R\. Reich, K\. Gruteke Klein, Y\. Angel, P\. Prasse, L\. A\. Jäger, and Y\. Berzak\(2025\)EyeBench: predictive modeling from eye movements in reading\.InThe Thirty\-Ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track,External Links:[Link](https://openreview.net/pdf?id=LhbYJJ3MFd/)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p1.1),[§I](https://arxiv.org/html/2607.08152#S1.p2.5),[§II](https://arxiv.org/html/2607.08152#S2.p1.1),[§III\-A](https://arxiv.org/html/2607.08152#S3.SS1.p1.2),[§IV](https://arxiv.org/html/2607.08152#S4.p1.1),[§V](https://arxiv.org/html/2607.08152#S5.p1.5),[§V](https://arxiv.org/html/2607.08152#S5.p2.6)\.
- \[15\]N\. Siegelman, S\. Schroeder, C\. Acartürk, H\. Ahn, S\. Alexeeva, S\. Amenta, R\. Bertram, R\. Bonandrini, M\. Brysbaert, D\. Chernova, S\. M\. Da Fonseca, N\. Dirix, W\. Duyck, A\. s\. Fella, R\. Frost, C\. A\. Gattei, A\. Kalaitzi, N\. Kwon, K\. Lõo, M\. Marelli, T\. C\. Papadopoulos, A\. Protopapas, S\. Savo, D\. E\. Shalom, N\. Slioussar, R\. Stein, L\. Sui, A\. Taboh, V\. Tønnesen, K\. A\. Usal, and V\. Kuperman\(2022\)Expanding horizons of cross\-linguistic research on reading: the multilingual eye\-movement corpus \(meco\)\.Behavior Research Methods54,pp\. 2843–2863\.External Links:[Link](https://doi.org/10.3758/s13428-021-01772-6)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p1.1)\.
- \[16\]N\. J\. Smith and R\. Levy\(2013\)The effect of word predictability on reading time is logarithmic\.Cognition128\(3\),pp\. 302–319\.External Links:[Document](https://dx.doi.org/http%3A//dx.doi.org/10.1016/j.cognition.2013.02.013)Cited by:[§II](https://arxiv.org/html/2607.08152#S2.p2.1)\.
- \[17\]R\. Speer, J\. Chin, A\. Lin, S\. Jewett, and L\. Nathan\(2018\)LuminosoInsight/wordfreq: v2\.2\.Zenodo\.External Links:[Document](https://dx.doi.org/10.5281/zenodo.1443582),[Link](https://doi.org/10.5281/zenodo.1443582)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p3.1),[2nd item](https://arxiv.org/html/2607.08152#S3.I1.i2.p1.1)\.
- \[18\]F\. Wilcoxon\(1945\)Individual comparisons by ranking methods\.Biometrics Bulletin1\(6\),pp\. 80–83\.External Links:[Document](https://dx.doi.org/10.2307/3001968)Cited by:[§IV](https://arxiv.org/html/2607.08152#S4.p3.6)\.
- \[19\]D\. Yang and N\. Hollenstein\(2023\)PLM\-AS: pre\-trained language models augmented with scanpaths for sentiment classification\.InProceedings of the Northern Lights Deep Learning Workshop,External Links:[Link](https://doi.org/10.7557/18.6797)Cited by:[§I](https://arxiv.org/html/2607.08152#S1.p2.5),[§II](https://arxiv.org/html/2607.08152#S2.p1.1)\.