{"id":"204190a4-5035-4bce-a4c9-c4f589d9f06d","arxiv_id":"2412.04184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A spectral-loss-regularized LSTM-CNN GAN generates synthetic eye-gaze velocity trajectories whose statistical moments and autocorrelation closely match real data, outperforming a four-state HMM in this comparison.","lead":"This paper tests a GAN-based system that generates synthetic eye-gaze velocity data, adding a frequency-based spectral loss to improve realism. The best variant, an LSTM generator paired with a CNN discriminator, matched real data statistics much more closely than a hidden Markov model did.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Spectral-loss benefit is measured with the spectral loss itself; no quantitative LG vs Lfinal DJS comparison supports the central novelty claim.","rationale":"The reader's weakest assumption is the absence of a held-out split. I agree that this is a genuine limitation, and it is part of the problem. However, the more load-bearing concern is that the paper's central contribution—spectral regularization—is evaluated using the spectral loss itself. Section 4.1 reports 'Spectral Score' and integrals of L_final and DJS for models trained with L_final, but it never gives quantitative DJS for the LG-only condition. Since L_final minimizes L_spectral by construction, observing a lower spectral score is not evidence of improved data fidelity. The DJS values that are reported (e.g., 0.000322 in Table 2) may be computed on training data, further weakening the generalization claim. These issues can be fixed by rerunning the comparison with held-out data and reporting DJS/ACF for both loss conditions, which is why the verdict should remain CONDITIONAL rather than move to REJECT. The paper does provide code and a clear experimental setup, which allows this check to be performed independently.","tokens_in":17141,"tokens_out":8082,"duration_ms":81622,"concrete_test":"Re-run the four architectures under both LG and L_final, and compute DJS and ACF mean-absolute error on a held-out test split (e.g., 20% of sequences not used in training), reporting values per condition and per architecture. If L_final-trained LSTM-CNN does not yield a lower held-out DJS than LG-trained LSTM-CNN by a margin consistent with the paper's claim, the spectral-regularization benefit is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is the spectral loss L_final = L_G + λ L_spectral (Eq. 10, §2.2). To show it improves fidelity, §4.1 compares the four architectures under LG and Lfinal, but the only quantitative scores presented in Fig. 4 are the spectral score (L_spectral) and integrals of L_final and DJS for models trained with L_final. Because L_final directly penalizes L_spectral, reporting lower L_spectral for L_final-trained models is circular; it does not demonstrate that the generated distribution is closer to real data. The independent metric DJS is shown in Fig. 4, but no DJS values are reported for the LG-only variants of each architecture, so the improvement claim rests on visual inspection of Fig. 3. Moreover, Section 3 describes segmentation into training batches but never describes a held-out test split; DJS and ACF appear to be computed on the same sequences used for training, so the headline DJS=0.000322 does not establish generalization beyond the training distribution. Either issue alone undermines the quantitative support for the abstract's claim; the circular spectral-loss metric is the more direct threat to the paper's contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generative-adversarial-network (GAN) framework with LSTM and CNN generators/discriminators for synthetic eye-gaze velocity trajectories, augmented with a spectral loss term L_spectral (Eq. 8) that penalizes differences in log-magnitude Fourier spectra between real and generated sequences, combined as L_final = L_G + λ L_spectral (Eq. 10). Four GAN architectures (CNN-CNN, LSTM-CNN, CNN-LSTM, LSTM-LSTM) are compared under adversarial-only and spectral-regularized losses, and the best architecture (LSTM-CNN) is compared against a four-state HMM using Jensen-Shannon divergence, autocorrelation, and distribution moments. The central claim is that the spectrally regularized LSTM-CNN GAN 'closely matches the real data' (DJS = 0.000322) while the HMM 'significantly diverges' (DJS = 0.01244), positioning the GAN as a high-fidelity synthetic gaze data generator.","tokens_in":17383,"tokens_out":3433,"duration_ms":33312,"significance":"If the central claim were fully supported, the paper would offer a practical, lightweight GAN-based alternative to traditional Markov models for generating realistic eye-gaze trajectories, with applications in HCI, simulation, and diagnostic tool development. Strengths include the public release of code and data-generation package, the comparison against a standard HMM baseline, and the use of several evaluation metrics (DJS, ACF, moments) that are at least partly independent of the training objective. However, the significance is currently bounded by methodological gaps: the spectral-loss contribution is not established through independent quantitative comparison, the evaluation appears to lack a held-out test set, and no variability measures are reported. These issues are fixable but require additional experiments.","major_comments":[{"comment":"The central novelty claim—that spectral regularization improves fidelity—is not quantitatively demonstrated. The only quantitative scores in Fig. 4 are the 'Spectral Score' (essentially L_spectral from Eq. 8) and integrals of L_final and DJS for models trained with L_final. Because L_final directly minimizes L_spectral, reporting a lower L_spectral for L_final-trained models is circular and does not show that the generated distribution is closer to the real data. The independent metric DJS is shown in Fig. 4 only for L_final-trained models; no DJS, moment, or ACF values are reported for the LG-only variants of each architecture. To support the abstract's claim, the authors should report independent metrics (DJS, moments, ACF) for both LG-only and L_final conditions for every architecture, with confidence intervals over repeated runs.","section":"Section 4.1 and Eq. (10)"},{"comment":"The evaluation does not describe a train/test split. Section 3 states that velocity data were segmented into sequences of 200 points and supplied to the discriminator, but it never specifies a held-out set of real sequences for evaluation. The DJS and ACF calculations appear to use the same sequences that were used for GAN training. If the comparison set overlaps the training set, the reported DJS = 0.000322 may reflect overfitting or memorization rather than generalization to new gaze trajectories. The authors must separate training and test sequences, describe how synthetic sequences are generated (length, number, initialization), and report DJS/ACF on the held-out real data.","section":"Sections 3 and 4"},{"comment":"The reported HMM divergence values are internally inconsistent. Table 2 lists the HMM average DJS as 0.01244, while Section 4.2 states the four-state HMM has DJS = 0.0131 and the five-state HMM has DJS = 0.013245. A later paragraph in the same section refers to a 'higher JS divergence (0.00714)'. These three numbers cannot all be correct for the same model and evaluation. The authors should correct the inconsistencies and specify exactly which DJS value corresponds to the four-state HMM used in the comparison, along with the log base used in Eq. (11) (the claimed bound 0 ≤ DJS ≤ 1 requires base-2 logarithms).","section":"Table 2, Section 4.2"},{"comment":"The article reports no repeated-seed variability or statistical significance tests. Single-run point estimates for DJS, moments, and ACF do not support ranking of architectures or the claim that the HMM 'significantly diverges' from the real data. Given the stochasticity of GAN training, the authors should train each configuration with multiple random seeds, report means and standard deviations (or confidence intervals), and perform a statistical test (e.g., bootstrap or paired test) for the LSTM-CNN versus HMM differences.","section":"Section 4.1 and Table 2"}],"minor_comments":[{"comment":"The text says 'The CNN-LSTM GAN utilized a random noise vector...' but the model being evaluated is the LSTM-CNN architecture; this naming inconsistency should be fixed.","section":"Section 4.2"},{"comment":"The caption and text refer to a 'red trajectory' for the spectral loss and a 'blue line' for DJS, but the figure legend and axis labels are unclear. Please ensure the plot clearly distinguishes the loss components and DJS, and define what 'Spectral Score' means relative to Eq. (8).","section":"Fig. 4(a)"},{"comment":"The data description lacks participant information: number of participants, total recording duration, number of sequences, and how left/right eye data were combined or treated. This information is needed for reproducibility.","section":"Section 3"},{"comment":"The HMM implementation details are incomplete: no mention of the number of training sequences, initial parameter values for the EM algorithm, convergence criteria, or how synthetic HMM sequences were generated (e.g., from the fitted emission distributions).","section":"Section 2.4 and Appendix A"},{"comment":"The reference to 'reproducing stochastic processes' in Lencastre et al. (2023) is relevant, but the introduction does not clearly state which concrete limitations of prior GANs (e.g., mode collapse, poor tail capture) the spectral loss specifically targets. A more precise motivation would strengthen the paper.","section":"Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a practical and timely problem, and the code release is a plus. However, the present evidence is insufficient for the central claims: the spectral-loss benefit is not established via independent metrics, the lack of a held-out test split undermines the generalization claim, and there are internal numeric inconsistencies in the HMM comparison. These are fixable with additional experiments and careful reporting, so I recommend major revision rather than rejection. I would also suggest the editor check whether the journal scope expects a deeper methodological comparison with state-of-the-art time-series GANs (e.g., TimeGAN, RCGAN), as the authors explicitly defer such comparison to future work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent empirical study with a clear result: an LSTM-CNN GAN with spectral regularization beats a four-state HMM at matching the distribution and autocorrelation of eye-gaze velocity data. The new piece is the application of spectral loss to this time-series setting and the systematic architecture comparison. Credit where due: the HMM baseline is properly fitted via Baum-Welch, the comparison uses independent metrics (moments, DJS, ACF), and the gap is large. The code and data availability are a plus, and the writing is clear.\n\nThe soft spots are real but not fatal. The central novelty claim—that spectral loss improves fidelity—is under-supported because the spectral score in Fig. 4 is the very quantity minimized by Lfinal. Reporting lower spectral loss for Lfinal-trained models is partly circular. The paper does not report DJS values for LG-only variants, so the improvement rests on visual inspection of Fig. 3. Also, no train/test split is described; DJS and ACF are computed on the same sequences used for training, so the headline DJS=0.000322 does not establish generalization beyond the training set. That said, the HMM comparison is independent and credible, and the GAN clearly wins on that comparison.\n\nThere is a numeric inconsistency: the text in Section 4.2 reports HMM DJS as 0.00714, while Table 2 says 0.01244 and Fig. 5(c) shows about 0.013 for four hidden states. That needs fixing. Minor: Section 4.2 refers to \"CNN-LSTM GAN\" when the experiment uses LSTM-CNN.\n\nWho is this for? Researchers working on synthetic eye-tracking data, GANs for time series, or comparisons with Markov models. It deserves a serious referee because the empirical question is worthwhile and the main comparison is sound, but it needs revision to support the spectral-loss claims. Send it to peer review with a request for held-out evaluation, DJS values for LG-only variants, and correction of the inconsistency.","headline":"Competent empirical study whose headline GAN-vs-HMM result is credible, but the spectral-loss improvement claim is under-supported by a circular metric and a missing held-out split.","tokens_in":17937,"tokens_out":2199,"would_cite":false,"duration_ms":21188,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A spectrally regularized LSTM-CNN GAN produces synthetic eye-gaze velocity trajectories that match real data far more closely than a hidden Markov model on the reported distributional and temporal metrics.","keywords":["Generative Adversarial Networks","eye gaze velocity","spectral loss","LSTM","CNN","Hidden Markov Model","synthetic time series","Jensen-Shannon divergence"],"falsifier":"Train the LSTM-CNN GAN on one set of gaze-velocity sequences and compute Jensen–Shannon divergence and autocorrelation on a disjoint set from different participants or recording sessions; if $D_{JS}$ on the held-out data rises to the HMM's level or above, the central fidelity claim is not established.","tokens_in":16935,"feed_emoji":"👁","tokens_out":8664,"duration_ms":73599,"temperature":0.7,"pith_summary":"This paper argues that a lightweight GAN pairing an LSTM generator with a CNN discriminator, trained with a spectral loss that compares the log Fourier magnitudes of real and generated sequences, can generate synthetic eye-gaze velocity trajectories that match recorded data in distribution and autocorrelation. The authors report a Jensen–Shannon divergence of $0.000322$ for this model, versus $0.01244$ for a four-state HMM, and show that spectral regularization improves fidelity over adversarial-only training. If correct, the result provides a practical generative tool for simulation environments, training systems, and human-computer interaction research that needs realistic gaze data, and it narrows a previously reported gap in which simple Markov models often beat GANs on stochastic trajectories.","feed_headline":"A hybrid GAN with spectral loss reproduces real gaze velocity data","feed_subtitle":"Fourier-based spectral loss lets an LSTM-CNN GAN outperform a 4-state HMM on gaze-velocity fidelity.","key_machinery":"The load-bearing mechanism is the spectral loss $L_{\\mathrm{spectral}} = \\sum_{k=0}^{N-1} [\\log(|F(x)_k|) - \\log(|F(\\hat{x})_k|)]^2$, which penalizes differences in the log magnitudes of the discrete Fourier transforms of real and generated velocity sequences, normalizing amplitude disparities and focusing on relative spectral energy. It is added to the non-saturating adversarial generator loss as $L_{\\mathrm{final}} = L_G + \\lambda L_{\\mathrm{spectral}}$ with $\\lambda = 0.1$. The generator and discriminator pair LSTM and CNN layers: the LSTM generator converts random noise into a sequence of hidden states that become velocity displacements, capturing temporal dependencies, while the CNN discriminator uses strided convolutions and a sigmoid output to classify trajectory realism. Training uses 200-point sequences, 500 epochs, a 0.0002 learning rate, and fidelity is measured with Jensen–Shannon divergence and autocorrelation.","core_discovery":"The central claim is that a GAN trained to minimize $L_{\\mathrm{final}} = L_G + \\lambda L_{\\mathrm{spectral}}$, where $L_{\\mathrm{spectral}} = \\sum_{k=0}^{N-1}[\\log(|F(x)_k|) - \\log(|F(\\hat{x})_k|)]^2$ and $x$, $\\hat{x}$ are real and generated velocity sequences, reproduces the statistical and spectral properties of real eye-gaze velocity time series. Among the four tested architectures, the LSTM-CNN model showed the closest alignment to the real data distribution, with $D_{JS}=0.000322$; the HMM diverged with $D_{JS}=0.01244$. The paper also reports that LSTM-CNN matches the mean, standard deviation, skewness, and kurtosis of the real data more closely than the HMM and captures the early-lag autocorrelation function. The gain is attributed to the LSTM generator's long-term temporal modeling, the CNN discriminator's local feature extraction, and spectral regularization's emphasis on high- and low-frequency components.","pith_inferences":["The reported fidelity statistics are computed on sequences used to train the GAN; a held-out evaluation on previously unseen participants or recording sessions would test whether the model generalizes or partly memorizes the training data.","The log-magnitude Fourier loss is scale-sensitive in a way that suits heavy-tailed velocity distributions; the same regularization could transfer to other non-stationary physiological time series, such as heart-rate variability or EEG, where spectral shape and tail behavior matter.","The manuscript reports two different HMM divergence values (0.00714 in the comparison text and 0.01244 in Table 2); the magnitude of the claimed gap depends on which value is used.","Because the GAN starts from uniform noise and produces 200-point segments, the generated trajectories are stationary in distribution; assessing whether the model captures longer-range, non-stationary gaze behavior would require comparing on multi-segment concatenated sequences."],"forward_implications":["Synthetic gaze-velocity data with realistic heavy tails and temporal autocorrelation can be produced by a comparatively small GAN, making it usable in simulation environments and eye-tracking training systems.","Adding spectral regularization stabilizes GAN training and improves distributional fidelity for eye-gaze trajectories relative to adversarial loss alone.","Among the four tested architectures, the LSTM-CNN combination gives the best balance between fidelity (low $D_{JS}$ and spectral loss) and per-epoch computation time.","A four-state HMM fitted by Baum-Welch does not reproduce the mean, standard deviation, skewness, kurtosis, or early-lag autocorrelation of real gaze velocity, whereas the spectrally regularized LSTM-CNN GAN does on the reported metrics.","The result directly addresses the earlier finding that Markov models can outperform GANs on stochastic trajectory data, by showing that a suitably regularized GAN closes that gap."],"supporting_citations":[{"why":"Defines the adversarial generator-discriminator game on which the GAN training is built.","marker":"16"},{"why":"Supplies the non-saturating generator loss used as $L_G$ in the combined objective.","marker":"25"},{"why":"Earlier comparison where Markov models beat GANs on stochastic trajectories; the baseline this paper aims to surpass.","marker":"32"},{"why":"Introduces the spectral-loss perspective from image super-resolution that motivates Fourier-magnitude comparison.","marker":"35"},{"why":"Shows CNN generators consistently fail to reproduce spectral distributions, motivating the spectral regularization term.","marker":"36"},{"why":"Provides the DCGAN convolutional architecture adopted for the CNN generator and discriminator.","marker":"43"},{"why":"Supplies the HMM formalism and forward-backward procedure used for the Markov baseline.","marker":"45"},{"why":"Supplies the Baum-Welch algorithm used to fit the HMM baseline.","marker":"46"}],"fun_headline_variants":["Spectrally regularized GAN outperforms HMM on gaze fidelity","GAN with spectral loss reproduces eye gaze velocity distributions","LSTM-CNN GAN with spectral loss matches real gaze data","Spectral loss sharpens GAN's eye gaze trajectories","Hybrid GAN with spectral loss: high-fidelity gaze modeling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper compares generated data against the same gaze-velocity sequences used to train the GAN, with no described held-out split, so the reported closeness could reflect overfitting rather than an ability to generate new-looking gaze trajectories.","fun_headline_variants_meta":{"raw":{"variants":["Spectrally regularized GAN outperforms HMM on gaze fidelity","GAN with spectral loss reproduces eye gaze velocity distributions","LSTM-CNN GAN with spectral loss matches real gaze data","Spectral loss sharpens GAN's eye gaze trajectories","Hybrid GAN with spectral loss: high-fidelity gaze modeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000616,"raw_usage":{"total_tokens":2927,"prompt_tokens":1076,"completion_tokens":1851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":1765}},"tokens_in":692,"tokens_out":1851,"duration_ms":13396,"temperature":1.0,"reasoning_tokens":1765,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:40:11.080994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the LSTM-CNN GAN on one set of gaze-velocity sequences and compute Jensen–Shannon divergence and autocorrelation on a disjoint set from different participants or recording sessions; if $D_{JS}$ on the held-out data rises to the HMM's level or above, the central fidelity claim is not established.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the non-saturating generator loss used as $L_G$ in the combined objective."},{"cited_title":"R., Yazidi, A","cited_arxiv_id":null,"evidence_quote":"Earlier comparison where Markov models beat GANs on stochastic trajectories; the baseline this paper aims to surpass."}],"review_version":1}