{"id":"82c24a64-4c34-400b-8d3a-9bd28297700c","arxiv_id":"2611.17013","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DiffSSMEye, a diffusion model with bidirectional state-space layers, synthesizes five-second, 1,000 Hz gaze-velocity sequences with the lowest median spatial error and the highest EKYT embedding similarity among compared GAN, VAE, and diffusion baselines.","lead":"A new AI model generates fake eye-tracking recordings at 1,000 Hz that imitate both a person's gaze style and the task being performed, beating earlier synthetic-gaze systems on spatial accuracy. Such data could help researchers test gaze-based systems and explore privacy-conscious sharing, though the paper's main biometric score measures the model against its own training signal.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"EKYT evaluation metric is the same encoder used for conditioning and loss, making the biometric claim a self-consistency check; spatial accuracy claim is independent and supported.","rationale":"The reader's weakest_assumption correctly identifies the circularity of using EKYT for conditioning, loss, and evaluation. The paper's own limitation statement (b240) confirms this. My stress-test agrees that this is the most load-bearing concern because the biometric half of the central claim is not independently validated. The spatial accuracy claim is separate and better supported, though Tables 1 and 2 lack error bars. The concrete test I propose is a straightforward way to settle whether the EKYT similarity scores reflect genuine identity preservation or merely self-consistency. The verdict remains CONDITIONAL because the spatial claim is plausible but the biometric claim requires independent verification before acceptance. I agree with the reader's assessment.","tokens_in":16609,"tokens_out":531,"duration_ms":7903,"concrete_test":"Re-evaluate the 59 held-out subjects using a different, independently trained biometric encoder (e.g., a separate EKYT ensemble trained on a disjoint cohort, or a different architecture like a CNN-based gaze authentication model). If the cosine similarity drops substantially (e.g., below 0.85) or the rank ordering against baselines changes, then the current EKYT-based metric is not a reliable proxy for identity preservation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim of \"biometric feature consistency\" (Abstract, b8) rests on EKYT cosine similarity scores of 0.91–0.95. However, the identical frozen EKYT encoder is used for three distinct roles: (i) conditioning input (Eq. 3), (ii) identity-preservation loss (Eq. 12), and (iii) evaluation metric (Eq. 20). The paper acknowledges this circularity (Section 7, b240: \"biometric assessment is not independent\"), but the headline claim is stated without this qualification. The scores therefore measure how well the generator reproduces its own conditioning signal, not independent biometric fidelity. No alternative encoder, downstream authentication task, or cross-dataset identity check is provided. The spatial accuracy half of the claim (lowest median target-alignment error on HSS and RAN) is independent of this concern and is supported by the reported tables and statistical tests, though Tables 1 and 2 lack error bars or significance tests for the median comparisons.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces DiffSSMEye, a conditional denoising diffusion probabilistic model with bidirectional Structured State Space Sequence (S4) layers for generating five-second gaze-velocity sequences at 1,000 Hz. The model conditions on a coarse gaze reference, user identity (via frozen EKYT embeddings), and task. It combines noise prediction with identity-preservation, multi-resolution spectral, and integrated-drift losses. Subject-disjoint evaluation on GazeBase across seven tasks reports mean real–synthetic EKYT embedding cosine similarities of 0.91–0.95 and the lowest median spatial errors among compared generators for horizontal and random saccade tasks. The paper claims improved spatial accuracy and biometric feature consistency.","tokens_in":16880,"tokens_out":1548,"duration_ms":11499,"significance":"If the spatial accuracy claims hold, DiffSSMEye advances the state of the art in high-frequency gaze synthesis by achieving lower median target-alignment errors than prior generative models (DiffEyeSyn, SP-EyeGAN, VAE) on two saccade tasks. The use of bidirectional S4 layers for temporal modeling in this domain is novel. The paper provides a subject-disjoint evaluation protocol, which is appropriate for assessing generalization to unseen users. The statistical comparison on RAN RMSE (Table 4) with Holm-corrected p-values is a strength. However, the biometric claim is undermined by the circularity of using the same frozen EKYT encoder for conditioning, loss, and evaluation, which the authors acknowledge in Section 7. The spatial accuracy results are independent and supported, but the headline claim of 'biometric feature consistency' is not independently validated.","major_comments":[{"comment":"The biometric claim (headline 'biometric feature consistency') rests entirely on EKYT cosine similarity scores (Table 3). However, the same frozen EKYT encoder serves as (i) the conditioning input (Eq. 3), (ii) the identity-preservation loss (Eq. 12), and (iii) the evaluation metric (Eq. 20). The paper acknowledges this circularity in Section 7 (b240: 'biometric assessment is not independent'), but the abstract and conclusion state the result without this qualification. The scores therefore measure how well the generator reproduces its own conditioning signal, not independent biometric fidelity. No alternative encoder, downstream authentication task, or cross-dataset identity check is provided. This is a load-bearing issue for the biometric half of the central claim. The spatial accuracy claims are independent and not affected.","section":"§5.3, Eq. (20); §3.5.2, Eq. (12); §3.3, Eq. (3)","authors_acknowledge":"Because EKYT is used for conditioning, supervision, and evaluation, biometric assessment is not independent, and the current metrics do not establish downstream authentication utility or complete physiological fidelity.","authors_acknowledge_p":0.94},{"comment":"Tables 1 and 2 report U50|E50 and U95|E95 values without any measure of variability (e.g., confidence intervals, standard deviations, or error bars). The statistical analysis in §5.5 (Table 4) provides bootstrap CIs and p-values for RAN RMSE, but the headline spatial accuracy results (U50|E50 for HSS and RAN) lack such quantification. Given that the improvement over DiffEyeSyn is modest (3.20 vs 3.73 for HSS; 3.22 vs 4.06 for RAN), it is unclear whether these differences are statistically significant. The authors should provide confidence intervals or significance tests for the median comparisons in Tables 1 and 2.","section":"Table 1 and Table 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The statistical analysis uses RMSE over the first 5,000 samples as the metric, which is a different measure from the U50|E50 spatial accuracy reported in Table 1. While the RMSE result supports improved reconstruction accuracy, it does not directly validate the headline spatial accuracy claim (target-alignment error). The authors should clarify how RMSE relates to the U50|E50 metric and whether the significant improvement in RMSE translates to significant improvement in target-alignment error.","section":"§5.5, Table 4","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The sine normalization uses sin(π v^T / (2 v_max)). The superscript T appears to be a transpose, but v is a vector; please clarify the intended operation (element-wise sine of the scaled vector).","section":"§3.2, Eq. (2)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The drift loss uses L1 norm of the cumulative sum of velocity differences. The notation ||·||_1 is clear, but the equation formatting has extra braces and backslashes that should be cleaned.","section":"§3.5.4, Eq. (14)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The direction vector formula uses tan(π x/180) etc. This is the standard spherical projection, but the derivation of the normalization factor should be briefly justified or cited.","section":"§4.3.1, Eq. (16)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The standard deviations for DiffSSMEye (e.g., 0.09, 0.11, 0.12) are substantially larger than those for DiffEyeSyn (e.g., 0.04, 0.05, 0.03). This higher variance should be discussed, as it may indicate less consistent biometric alignment across subjects despite higher mean.","section":"Table 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The ablation study reports EKYT cosine similarity for different conditioning configurations. Since the evaluation metric is the same as the conditioning signal, the ablation results are expected to show improvement when identity conditioning is added. This does not provide independent evidence of biometric fidelity; it only shows that the model uses the conditioning signal. The authors should acknowledge this limitation in the ablation discussion.","section":"§5.6, Table 5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The limitations section is thorough and acknowledges the circularity issue. However, the abstract and conclusion should be revised to qualify the biometric claim accordingly, e.g., 'consistency with the conditioning representation' rather than 'biometric feature consistency'.","section":"§7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper uses 'biometric feature consistency' and 'identity preservation' interchangeably. Since the evaluation is not independent, consider using more precise terminology such as 'embedding consistency' or 'conditioning alignment'.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Hey,\n\nThis paper is a clear step forward for high-frequency gaze synthesis. The spatial accuracy numbers on HSS and RAN are the best I've seen from a learned generator, and the statistical test on RAN RMSE (Holm-adjusted p=0.006 vs DiffEyeSyn) gives me confidence the improvement is real, not just a lucky seed. The architecture is a sensible recombination of published pieces—bidirectional S4 denoiser from SSSD, conditioning from DiffEyeSyn, auxiliary losses from audio—but the configuration for 1,000 Hz gaze is new and the ablation shows identity conditioning is doing the heavy lifting.\n\nNow the soft spot, which the reader and stress-test both caught: the biometric claim. The EKYT encoder is used for conditioning (Eq. 3), identity-preservation loss (Eq. 12), and evaluation (Eq. 20). The paper discloses this in Section 7 (b240: \"biometric assessment is not independent\"), but the abstract and conclusion state \"biometric feature consistency\" without that qualification. The 0.91–0.95 cosine similarities therefore measure how well the generator reproduces its own conditioning signal, not independent biometric fidelity. No alternative encoder, downstream authentication task, or cross-dataset check is provided. This doesn't sink the paper—the spatial half is independent and supported—but it means the headline claim is overstated. The authors are honest about the limitation in the text, but they don't neutralize it.\n\nOther concerns are minor: Tables 1 and 2 lack error bars for the median comparisons (though the RAN RMSE test partially addresses this), the two new loss terms (spectral and drift) are never isolated in the ablation, and code/checkpoints are promised rather than shipped. The EKYT training-cohort overlap with test subjects is also unspecified.\n\nWho this is for: researchers working on gaze synthesis or biometric data augmentation. The spatial results are worth engaging with. The biometric claim needs independent verification before being taken at face value.\n\nRecommendation: I'd send this to peer review. The spatial contribution is solid, the architecture is clearly described, and the limitations are honestly stated (even if the abstract oversells). A good reviewer will push for an independent biometric evaluation and error bars on the main tables, but the core work is worth publishing.","headline":"Solid spatial results, but the biometric claim is a self-consistency check—authors disclose this, but the abstract doesn't.","tokens_in":17489,"tokens_out":566,"duration_ms":7261},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"DiffSSMEye: A conditional diffusion model with structured state-space layers achieves the highest real–synthetic embedding similarity on six of seven eye-movement tasks and the lowest median spatial errors for horizontal and random…","keywords":["eye movement synthesis","gaze synthesis","diffusion model","state-space model","S4","biometric consistency","conditional generation","eye tracking"],"falsifier":"Train an independent biometric gaze classifier (e.g., a ResNet- or Transformer-based model) on real GazeBase data from the 263 training subjects, then evaluate its accuracy on real vs. synthetic sequences from the 59 held-out subjects. If DiffSSMEye's synthetic sequences do not yield significantly higher authentication accuracy than those from DiffEyeSyn or a simple baseline, the claim of improved biometric feature consistency would be undermined. Alternatively, if the independent classifier achieves near-chance accuracy on synthetic sequences, the EKYT cosine similarity results would be exposed as self-consistency rather than true identity preservation.","tokens_in":18010,"feed_emoji":"👁️","tokens_out":1883,"duration_ms":13067,"temperature":0.7,"pith_summary":"The paper introduces DiffSSMEye, a conditional denoising diffusion probabilistic model that generates five-second gaze-velocity sequences at 1,000 Hz. It uses bidirectional Structured State Space Sequence (S4) layers in the denoiser, conditioning on a coarse gaze reference, a frozen biometric embedding from the EKYT encoder, and a learned task embedding. The authors argue that this architecture, combined with auxiliary losses for identity preservation, multi-resolution spectral matching, and integrated-drift penalty, produces synthetic eye movements that are more spatially accurate and more consistent with a subject's biometric features than prior GAN, VAE, or diffusion baselines. If true, this would provide a method for generating high-fidelity synthetic gaze data that could supplement limited recordings, support controlled experiments, and reduce privacy risks associated with sharing real eye-tracking data, while retaining subject-specific oculomotor characteristics.","feed_headline":"Diffusion with state-space layers improves gaze synthesis accuracy and biometric…","feed_subtitle":"DiffSSMEye generates 1000 Hz gaze sequences that better match real subject-specific eye movement patterns than prior GAN, VAE, and…","key_machinery":"The key machinery is the denoiser architecture: 30 residual blocks each containing two bidirectional Structured State Space Sequence (S4) layers. These S4 layers model long-range temporal dependencies through a discretized state-space model that is equivalent to a convolution with sequence-spanning kernels, allowing the network to incorporate context from both past and future timesteps. Conditioning is applied via feature-wise linear modulation (FiLM) using the EKYT identity embedding and the task embedding, and a coarse gaze reference is concatenated with the noisy input. The auxiliary losses (identity-preservation via cosine distance in EKYT space, multi-resolution STFT magnitude loss, and integrated-drift penalty) constrain the output beyond the standard noise-prediction objective.","core_discovery":"The central claim is that DiffSSMEye, by integrating bidirectional S4 layers into a conditional DDPM denoiser and augmenting the noise-prediction loss with identity-preservation, multi-resolution spectral, and integrated-drift objectives, can synthesize five-second, 1000 Hz gaze-velocity sequences that achieve lower median spatial error on horizontal and random saccade tasks and higher real–synthetic embedding cosine similarity on six of seven tasks compared to SP-EyeGAN, a VAE baseline, and DiffEyeSyn. The authors state that these results demonstrate improved spatial accuracy and biometric feature consistency for subject-specific gaze synthesis.","pith_inferences":["Because the EKYT embedding is used for conditioning, supervision, and evaluation, the headline cosine similarity scores (0.91–0.95) partly measure the generator's ability to reproduce the signal that guides it, not an independent measure of identity preservation. An independent biometric encoder or a downstream authentication task would provide a stronger test of identity fidelity.","The spatial accuracy improvements over DiffEyeSyn are modest (0.53 dva for HSS, 0.84 dva for RAN at U50|E50), and at the 95th percentile DiffEyeSyn actually outperforms DiffSSMEye on RAN. This suggests that the main advantage of the S4 architecture may lie in modeling the central tendency of gaze placement rather than eliminating extreme outliers, and that the choice of evaluation percentile can c","The paper evaluates only on GazeBase, which has a specific sampling rate (1000 Hz), tracker (EyeLink 1000), and task set. Generalization to other trackers, sampling rates (e.g., 120 Hz, 2000 Hz), or clinical populations (e.g., dyslexia, ADHD) is untested, and the model's performance may degrade if the coarse-reference construction or EKYT encoder is not well-matched to the new data.","The integrated-drift loss penalizes endpoint displacement over 5-second windows, which may encourage the model to avoid low-frequency drift at the expense of higher-frequency micro-movements that are characteristic of individual oculomotor signatures. A comparison of power spectral density or microsaccade rates between real and synthetic data would clarify whether spectral fidelity is maintained a"],"forward_implications":["If DiffSSMEye's spatial accuracy and biometric consistency hold across diverse datasets, synthetic gaze could serve as a drop-in replacement for real data in training and evaluating gaze-based interfaces, biometric authentication systems, and clinical assessment tools, reducing the need for expensive and time-consuming data collection.","The method's ability to preserve subject-specific oculomotor dynamics at 1000 Hz suggests that synthetic data could be used to augment small or imbalanced datasets for machine learning models, potentially improving their robustness and fairness across demographic groups.","The framework's conditioning on a coarse reference trajectory and task embedding implies that users could generate gaze sequences for novel task conditions or imagined scenarios by supplying a new coarse reference and task label, enabling controlled experiments that are difficult to conduct with human participants.","The explicit modeling of long-range dependencies via S4 layers may generalize to other high-frequency physiological signals (e.g., ECG, EEG, EMG) where preserving subject-specific temporal structure is critical, offering a template for conditional generation in biometry.","The paper's ablation study shows that identity conditioning provides the largest individual gain in embedding similarity, suggesting that biometric encoders are a powerful conditioning signal for subject-specific generation, but the authors acknowledge that the evaluation metric is not independent of the conditioning signal, so the true identity-preservation capability remains to be verified with "],"supporting_citations":[{"why":"Supplies the bidirectional Structured State Space Sequence (S4) layer architecture that forms the core of the denoiser.","marker":"[18]"},{"why":"Supplies the frozen EKYT encoder that provides the identity embedding used for conditioning, identity-preservation loss, and evaluation metric.","marker":"[4]"},{"why":"DiffEyeSyn is the primary diffusion baseline; the paper compares against it and claims improved spatial accuracy and embedding similarity.","marker":"[17]"},{"why":"SP-EyeGAN is a GAN-based baseline for eye movement synthesis; the paper compares against it.","marker":"[15]"},{"why":"VAE baseline for eye movement generation; the paper compares against it.","marker":"[14]"},{"why":"DiffWave architecture provides the residual/skip connection design that the denoiser builds upon.","marker":"[38]"},{"why":"Demonstrates diffusion with structured state-space models for time-series imputation and forecasting, providing the SSSD design pattern that DiffSSMEye adapts.","marker":"[19]"},{"why":"GazeBase is the dataset used for training and evaluation; the paper's results are specific to this dataset.","marker":"[12]"},{"why":"FiLM conditioning layer is used to incorporate identity and task embeddings into the denoiser.","marker":"[21]"}],"fun_headline_variants":["State-space diffusion sharpens synthetic eye movements","Bidirectional S4 layers boost gaze synthesis fidelity","ID-preserving loss yields realistic 1000 Hz gaze data","DiffSSMEye: higher biometric consistency in gaze synthesis","Conditional diffusion model bests GANs for eye movement synthesis"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The biometric consistency claim depends on the assumption that cosine similarity between real and synthetic EKYT embeddings is a valid measure of identity preservation for unseen subjects, but the EKYT encoder is used for conditioning, supervision, and evaluation, so the metric partly measures self-consistency rather than independent biometric fidelity.","fun_headline_variants_meta":{"raw":{"variants":["State-space diffusion sharpens synthetic eye movements","Bidirectional S4 layers boost gaze synthesis fidelity","ID-preserving loss yields realistic 1000 Hz gaze data","DiffSSMEye: higher biometric consistency in gaze synthesis","Conditional diffusion model bests GANs for eye movement synthesis"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":8.2e-05,"raw_usage":{"cost":8.20092e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":8.20092e-05,"upstream_inference_prompt_cost":6.76312e-05,"upstream_inference_completions_cost":1.4378e-05},"total_tokens":1027,"prompt_tokens":948,"completion_tokens":79,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":0}},"tokens_in":948,"tokens_out":79,"cost_source":"openrouter_usage","duration_ms":3448,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":0,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T07:31:32.763517+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Train an independent biometric gaze classifier (e.g., a ResNet- or Transformer-based model) on real GazeBase data from the 263 training subjects, then evaluate its accuracy on real vs. synthetic sequences from the 59 held-out subjects. If DiffSSMEye's synthetic sequences do not yield significantly higher authentication accuracy than those from DiffEyeSyn or a simple baseline, the claim of improved biometric feature consistency would be undermined. Alternatively, if the independent classifier achieves near-chance accuracy on synthetic sequences, the EKYT cosine similarity results would be exposed as self-consistency rather than true identity preservation.","supporting_citations":[{"cited_title":"Efficiently Modeling Long Sequences with Structured State Spaces","cited_arxiv_id":"2111.00396","evidence_quote":"Supplies the bidirectional Structured State Space Sequence (S4) layer architecture that forms the core of the denoiser."},{"cited_title":"Eye know you too: Toward viable end-to-end eye movement biometrics for user authentication.IEEE Transactions on Information Forensics and Security, 17:3151–3164, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen EKYT encoder that provides the identity embedding used for conditioning, identity-preservation loss, and evaluation metric."},{"cited_title":"DiffEyeSyn: Diffusion-based User-specific Eye Movement Synthesis","cited_arxiv_id":"2409.01240","evidence_quote":"DiffEyeSyn is the primary diffusion baseline; the paper compares against it and claims improved spatial accuracy and embedding similarity."},{"cited_title":"Sp-eyegan: Generating synthetic eye movement data with generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"SP-EyeGAN is a GAN-based baseline for eye movement synthesis; the paper compares against it."},{"cited_title":"Fully convolutional neural networks for raw eye tracking data segmentation, generation, and reconstruction","cited_arxiv_id":null,"evidence_quote":"VAE baseline for eye movement generation; the paper compares against it."},{"cited_title":"DiffWave: A Versatile Diffusion Model for Audio Synthesis","cited_arxiv_id":"2009.09761","evidence_quote":"DiffWave architecture provides the residual/skip connection design that the denoiser builds upon."},{"cited_title":"Diffusion-based Time Series Imputation and Forecasting with Structured State Space Models","cited_arxiv_id":"2208.09399","evidence_quote":"Demonstrates diffusion with structured state-space models for time-series imputation and forecasting, providing the SSSD design pattern that DiffSSMEye adapts."},{"cited_title":"Gazebase, a large-scale, multi-stimulus, longitudinal eye movement dataset.Scientific Data, 8(1):184, 2021","cited_arxiv_id":null,"evidence_quote":"GazeBase is the dataset used for training and evaluation; the paper's results are specific to this dataset."},{"cited_title":"Film: Visual reasoning with a general conditioning layer","cited_arxiv_id":null,"evidence_quote":"FiLM conditioning layer is used to incorporate identity and task embeddings into the denoiser."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}