{"id":"2d5d452c-54bc-40ae-bcd2-364832a829b0","arxiv_id":"2501.01212","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"MM-PTGNN uses cross-modal alignment to transfer personalized cybersickness traits from non-invasive sensors into a video encoder, enabling video-only inference at 88.4% reported accuracy.","lead":"This paper trains a VR sickness predictor that needs only the video feed at run time, after learning each user's head, eye, and physiological signals during training. It claims 88.4% accuracy, near EEG-based systems, at about 90ms latency, which would make real-time consumer-grade cybersickness prediction possible.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 5's Full Model accuracy (73.1%) contradicts the headline 88.4% under the paper's own 'identical settings' claim, so the central video-only accuracy is unsupported.","rationale":"The central claim is specific and falsifiable: video-only inference at 88.4% Top-1, near EEG-based 89.16%, with real-time latency. For this claim to hold, the paper's own reported results must be internally consistent. They are not. Section 4.4.1 Table 3 gives Ours Video Top-1 as 88.4%, while Section 4.5 Table 5 reports Full Model (ours) Top-1 as 73.1% under 'identical settings' and 'the same video-only inference setup.' These cannot both describe the same configuration unless some undisclosed condition differs, such as different folds, window size, or label granularity. The text provides no such distinction, and the abstract and headline cite only 88.4%. This is not a matter of consensus or framing; it is an internal contradiction in the evidence for the main claim. Even setting that contradiction aside, the evaluation protocol is under-specified. Section 4.2 says the metrics are reported on a held-out test set averaged over 5-fold cross-validation, but it never states whether the folds are subject-wise or window-wise. With 27 participants and overlapping sliding windows, random window splits create near-duplicate training and test windows and permit identity memorization. The reported sensor accuracy of 84.24% and video-only accuracy of 88.4% could then reflect user identification rather than cybersickness prediction. The paper also reports no error bars or per-subject results, and the stated code repository is not released, so neither number can be independently checked. The feature-alignment evidence in Table 2 is partly circular, since the alignment loss directly minimizes MSE between the same embedding pair being evaluated. A corrected version with consistent tables, explicit subject-wise splits, error bars, and released artifacts could make the central claim credible. As submitted, the central claim is unsupported. I therefore maintain the Reader's REJECT without adjustment.","tokens_in":13458,"tokens_out":2789,"duration_ms":27159,"concrete_test":"Obtain from the authors, or from released code and dataset, the exact split definition and per-fold results for the Full Model. Re-run the protocol described in Section 4.5 with subject-wise folds (all windows of a participant in one fold) and report Top-1 on held-out subjects. If the resulting accuracy is 73.1% or lower, the 88.4% headline is unsupported; if it is 88.4%, then Table 5's full-model row must be corrected. A secondary check: train a trivial identity-only classifier under the same splits; if it exceeds chance, the split is leaking identity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—video-only inference at 88.4%, closely matching EEG's 89.16%—is directly contradicted by the paper's own ablation. Table 3 lists 'Ours Video' at 88.4% Top-1; Section 4.5 states that all ablations are trained and evaluated under identical settings using the same video-only inference setup, yet Table 5 reports Full Model (ours) at 73.1% Top-1 and Macro-F1 68.2%. If Table 5 is accurate, the headline should be near 73%, not 88.4%; if Table 3 is accurate, Table 5's full-model row is wrong. The paper offers no reconciliation, no confidence intervals, and no released code or dataset to adjudicate. Additionally, Section 4.2 reports a 'held-out test set averaged over 5-fold cross-validation' without stating whether folds are subject-wise or window-wise. With 27 participants and overlapping sliding windows, random window splits would let the model memorize identity, inflating personalization and inflating any accuracy figure. Both issues are load-bearing: the headline number is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MM-PTGNN, a framework for personalized cybersickness prediction that trains a video encoder to imitate sensor-derived embeddings (head motion, eye tracking, physiological signals) via a cross-modal alignment loss, then performs video-only inference at deployability. The headline claims are 88.4% Top-1 accuracy matching EEG-based methods (89.16%), with roughly 90 ms inference latency. The manuscript reports experiments on a 27-participant dataset with 5-fold cross-validation, and includes ablation studies, feature-similarity diagnostics, and latency measurements. The core technical idea—training a video-only model by aligning with non-invasive multimodal signals—is plausible and potentially useful, but the reported results contain a direct internal contradiction in the central accuracy number, and the evaluation protocol is underspecified in a way that could invalidate the personalization claim.","tokens_in":13640,"tokens_out":3380,"duration_ms":33534,"significance":"If the claims were reliably established, the framework would be practically significant: it would offer a wearable-free, real-time cybersickness predictor for consumer VR, with accuracy comparable to EEG-based approaches. The paper targets a relevant problem, and the cross-modal alignment strategy is a reasonable route. However, the central numerical claim is not currently supported: the paper's own ablation table reports a Full Model accuracy of 73.1% for the same video-only inference setup that the abstract and Table 3 report as 88.4%, with no reconciliation. In addition, the cross-validation protocol does not state whether folds are subject-wise or window-wise, which is load-bearing for the personalization claim. No code or data are provided to adjudicate the contradiction. These issues are central rather than peripheral.","major_comments":[{"comment":"The central accuracy claim is contradicted by the paper's own ablation results. The abstract and Table 3 report 88.4% Top-1 accuracy for 'Ours Video', while Table 5 reports 'Full Model (ours)' at 73.1% Top-1 and 68.2% Macro-F1, and §4.5 states that all models are trained and evaluated under identical settings using the same video-only inference setup. These two sets of numbers cannot both be correct. The manuscript offers no explanation for the discrepancy, no confidence intervals, and no error bars. As written, the headline claim of 88.4% accuracy is unsupported; either the headline or the ablation must be wrong.","section":"§4.5, Table 5 vs. Abstract and Table 3"},{"comment":"The paper reports 'held-out test set averaged over 5-fold cross-validation' but never states whether the splits are subject-wise or window-wise. With only 27 participants and sliding-window time series, overlapping windows from the same participant are highly correlated; a window-wise random split would allow the model to memorize user identity and thereby inflate both the accuracy and the personalization claims. The authors must specify the split strategy and, in particular, confirm that all windows of a given participant are confined to either the training or test fold. Without this detail, the reported 88.4% accuracy cannot be interpreted as evidence of cross-user personalization.","section":"§4.2, Evaluation Metrics and 5-fold cross-validation"},{"comment":"The alignment-quality evidence in Table 2 is circular as evidence of personalization. The cosine similarity and MSE between z_v and z_p are direct measurements of the objective minimized by L_align in Eq. (15); high similarity between the video embedding and the sensor embedding is expected by construction when that loss is optimized. The randomized baseline is a useful sanity check, but the metric itself cannot independently demonstrate that the video encoder has learned personalized traits. The subsequent accuracy comparison against external baselines partially grounds the claim, but the feature-level table should not be presented as independent validation.","section":"§4.3.2, Table 2 and Eq. (15)"}],"minor_comments":[{"comment":"The abstract states an average inference latency of 90 ms, but Table 4 reports 93.6 ms for the proposed video-only model; these should be reconciled.","section":"Abstract and Table 4"},{"comment":"There is a typo in the abstract: 'relesed' should be 'released'.","section":"Abstract"},{"comment":"The text lists 'HR, BMP, EDA, and SKT' while the earlier introduction and Table 1 refer to BVP (blood volume pulse); the abbreviation should be consistent.","section":"§4.1 and Table 1"},{"comment":"The sentence 'Our model can achieve as fast as Kundo' is inconsistent with Table 4, which lists Kundo at 90 ms and the proposed model at 93.6 ms; the wording should be corrected to avoid implying the proposed model is faster.","section":"§4.4"},{"comment":"The notation H'_t in Eq. (4) is not defined; the equation uses E_t in one line and H'_t in the next, and the relationship between these variables should be clarified.","section":"§3.1.2, Eq. (4)"}],"recommendation":"reject","confidential_remarks":"The internal contradiction between Table 3 and Table 5 is severe enough that the central result cannot be accepted without access to the code and data, and the missing split-strategy specification compounds the concern. I would be open to reconsidering a revised manuscript that resolves the accuracy discrepancy, specifies subject-wise cross-validation, reports error bars, and releases the code. However, as submitted, the headline claim is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth taking seriously: instead of needing EEG or worn sensors at inference, train a video encoder with a cross-modal alignment loss against non-invasive signals so that video alone can predict individual cybersickness. That is a real step toward consumer-grade adaptive VR comfort, and the difference-attention module is a reasonable touch. The paper is also honest about the data source and reports latency. So it is not a crackpot paper; it is a plausible research direction with a serious execution problem.\n\nThe execution problem is the 88.4% vs 73.1% discrepancy. Table 3 lists \"Ours Video\" at 88.4% Top-1, while Table 5, which is described as the same video-only inference setup, reports the \"Full Model (ours)\" at 73.1% Top-1 and 68.2% Macro-F1. That is not a minor footnote; it is the paper's central claim. The authors never reconcile it, and there are no confidence intervals or significance tests to help a reader adjudicate. Add to that the undisclosed split strategy: with 27 participants and overlapping sliding windows, a window-wise split would let the model memorize user identity, which would inflate the personalization numbers in the alignment table and the accuracy itself. The alignment table also has a circularity problem, since cosine similarity and MSE between the two embeddings are literally what the alignment loss minimizes; high similarity just tells you the loss worked, not that the encoded traits are genuinely personalized.\n\nOn the positive side, the architecture is clearly described, the related work is relevant, and the runtime claim (93.6 ms per sample in Table 4) is plausible for video-only inference. The problem is not the idea; it is that the evidence as presented cannot separate real success from identity leakage or a typo.\n\nWho is this for? Researchers working on cybersickness prediction or adaptive VR who want to know that video-only transfer is plausible. They should read it, but they should not rely on the numbers until the authors run a proper subject-wise split, reconcile Tables 3 and 5, and release code and data. I would not cite it in its current form.\n\nFor peer review: this deserves referee time, not a desk reject, because the underlying contribution is useful and the flaws appear fixable. But the referee report should push hard for a consistent numerical story and a split that prevents leakage. If the authors can produce that, the paper becomes credible.","headline":"A genuinely useful video-only cybersickness idea, but the headline accuracy is contradicted by the paper's own ablation; requires fixes and released artifacts.","tokens_in":14210,"tokens_out":1626,"would_cite":false,"duration_ms":19643,"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 video encoder trained to imitate non-invasive sensor embeddings can predict VR cybersickness from video alone at 88.4%, matching EEG-based predictors.","keywords":["cybersickness prediction","virtual reality","video-only inference","cross-modal alignment","graph neural network","difference attention","personalized comfort","non-invasive sensors"],"falsifier":"Run the same experiment with subject-wise cross-validation, where all clips from each participant stay in either training or test, and compare top-1 accuracy; if accuracy drops sharply under this split, the video-only personalization claim is an artifact of identity leakage rather than true generalization.","tokens_in":13224,"feed_emoji":"🕶️","tokens_out":10803,"duration_ms":93501,"temperature":0.7,"pith_summary":"This paper tries to establish that personalized cybersickness prediction for virtual reality does not need EEG or other wearables at inference time. The authors propose training a video encoder jointly with a multimodal graph network over non-invasive signals—head motion, eye tracking, EDA, blood volume pulse, and skin temperature—so the video encoder learns user-specific patterns through a cross-modal alignment loss. At run time only the video stream is used, and the paper reports 88.4% top-1 accuracy, close to the EEG-based comparison of 89.16%, with an average inference latency of about 90ms. If true, consumer VR headsets could adapt comfort settings to each user in real time without extra hardware.","feed_headline":"Video alone predicts VR sickness at 88.4% accuracy","feed_subtitle":"Video-only model matches EEG accuracy after training on head, eye, and physiological signals","key_machinery":"The load-bearing mechanism is cross-modal alignment: an MSE loss $L_{\\text{align}} = \\frac{1}{N}\\sum_i \\|z_v^{(i)} - z_p^{(i)}\\|^2$ pulls a video embedding $z_v$ toward the personalized embedding $z_p$ computed from non-invasive sensor signals. The sensor embedding comes from a modality-specific graph convolutional network whose nodes are features within head, eye, and physiological modalities, followed by a Difference Attention Module that computes attention from local central-difference signal changes $\\Delta X_t = X_t - \\frac{1}{2k+1}\\sum_{\\tau=t-k}^{t+k} X_\\tau$, so dynamic changes drive the fusion and baseline drift is downweighted. This machinery turns the deployment problem into a representation-transfer problem: the video encoder inherits per-user traits during joint training, and at inference the sensors are no longer needed.","core_discovery":"The central claim is that a video-only model can match EEG-based cybersickness prediction by transferring personalization from non-invasive sensors during training. The model, MM-PTGNN, builds per-modality graphs over head, eye, and physiological time series, fuses them with a Difference Attention Module that emphasizes transient changes rather than baselines, and then forces the video embedding to approximate the sensor-derived embedding with a mean squared error alignment loss. After training, the sensor branch is discarded and prediction runs on VR video alone, reaching 88.4% accuracy versus 89.16% for the EEG baseline while keeping inference around 90ms per sample. The paper interprets the gap of less than one point as evidence that non-invasive signals carry the personalized information needed for consumer-grade deployment.","pith_inferences":["It is an open question whether the video-only model generalizes to VR content and demographics far from the 27-participant dataset; the alignment may encode stable user traits or may memorize dataset-specific video-sensor correlations.","Because alignment transfers sensor-derived traits into video features, the same recipe could distill other expensive measurements, such as physiological biomarkers, into camera-only models for health and engagement monitoring.","The difference-attention idea—attend to local signal changes rather than raw values—could transfer to other wearable time-series tasks, such as stress detection or seizure monitoring, where baseline drift is a known problem.","A direct extension would be few-shot personalization: train the alignment on a large multi-user corpus, then adapt the video encoder with a few minutes of each new user's headset signals without retraining the full model."],"forward_implications":["Consumer VR headsets could run personalized cybersickness prediction using only the rendered video stream, with no EEG caps or skin-contact sensors.","The reported 88.4% video-only accuracy is within about one point of the EEG-based 89.16% baseline, so the main barrier to comfort prediction becomes software rather than hardware.","At roughly 90ms per sample on a single GPU, the model is fast enough to support real-time adaptive responses, such as adjusting field of view or content difficulty mid-session.","A 42.3MB model size makes the approach plausible for edge devices, not just server-side inference."],"supporting_citations":[{"why":"Provides the EEG-based 89.16% accuracy baseline that the paper's video-only model is designed to match without EEG hardware.","marker":"[17]"},{"why":"Supplies the dataset of synchronized VR video, head/eye motion, and physiological signals, and serves as the lightweight real-time comparison.","marker":"[20]"},{"why":"Serves as a video-plus-sensor deep learning baseline for motion sickness prediction.","marker":"[22]"},{"why":"Demonstrates that integrated HMD eye- and head-tracking sensors can support cybersickness prediction, motivating the non-invasive sensor set.","marker":"[9]"},{"why":"Offers a video-only sickness estimation approach whose high latency is compared against the proposed real-time model.","marker":"[25]"},{"why":"Provides a sensor-only physiological baseline for cybersickness detection.","marker":"[24]"},{"why":"Is the closest video-only optical-flow baseline, at 86.2% accuracy.","marker":"[16]"},{"why":"Is the closed-loop sensor-plus-video framework used for inference latency comparison.","marker":"[8]"},{"why":"Is the EEG-plus-sensor transformer baseline whose latency and accuracy contextualize the proposed method.","marker":"[12]"}],"fun_headline_variants":["Video-only model predicts VR sickness at 88.4% accuracy","No EEG needed: video predicts VR dizziness at 88.4% accuracy","Video rivals EEG for VR sickness prediction: 88.4% without sensors","Real-time video-only cybersickness forecast: 88.4% match to EEG","Drop the sensors: video alone predicts VR sickness at 88.4%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the test participants are people the model has never been trained on; if the same person's video clips appear in both training and testing, the reported personalization accuracy could be inflated by the model memorizing that person.","fun_headline_variants_meta":{"raw":{"variants":["Video-only model predicts VR sickness at 88.4% accuracy","No EEG needed: video predicts VR dizziness at 88.4% accuracy","Video rivals EEG for VR sickness prediction: 88.4% without sensors","Real-time video-only cybersickness forecast: 88.4% match to EEG","Drop the sensors: video alone predicts VR sickness at 88.4%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000945,"raw_usage":{"total_tokens":4024,"prompt_tokens":920,"completion_tokens":3104,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":3003}},"tokens_in":536,"tokens_out":3104,"duration_ms":20824,"temperature":1.0,"reasoning_tokens":3003,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:32:15.556578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same experiment with subject-wise cross-validation, where all clips from each participant stay in either training or test, and compare top-1 accuracy; if accuracy drops sharply under this split, the video-only personalization claim is an artifact of identity leakage rather than true generalization.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EEG-based 89.16% accuracy baseline that the paper's video-only model is designed to match without EEG hardware."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dataset of synchronized VR video, head/eye motion, and physiological signals, and serves as the lightweight real-time comparison."},{"cited_title":"In 2023 IEEE International Symposium on Mixed and Augmented Reality (ISMAR)","cited_arxiv_id":null,"evidence_quote":"Offers a video-only sickness estimation approach whose high latency is compared against the proposed real-time model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the closest video-only optical-flow baseline, at 86.2% accuracy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the EEG-plus-sensor transformer baseline whose latency and accuracy contextualize the proposed method."}],"review_version":1}