{"id":"02e8348f-dd5d-41c1-a803-07162fbd2a7d","arxiv_id":"2412.08223","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"In VR, task load, red light, and slow music bias users toward underestimating elapsed time, and relative underestimation correlates with better user experience.","lead":"This VR study tests how light color, music tempo, and task load change people's sense of time, and trains a neural network to guess time perception from EEG and heart signals. It finds that tasks, red light, and slow music make people underestimate time, and that underestimating time tracks with better reported experience.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RQ3's labels are deterministic threshold functions of the input trial length, so TPM-Net's 83.33% accuracy may reflect duration/padding rather than physiological content.","rationale":"The reader's weakest assumption identifies the same load-bearing gap. The concern is not that the effect sizes are small or the statistics non-standard; it is structural: the input contains the label-generating quantity. The RSTC/UX correlation is a separate empirical claim and may survive, but RQ3's modeling claim and the associated accuracy numbers are uninterpretable without a duration control. Because the control is straightforward, the current CONDITIONAL verdict is appropriate; no change is needed, but the revision should require the duration-control experiment rather than cosmetic edits.","tokens_in":19720,"tokens_out":5274,"duration_ms":61410,"concrete_test":"Train a three-class classifier on the single feature 'pre-padding sequence length' (number of valid frames) using the same 85/15 split and labels from Section 3.3; report its test accuracy. Since labels are threshold functions of t, a near-perfect score would prove the network can solve RQ3 without any physiological content, confirming the duration confound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TPM-Net models time perception from multimodal physiological data is not supported because the classification target is a deterministic function of the input duration. Section 3.3 defines labels as underestimated for t >= 69, acceptably accurate for 51 <= t < 69, and overestimated for t < 51, where t is the participant's self-terminated trial duration. The same section states that 'the length of each trial varies depending on the duration taken by participants,' and the preprocessing stage pads sequences within each batch. The physiological and Zei sequences therefore carry the trial length either in their number of unpadded frames or in the padding pattern, and the label is a threshold function of that length. A model that ignores all signal values and counts non-padding frames can reproduce the labels almost exactly, with errors only near the 51/69 boundaries. The Table 5 comparison against NB, SVM, CNN, and BiLSTM does not control for this, so it tests the architecture's ability to exploit duration, not whether EEG, HR, or human-state content is informative. The abstract's 86.11% versus Table 5's 83.33% adds a reporting inconsistency, but the missing duration control is the decisive gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a between-subject VR experiment (N = 56) that varies three zeitgebers—light color (red/blue/white), music tempo (140 BPM/70 BPM/none), and task factor (spatial rotation task vs. none)—under a prospective 60-s time-estimation task. The authors propose the Relative Subjective Time Change (RSTC) method, which compares a participant's mean subjective time estimate in an experimental condition with their own baseline, to analyze how relative time perception relates to questionnaire-based user experience. They also propose TPM-Net, a CNN-Transformer classifier that takes multimodal physiological time series (EEG, brainwave components, heart rate, human-state signals, and zeitgeber flags) and classifies each trial as underestimated, acceptably accurate, or overestimated. The main reported findings are that the task factor significantly increases underestimation, red light and slow-tempo music further bias toward relative underestimation, relative underestimation strongly correlates with overall user experience, presence, and engagement, and TPM-Net outperforms classical and deep baselines with roughly 83–86% accuracy.","tokens_in":20022,"tokens_out":7552,"duration_ms":67534,"significance":"The strength of the paper is its systematic experimental setup and the RSTC baseline-relative analysis, which offers a practical way to link subjective time distortion to user experience without requiring accurate absolute time judgments; the reported RQ2 correlations are substantial (e.g., r = 0.65 for relative time perception vs. relative overall user experience). However, the RQ3 modeling claim is currently not supported because the classification label is a deterministic threshold function of the trial's actual duration, and the physiological input sequences span exactly that duration, so the network can solve the task from sequence length alone. In addition, the reported accuracy in text differs from Table 5, and several ANOVA statistics in Figure 6 are internally inconsistent. If these issues are resolved, the paper would constitute a meaningful contribution to VR time-perception research; at present the empirical conclusions for RQ1 and RQ2 are plausible but the headline RQ3 result requires substantially more evidence.","major_comments":[{"comment":"The labels in the Classification stage are deterministic threshold functions of the trial duration t (underestimated for t ≥ 69 s, acceptably accurate for 51 s ≤ t < 69 s, overestimated for t < 51 s), and the input sequences are physiological and Zei recordings covering exactly that same trial, padded within each batch to a common length. A model that only counts non-padding frames (or otherwise attends to sequence length) can recover the label with errors only within ±3 s of the two boundaries. Table 5 does not include such a length-only baseline, nor any control that removes duration from the input (e.g., fixed-length cropping, using duration as a separate covariate, or randomizing the mapping between duration and labels). Therefore the reported 83.33% accuracy in Table 5 does not establish that EEG, HR, or human-state content carries information about time perception. Please add a duration-only control and report whether TPM-Net still exceeds it; this is essential for the RQ3 conclusion.","section":"3.3, 4.3"},{"comment":"The abstract and §4.3 state that TPM-Net achieved Accuracy = 86.11%, Macro-F1 = 83.10%, UAR = 79.45%, but Table 5 lists Accuracy = 83.33%, Macro-F1 = 74.84%, UAR = 71.32% for the same model. These are different results and must be reconciled; the discrepancy is too large to be a rounding artifact and affects the paper's headline claim. Please state which result corresponds to the reported 85%/15% split and which to the 5-fold cross-validation, and report both consistently.","section":"Abstract; §4.3; Table 5"},{"comment":"Several reported F statistics are inconsistent with their accompanying p-values. For example, with degrees of freedom (2, 39), an F value of 1.43 cannot yield p = 0.025 (the critical F at α = 0.05 is about 3.24), and F(2, 39) = 2.34 cannot yield p = 0.046. These appear in the panels for the music-without-task and color-without-task groups, where the claimed significant effects of music tempo and light color on subjective time estimation rest. Please correct the values or the statistics; if the reported p-values are wrong, the corresponding conclusions about which zeitgebers significantly affect time perception may change.","section":"§4.1, Figure 6"},{"comment":"The ±15% tolerance interval that defines the three classes is chosen after inspecting the collected data ('STD = 10.25' and 'typical variation observed'), so the class balance (243/194/49) and the attainable classification accuracy are partly produced by a threshold fitted to the same dataset. Please report the classification results for at least two alternative margins (e.g., ±10% and ±20%) and show that the relative ranking of TPM-Net against the baselines is stable. Without this, the reader cannot tell whether the reported accuracy is an artifact of an overly generous or specifically tuned label boundary.","section":"§3.3, §4.3"}],"minor_comments":[{"comment":"'Signal prepossess' in the Figure 4 caption (and in the stage description in §3.3) should read 'signal preprocessing'.","section":"Figure 4"},{"comment":"The evaluation protocol is described as an 85%/15% split with a 20% validation portion of training data, and also as 5-fold cross-validation; please state which result in Table 5 corresponds to each protocol, since the two procedures are not equivalent.","section":"§4.3"},{"comment":"The text refers to 'Groups 1 through 4' but Table 2 numbers the experiment groups 0–4; please harmonize the numbering.","section":"§3.2.3"},{"comment":"The exclusion of one participant due to cybersickness is described in §4.4, while §3.2.1 reports four exclusions due to protocol/technical issues; please clarify whether the cybersickness participant is included among those four.","section":"§4.4"},{"comment":"References [25] and [85] are patent documents cited with generic titles and no patent numbers; providing the full patent identifiers would allow readers to verify the human-state signal definitions.","section":"References [25], [85]"}],"recommendation":"major_revision","confidential_remarks":"The RQ3 duration-confound issue is the decisive technical concern; it is fixable in principle with a length-only baseline and fixed-length or duration-controlled inputs. The F/p inconsistencies in Figure 6 also require a careful reanalysis of RQ1. I would not reject the manuscript outright, but the authors must address both before the claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things: the empirical RQ1/RQ2 story is worth a look, and the RQ3 modeling claim is not supported as written. The paper reports a VR study with 56 participants, three zeitgeber classes, and a new baseline-relative method (RSTC). The finding that relative time underestimation tracks overall UX, presence, and engagement (r≈0.65) is a genuinely useful data point, and the RSTC approach is a sensible way to sidestep individual differences in absolute time estimation. The RQ1 results—task factor matters, red light and slow music push toward underestimation—are consistent with attention/arousal theory and are a reasonable extension of prior work. The self-citation to their own earlier VR paper is not a problem; these claims stand on their own.\n\nThe soft spot is RQ3, and it is load-bearing. Labels are defined as t≥69, 51≤t<69, t<51, where t is the participant's self-terminated trial duration. The input physiological sequences span exactly that same duration and are padded within each batch. A model that ignores signal content and counts non-padded frames can reproduce the labels almost exactly. So TPM-Net's reported accuracy (whether 86.11% or 83.33%—the text and Table 5 disagree) does not demonstrate that EEG, HR, or human-state content carries information about time perception. The paper needs a duration-controlled analysis: a length-only baseline, or a model trained on fixed-length windows, or at minimum a feature that removes sequence length. That is the decisive gap.\n\nSmaller issues: the RQ1 analysis uses t-tests on data they themselves report as non-normal, and the between/within participant comparisons across groups are a bit tangled. Those are fixable with non-parametric tests or clearer mixed-model reporting. The 86.11 vs 83.33 discrepancy is sloppy and should be reconciled.\n\nWho is this for? VR/HCI people working on time perception and UX. The empirical part is a legitimate extension, and the RSTC method may be worth citing. The modeling part currently overclaims. The paper deserves a serious referee because the empirical contribution is solid enough and the RQ3 flaw is identifiable and correctable—not a desk-reject. A careful revision could turn this into a useful paper.","headline":"RQ1/RQ2 empirical results are plausible and useful, but RQ3's deep-learning claim is compromised by a trial-duration confound and a numeric inconsistency.","tokens_in":20486,"tokens_out":2277,"would_cite":true,"duration_ms":24917,"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":"Underestimating time in VR tracks better user experience, and a brain-signal model can predict the distortion.","keywords":["virtual reality","time perception","zeitgebers","user experience","EEG","Transformer","physiological signals","Relative Subjective Time Change"],"falsifier":"Remove all duration information from TPM-Net's input—truncate or pad every trial to a fixed length, or feed only the first N seconds—and retrain; if accuracy does not fall far below the reported 83-86%, physiological content is doing the work, and if it collapses, the model is mainly classifying trial length.","tokens_in":19554,"feed_emoji":"🥽","tokens_out":10268,"duration_ms":88188,"temperature":0.7,"pith_summary":"The paper argues that how long VR users think a fixed 60-second interval lasted can be systematically shifted by controllable environment cues—task demand, light color, and music tempo—and that the direction of the shift is a usable signal about how the user experienced the scene. It introduces the Relative Subjective Time Change (RSTC) method, which compares each condition against the user's own baseline instead of against physical time, and finds that relatively underestimating time is strongly correlated with higher overall user experience, presence, and engagement. It then claims that a hybrid CNN-Transformer network called TPM-Net can classify time perception as underestimation, acceptable accuracy, or overestimation from multimodal physiological data (EEG, brainwaves, human states, heart rate) plus zeitgeber flags, with reported accuracy near 86% in the text and 83% in the comparison table. If these claims hold, designers could infer user experience from physiological signals without interrupting the experience, and time perception could serve as a proxy in VR therapy and training.","feed_headline":"Time distortion in VR predicts better experience; AI model reads EEG","feed_subtitle":"Underestimating a 60-second VR interval matched better presence and engagement; a model read it from brain signals.","key_machinery":"The two load-bearing devices are the RSTC method and TPM-Net. RSTC replaces raw subjective estimates with the difference $\\Delta\\hat{t} = \\hat{t}_{\\text{condition}} - \\hat{t}_{\\text{baseline}}$ between a condition and that participant's own baseline, so correlations with user-experience deltas remove stable individual differences in time judgment. TPM-Net is a five-stage pipeline: multimodal input (raw EEG; brainwave bands $\\alpha$, low $\\beta$, high $\\beta$, $\\theta$, $\\gamma$; human states stress, awareness, drowsiness, meditation; heart rate; and a 4-dimensional zeitgeber flag), preprocessing with low-pass filtering and empirical mode decomposition, per-channel 1D CNN temporal encoding, feature fusion into a 128-dimensional embedding, a Transformer for long-range dependencies, and an MLP classifier over three labels (underestimated $t\\ge 69$ s, acceptably accurate $51\\le t<69$ s, overestimated $t<51$ s). The experiment itself uses an event-triggered prospective design: participants are told to estimate 60 s and press a button when they judge the time has elapsed.","core_discovery":"On its own terms, the study establishes that in a 60-second prospective timing task inside VR, participants who performed a spatial rotation task pressed the stop button later than those who did not (mean 83.29 s versus 64.86 s, $p<0.01$), meaning they underestimated the elapsed time; red light and 70 BPM music pushed estimates further toward underestimation relative to baseline, while blue light and 140 BPM music moved them toward overestimation. Using the RSTC baseline-relative measures, the paper reports that the more a participant relatively underestimated time, the better they rated overall user experience, presence, and engagement ($r=0.65$, $r=0.55$, $r=0.52$, all $p<0.001$). It further claims that TPM-Net, a CNN temporal encoder feeding a Transformer over fused EEG, brainwave, human-state, heart-rate, and zeitgeber embeddings, classifies the three time-perception categories with accuracy 86.11%, macro-F1 83.10%, and UAR 79.45% in the text, outperforming naive Bayes, SVM, CNN, and BiLSTM baselines in the table.","pith_inferences":["If the accuracy survives removal of trial-duration cues, the same model could run in real time, adapting a VR scene when physiological signals indicate the user is losing track of time.","The RSTC baseline-differencing logic could transfer to other subjective measures, making relative self-reports comparable across people who use different absolute scales.","Because task presence differed between participant groups, a within-subject replication would test whether the reported task effect on time perception is causal rather than coincidental.","A direct extension would pair music tempo with task rhythm or task-object color to see whether task-linked zeitgebers, which the paper finds stronger, can be tuned independently of the task itself."],"forward_implications":["VR designers should expect time to compress: 71.35% of participants underestimated a 60-second interval overall, rising to 86.42% in the task group.","Comparing each user against their own baseline roughly doubles the explained variance in overall user experience (R2 from 0.191 to 0.424), so relative time change is a stronger UX signal than raw time estimation.","Relative underestimation is a proxy for presence and engagement: larger relative underestimation accompanies higher presence and engagement scores, while emotion and cognitive load show no significant correlation.","TPM-Net can coarsely infer time-perception category from EEG, brainwaves, human states, and heart rate with reported accuracy around 83-86%, suggesting passive physiological monitoring of time perception is possible."],"supporting_citations":[{"why":"supplies the presence, engagement, immersion, and emotion questionnaire dimensions the study adapts for in-VR user experience scores.","marker":"[68]"},{"why":"supplies the event-triggered prospective timing paradigm in which users press a button when they judge the target duration has passed.","marker":"[69]"},{"why":"supports the claim that most people cannot accurately estimate physical time, motivating the RSTC baseline-relative approach and the ±15 percent accuracy tolerance.","marker":"[73]"},{"why":"provides the ipRGC/blue-light mechanism the paper uses to explain why red versus blue light shifts time perception.","marker":"[80]"},{"why":"provides the neural-entrainment account of music tempo effects on time perception.","marker":"[53]"},{"why":"links engagement and time perception in VR, supporting the task-factor condition.","marker":"[49]"},{"why":"provides the 'time flies when having fun' evidence the paper extends to VR user experience.","marker":"[56]"},{"why":"supplies the convolution-augmented transformer design for multimodal physiological signals that TPM-Net adapts.","marker":"[79]"},{"why":"provides the cubic-spline resampling procedure and the BiLSTM baseline compared in the modeling results.","marker":"[86]"},{"why":"supplies the empirical mode decomposition preprocessing applied to raw EEG before temporal encoding.","marker":"[83]"}],"fun_headline_variants":["VR time distortion signals better UX; AI model decodes brain waves","Task, red light, slow tempo distort VR time; Transformer predicts it","Underestimating VR time boosts engagement; AI model reads EEG","VR time perception modeled by Transformer from EEG and context"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption enters in Section 3.3: the classification label is defined from the actual trial duration, and the input sequence spans exactly that same duration, so the network may learn from sequence length or padding rather than from EEG/HR patterns.","fun_headline_variants_meta":{"raw":{"variants":["VR time distortion signals better UX; AI model decodes brain waves","Task, red light, slow tempo distort VR time; Transformer predicts it","Underestimating VR time boosts engagement; AI model reads EEG","VR time perception modeled by Transformer from EEG and context"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3274,"prompt_tokens":1045,"completion_tokens":2229,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":2169}},"tokens_in":661,"tokens_out":2229,"duration_ms":16661,"temperature":1.0,"reasoning_tokens":2169,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:04:14.818584+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove all duration information from TPM-Net's input—truncate or pad every trial to a fixed length, or feed only the first N seconds—and retrain; if accuracy does not fall far below the reported 83-86%, physiological content is doing the work, and if it collapses, the model is mainly classifying trial length.","supporting_citations":[{"cited_title":"Tcha-Tokey, E","cited_arxiv_id":null,"evidence_quote":"supplies the presence, engagement, immersion, and emotion questionnaire dimensions the study adapts for in-VR user experience scores."},{"cited_title":"Unruh, D","cited_arxiv_id":null,"evidence_quote":"supplies the event-triggered prospective timing paradigm in which users press a button when they judge the target duration has passed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supports the claim that most people cannot accurately estimate physical time, motivating the RSTC baseline-relative approach and the ±15 percent accuracy tolerance."},{"cited_title":"Yang, S.-i","cited_arxiv_id":null,"evidence_quote":"provides the ipRGC/blue-light mechanism the paper uses to explain why red versus blue light shifts time perception."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the neural-entrainment account of music tempo effects on time perception."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"links engagement and time perception in VR, supporting the task-factor condition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the 'time flies when having fun' evidence the paper extends to VR user experience."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the convolution-augmented transformer design for multimodal physiological signals that TPM-Net adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the cubic-spline resampling procedure and the BiLSTM baseline compared in the modeling results."},{"cited_title":"Zhang, X.-p","cited_arxiv_id":null,"evidence_quote":"supplies the empirical mode decomposition preprocessing applied to raw EEG before temporal encoding."}],"review_version":1}