{"id":"8d093d97-d9a6-4d3f-894b-b2a2a7e2429d","arxiv_id":"2412.19814","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A transformer predicts the next fMRI brain state and accurately forecasts about 5.04 seconds of future activity from 21.6 seconds of resting-state data, while longer synthetic sequences preserve functional connectivity.","lead":"Researchers trained a transformer on resting-state fMRI data from 1,003 people and found it can predict the next brain state and about five seconds of future activity from the previous 21.6 seconds. The model also generates long synthetic fMRI sequences whose connectivity patterns resemble real brains, suggesting transformers may serve as generative models for brain activity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No persistence or linear AR baseline: low short-horizon MSE may reflect bandpass autocorrelation rather than learned dynamics.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: absence of trivial baselines. The paper's own reported shuffle control (Section 3.2) demonstrates order sensitivity but not predictive advantage over persistence or linear autoregression. Because the preprocessing includes bandpass filtering and z-scoring, adjacent and short-lag brain states are strongly autocorrelated; low MSE over 5 seconds is therefore not, by itself, evidence of learned nonlinear dynamics. This is a correctness-risk issue, not merely a disagreement with consensus, and it is directly testable because the code and data splits are available. The rest of the paper—held-out subject evaluation, ten-fold cross-validation, and the shuffled-order control—gives some independent support, but it does not close the baseline gap. I therefore keep the reader's CONDITIONAL verdict unchanged rather than escalate to REJECT: the concern is serious but addressable, and a positive baseline comparison would settle it. No ad hominem is intended; the critique is limited to the missing control experiment.","tokens_in":7500,"tokens_out":4349,"duration_ms":45735,"concrete_test":"Using the released code and the same ten-fold subject split, evaluate on held-out subjects: (1) persistence forecast x_{t+1}=x_t and, for iterative multi-step prediction, x_{t+k}=x_t for k=1..7; (2) a linear AR model (e.g., ridge-regularized AR(30) per region, fit on the training subjects) with the same iterative rollout. Compare single-step MSE to 0.0013 and the mean MSE over the first seven predicted time points to the reported <0.15. If either baseline achieves statistically comparable or lower MSE (paired test across subjects), the paper's central accuracy claim is not established. If the transformer significantly beats both, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the transformer learns brain dynamics and accurately predicts future brain states is supported mainly by MSE values in Sections 3.2 and 3.3, yet the only temporal control is a shuffled-order condition. The HCP fMRI series were z-scored per region and bandpass-filtered to 0.01-0.1 Hz (Section 2.1), which induces strong autocorrelation at TR=0.72s. With such smoothing, a persistence forecast x_{t+1}=x_t or a linear autoregressive model can plausibly produce single-step MSE near 0.0013 and multi-step MSE below the paper's post hoc threshold of 0.15 for the first seven time points. The shuffle control only shows that temporal order matters to the trained transformer; it does not show that the transformer outperforms simple extrapolation. If a trivial baseline matches or beats the reported errors, the 'accurate prediction up to 5.04s' conclusion reduces to autocorrelation of the BOLD signal rather than learned brain dynamics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a transformer-based autoregressive model for predicting resting-state fMRI brain states. The model takes 30 consecutive time points (21.6 s at TR=0.72 s) of 379 regional BOLD signals from the HCP dataset and predicts the next time point; iterative application generates a synthetic time series of up to 1150 time points. The authors report a single-step MSE of 0.0013 (versus 0.97 under a shuffled-order control), MSE below 0.15 for the first seven multi-step predictions (5.04 s), and similarity between the functional-connectivity matrices computed from true and predicted data (spatial correlation 0.50–0.60). Ten-fold subject-level cross-validation on 1003 HCP subjects is used, the window size and epoch count were chosen in a preliminary experiment, and the code is publicly available.","tokens_in":7659,"tokens_out":10671,"duration_ms":94535,"significance":"If the claims hold, the paper would demonstrate a generative model that produces plausible future brain states from a short fMRI segment, with potential applications in scan-time reduction and in generating synthetic resting-state data. The evaluation has genuine strengths: held-out subjects at the cross-validation level, a shuffled-order control as a sanity check, and public code that supports reproducibility. The Markov-chain interpretation of error accumulation is clearly stated. However, the central claim that the transformer learns brain dynamics rather than exploiting the strong temporal autocorrelation of the BOLD signal is not yet supported, because trivial baselines (persistence, AR(1)) are absent and the 'accurate up to 5.04 s' criterion is defined post hoc. The functional-connectivity result is also not discriminating between temporal dynamics and mere spatial covariance reproduction. These gaps make the significance conditional on the requested experiments.","major_comments":[{"comment":"The central claim that the transformer 'learned and leveraged the temporal dependencies' and can accurately predict brain states up to 5.04 s is not established by the current evaluation, which uses only a shuffled-order control. The bandpass filter (0.01–0.1 Hz) applied at Section 2.1, combined with TR=0.72 s, induces strong temporal autocorrelation in the regional BOLD series, so a persistence forecast (x̂_{t+1}=x_t) or a per-region AR(1) model may plausibly reproduce the reported single-step MSE of 0.0013 and the low multi-step MSE over the first seven time points (Fig. 3b). The shuffled-input MSE of 0.97 is close to 1.0—the MSE of predicting the temporal mean of z-scored data—so the 700-fold difference mainly shows that order carries information, not that the transformer extracts dynamics beyond simple extrapolation. Please add persistence and AR(p) baselines evaluated on the same held-out subjects with the same metrics (MSE, correlation, and FC analysis), and report whether the transformer's errors are significantly lower.","section":"§3.2, §3.3 (with §2.1)"},{"comment":"The paper defines the accuracy window as 'the first seven predicted time points (within the time of 5.04s) have relatively low MSE (< 0.15)'. This threshold is chosen after observing the results (Fig. 3b), and no pre-specified criterion, noise ceiling, or baseline-calibrated value is given. The 'accurate up to 5.04 s' conclusion is therefore post hoc. Please justify the threshold a priori, or calibrate it against the trivial baselines above, and report the distribution (mean ± SD across subjects) of the MSE at each of the first seven horizons rather than only the group average.","section":"§3.3"},{"comment":"The comparison with BrainLM ('their 20-time-point MSE is 0.568, larger than ours') is not controlled: the authors' model uses a 30-time-point input, whereas the BrainLM evaluation cited in the Introduction uses a longer input (180 time points), and the preprocessing, parcellation, and data splits differ. Under these conditions the comparison does not support the claim of superiority. Either re-implement BrainLM on the same preprocessing and folds, or present the comparison only as indicative and state explicitly which factors could explain the difference.","section":"§3.3"},{"comment":"The functional-connectivity analysis does not distinguish learned temporal dynamics from reproduced spatial covariance. The group-average FC pattern in Fig. 4b is determined by the cross-regional covariance of the signals; any generator that draws from a multivariate distribution with the empirical covariance (for instance, a null model that temporally shuffles each region's series independently) would produce a similar FC similarity (r ≈ 0.5–0.6 in Fig. 4e) without capturing sequential structure. Please compare the FC of the transformer output against that of a spatial-covariance null model and of the persistence baseline, and report whether the similarity is significantly higher.","section":"§3.4"}],"minor_comments":[{"comment":"The caption contains grammatical slips (e.g., 'which is the MSE of the 51st time points'); the entire manuscript would benefit from careful copy-editing for typos and phrasing.","section":"Fig. 3 caption"},{"comment":"Model selection in Fig. 2 is based on single runs; no error bars, standard deviations across subjects, or multiple random seeds are reported, so the choice of window size 30 and 20 epochs is not shown to be robust.","section":"Fig. 2"},{"comment":"The decoder input description is ambiguous: 'the last time point of the encoder input combined with the encoder output' should be specified precisely, including tensor shapes and whether the decoder uses causal masking during the iterative autoregressive generation described in Section 2.4.","section":"§2.2"},{"comment":"It is unclear whether the MSE values in Sections 3.2 and 3.3 are averaged over the ten cross-validation folds or produced by a single final model; the manuscript should state this explicitly.","section":"§2.3"},{"comment":"In the text 'with MSE from 0.35 to 0.45 and correlation from 0.50 to 0.60', the quantity labeled MSE is the mean absolute difference between FC matrices shown in Fig. 4e; using 'MSE' here conflates two different error metrics and should be corrected.","section":"§3.4"},{"comment":"The claim that short-segment prediction 'can largely shorten the scanning time' is not directly demonstrated; a sentence acknowledging that the clinical benefit is currently speculative would be more accurate.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript is a clearly written but methodologically incomplete draft. The central claim—that a transformer learns brain dynamics from 30 time points of rs-fMRI—rests on MSE values that could plausibly be replicated by persistence or AR(1) forecasts given the bandpass-filtered BOLD autocorrelation. If those baselines match the transformer, the main conclusion reduces to a statement about autocorrelation; if the transformer wins, the paper would be a solid, publishable contribution. The BrainLM comparison and the FC analysis also need matched or null-model controls. I recommend major revision rather than rejection because the missing experiments are within the scope of the manuscript and the paper is honest about error accumulation. The code availability and subject-level cross-validation are strengths to preserve. The title and abstract currently state accuracy claims that are not yet supported and should be softened or substantiated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful empirical demonstration that a compact transformer can forecast a few seconds of rs-fMRI from a 21.6s window, with held-out-subject evaluation and a sane shuffle control. The headline number (5.04s) is real in the sense that the paper actually computes it, but the paper does not show that the transformer beats trivial extrapolation, so I would not yet take the 'learned brain dynamics' interpretation at face value.\n\nWhat is actually new: while BrainLM and SwiFT use transformers on fMRI for classification or with 180-token pretraining, this paper trains a small time-series transformer on 30 time points and reports both single-step and iterative multi-step MSE, plus the observation that the generated series preserves group-level functional connectivity. The code is on GitHub, which is more than many papers in this space do. The shuffle control is a good sanity check and the paired t-test is appropriate. The ten-fold cross-validation at the subject level is also solid.\n\nThe soft spots are the ones the reader flagged. Most important: there is no persistence or AR(1) baseline anywhere in the paper. The data are bandpass filtered to 0.01–0.1 Hz and z-scored per region; at TR = 0.72 s that produces very high lag-1 autocorrelation. A persistence forecast may well produce single-step MSE in the same ballpark as the reported 0.0013, and may stay under the post hoc MSE < 0.15 criterion for the first 7 steps. The shuffle control only shows that temporal order matters to the trained transformer, not that it outperforms simple extrapolation. The 0.15 threshold is chosen after looking at the curves, and the time-series plots have no error bars across subjects. The FC preservation result is suggestive but also lacks a null model; a persistence baseline would trivially preserve FC, so it does not discriminate.\n\nNone of these are fatal. The paper is what it claims to be: a promising initial result with a simple architecture, not a foundation model. The fix is straightforward: add persistence and AR(1) curves to Figures 2 and 3, pre-register the accuracy criterion, and run a spatial null for the FC result. If the transformer still beats persistence after that, the 5.04 s claim is credible.\n\nFor the right reader — a neuroimaging methodologist thinking about scan-time reduction or generative models for fMRI — this is worth engaging with. I would send it to peer review, because the empirical claim is concrete and the missing baseline is an addressable weakness rather than a fatal flaw. I would not cite it yet in my own work until the baseline comparison is in.","headline":"Compact transformer predicts 5.04s of rs-fMRI, but the missing persistence/AR baseline undercuts the 'learned dynamics' claim for now.","tokens_in":8218,"tokens_out":3755,"would_cite":false,"duration_ms":32914,"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 transformer trained on 30 resting-state fMRI time points can predict the next seven brain states, roughly 5.04 seconds, with MSE below 0.15.","keywords":["fMRI","brain states","transformer","prediction","resting-state fMRI","functional connectivity","autoregressive model","Human Connectome Project"],"falsifier":"Compute a persistence forecast on the same held-out HCP subjects, setting each predicted state equal to the previous true or predicted state, and measure the MSE over the first seven steps; if that baseline reaches or beats the transformer's 0.15, the short-horizon accuracy is explained by signal smoothness rather than by the model's learned dynamics.","tokens_in":7282,"feed_emoji":"🧠","tokens_out":6857,"duration_ms":63235,"temperature":0.7,"pith_summary":"The paper asks whether future brain states, represented by the mean blood-oxygen-level-dependent (BOLD) activity of 379 regions at each fMRI time point, can be predicted from a short history. It trains a transformer on resting-state fMRI from the Human Connectome Project and reports that, given 30 time points (21.6 seconds), the model predicts the next brain state with mean squared error (MSE) 0.0013 and the next seven states (5.04 seconds) with error below 0.15. Shuffling the input order raises the single-step error more than 700-fold, which the authors read as evidence that temporal structure, not static features, drives the prediction. Over longer horizons the error accumulates, yet the fully generated time series still reproduces the group-level functional connectivity of the brain. If these results hold, short fMRI scans could be extended synthetically and generative models of brain dynamics may become practical.","feed_headline":"Transformer predicts 5.04 seconds of future brain states","feed_subtitle":"With 21.6 seconds of resting-state fMRI as input, it forecasts seven states and the brain's connectivity pattern.","key_machinery":"The load-bearing object is a transformer encoder-decoder treated as an autoregressive time-series forecaster. Each token is one brain state, a 379-dimensional vector of mean regional BOLD signals defined by the multi-modal parcellation atlas, and sine/cosine positional encodings inject temporal order. The encoder reads 30 consecutive tokens; the decoder receives the encoder output plus the last true token and emits a single next state, trained with mean squared error. Self-attention is what allows the model to weight relationships among distant past states, and the paper's case that this matters rests on the contrast between low error on ordered sequences and high error on shuffled ones.","core_discovery":"The central claim is that a transformer, trained as an autoregressive model on regional resting-state fMRI, learns enough of the brain's temporal dynamics to predict the immediate future: a single-step MSE of 0.0013, the first seven multi-step predictions below 0.15, and a shuffled-order control error of 0.97 with p < $10^{-10}$. The paper interprets the shuffled control as evidence that sequential order is indispensable; otherwise a model with the same inputs but random order would perform similarly. For long-range synthesis, prediction error grows as predicted states are fed back into the model, but the group-average functional connectivity matrix computed from generated time series correlates with the true group-average matrix at 0.50 to 0.60, indicating that the model captures the population-level organization of functional networks.","pith_inferences":["The paper does not compare against a persistence baseline (x_{t+1} = x_t) or a linear autoregressive model; until such baselines are shown to do worse, the unique contribution of the transformer to short-horizon prediction remains an open question.","Because the model operates on z-scored regional signals, its reported MSE values are relative to unit-variance time series; comparing forecast errors across studies would require reporting errors in raw BOLD units or per-region variances.","The same encoder-decoder design could be tested on task-evoked fMRI and on clinical cohorts, where the forecasting horizon and functional-connectivity fidelity would need to be weighed against scan time in concrete diagnostic tasks."],"forward_implications":["A 21.6-second resting-state fMRI segment may suffice to generate roughly five seconds of plausible future brain activity, shortening the scan time needed for applications that depend on a short predicted window.","Long autoregressive generation, despite accumulating error, preserves the group-average functional connectome, so the model captures shared functional organization across individuals.","The 700-fold jump in error when input frames are shuffled implies temporal ordering is essential to the model's performance, not a side effect of static regional relationships.","Error accumulation limits reliable prediction to about seven time points, so further work on reducing cascading errors is required before longer forecasts could be used in practice."],"supporting_citations":[{"why":"Supplies the HCP young-adult resting-state fMRI data, the 1003 subjects and four 1200-frame scans per subject that the transformer is trained and tested on.","marker":"[28]"},{"why":"Defines the multi-modal parcellation whose 379 grey-matter regions form the coordinate space of each predicted brain-state vector.","marker":"[32]"},{"why":"Motivates the 0.01 to 0.1 Hz bandpass filter applied before prediction, the preprocessing step that shapes the smoothness of the signal the model must forecast.","marker":"[31]"},{"why":"Introduces the self-attention and positional-encoding machinery that the paper adapts for sequential brain-state prediction.","marker":"[18]"},{"why":"Provides the time-series transformer architecture that the authors re-designed into their single-step autoregressive encoder-decoder.","marker":"[27]"},{"why":"Supplies the BrainLM baseline whose 20-time-point MSE of 0.568 is compared against the reported 0.26.","marker":"[26]"}],"fun_headline_variants":["Transformer predicts brain states 5 seconds ahead","AI model forecasts future fMRI brain states","Transformer learns brain dynamics from fMRI data","Predicting brain states with transformer and fMRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the assumption that the low prediction error comes from learned brain dynamics rather than from the temporal smoothness of bandpass-filtered fMRI, and the paper reports no comparison against repeating the last observed state or a linear autoregressive model.","fun_headline_variants_meta":{"raw":{"variants":["Transformer predicts brain states 5 seconds ahead","AI model forecasts future fMRI brain states","Transformer learns brain dynamics from fMRI data","Predicting brain states with transformer and fMRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1280,"prompt_tokens":935,"completion_tokens":345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":292}},"tokens_in":551,"tokens_out":345,"duration_ms":4248,"temperature":1.0,"reasoning_tokens":292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:19:11.650192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute a persistence forecast on the same held-out HCP subjects, setting each predicted state equal to the previous true or predicted state, and measure the MSE over the first seven steps; if that baseline reaches or beats the transformer's 0.15, the short-horizon accuracy is explained by signal smoothness rather than by the model's learned dynamics.","supporting_citations":[{"cited_title":"Neuroimage, 2013","cited_arxiv_id":null,"evidence_quote":"Supplies the HCP young-adult resting-state fMRI data, the 1003 subjects and four 1200-frame scans per subject that the transformer is trained and tested on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the multi-modal parcellation whose 379 grey-matter regions form the coordinate space of each predicted brain-state vector."},{"cited_title":"Scientific reports, 2017","cited_arxiv_id":null,"evidence_quote":"Motivates the 0.01 to 0.1 Hz bandpass filter applied before prediction, the preprocessing step that shapes the smoothness of the signal the model must forecast."},{"cited_title":"bioRxiv, 2023: p","cited_arxiv_id":null,"evidence_quote":"Supplies the BrainLM baseline whose 20-time-point MSE of 0.568 is compared against the reported 0.26."}],"review_version":1}