{"id":"244dd88a-3b30-4484-ad95-16c79f51fbc7","arxiv_id":"2605.26434","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Reconstruction-based EEG foundation models preferentially encode aperiodic and low-frequency components over oscillatory structure, with embeddings capturing subject identity more than task-relevant information.","lead":"EEG foundation models using reconstruction pretraining capture aperiodic and low-frequency signal components while under-representing higher-frequency oscillations. This bias explains their limited gains over smaller supervised models in low-data regimes and points to a concrete direction for improving EEG representations.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"Synthetic EEG generation may introduce artifacts that exaggerate oscillatory under-representation, weakening attribution to reconstruction objective","rationale":"The reader's weakest_assumption is precisely the load-bearing point; the abstract provides no further detail on synthetic fidelity, so the concern stands as identified. No stronger internal inconsistency appears from the given text.","tokens_in":1734,"tokens_out":317,"duration_ms":19740,"concrete_test":"In the methods section, extract the exact synthetic generation procedure and parameters; recompute the embedding bias metrics after replacing the synthetic generator with one fitted to real EEG (e.g., multivariate AR(4) model estimated on the same dataset used for pretraining); if the aperiodic/oscillatory disparity shrinks by >15% or reverses sign, the original synthetic setup is the dominant source of the reported bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that synthetic inputs isolate the effect of the reconstruction loss on aperiodic vs. oscillatory components. The abstract states experiments use 'controlled, synthetically-generated EEG inputs' to demonstrate bias toward aperiodic and under-representation of high-frequency oscillations. If the generation procedure (e.g., independent superposition of 1/f noise and sinusoids without matching real EEG non-stationarity, phase-amplitude coupling, or channel correlations) does not reproduce the joint spectral statistics of real data, the observed mismatch could arise from the synthetic construction itself rather than the pretext task. This directly underpins the mechanistic account and the motivation for auxiliary losses.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that reconstruction-based pretraining in EEG foundation models induces a spectral bias favoring high-power aperiodic components and low-frequency content while under-representing low-power high-frequency oscillations. This is shown via controlled synthetic EEG inputs that isolate the effect, plus linear probe experiments on real BCI datasets demonstrating stronger encoding of subject identity than task information; the authors conclude this mismatch explains poor low-resource performance and motivate auxiliary losses targeting oscillatory structure.","tokens_in":1863,"tokens_out":484,"duration_ms":23665,"significance":"If the central attribution holds, the work supplies a concrete mechanistic account of why reconstruction objectives are mismatched to EEG statistics and supplies an actionable path (auxiliary losses) for improving foundation-model pretraining. The controlled synthetic setup, if shown to preserve real EEG joint statistics, would be a strength for causal isolation; the linear-probe results on subject vs. task encoding would further ground the practical relevance.","major_comments":[{"comment":"Synthetic EEG generation subsection (Experiments section): the central claim that the observed aperiodic bias is caused by the reconstruction objective requires that the synthetic inputs reproduce the joint spectral statistics of real EEG (non-stationarity, phase-amplitude coupling, channel correlations). The manuscript must supply the precise generation procedure and quantitative comparisons (e.g., PSD, coherence, PAC metrics) between synthetic and real data; absent these controls, the mismatch could originate in the synthetic construction itself rather than the pretext task.","section":"Experiments / synthetic EEG generation"},{"comment":"Linear probe evaluation paragraph (Results section): the claim that embeddings encode subject identity more strongly than task-relevant information is load-bearing for the practical implication. The manuscript should report the exact probe accuracies, the number of subjects/tasks, cross-validation scheme, and statistical comparison (e.g., paired t-test or effect size) between subject and task probes; without these numbers the strength of the bias remains unquantified.","section":"Results / linear probe evaluation"}],"minor_comments":[{"comment":"Abstract, final sentence: the phrasing 'thereby reinforcing the low-frequency and aperiodic component bias' is circular; reword to separate the empirical observation from the interpretive claim.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"Thank you for the constructive feedback on our manuscript. We address each major comment below and will incorporate revisions to strengthen the work.","responses":[{"response":"We agree that quantitative validation of the synthetic data against real EEG is necessary to attribute the observed bias specifically to the reconstruction objective. In the revised manuscript, we will expand the synthetic EEG generation subsection to provide the precise generation procedure along with direct comparisons of power spectral density (PSD), coherence, and phase-amplitude coupling (PAC) metrics between synthetic and real datasets. This addition will address the concern and reinforce the causal interpretation.","revision_made":"yes","referee_comment":"[Experiments / synthetic EEG generation] Synthetic EEG generation subsection (Experiments section): the central claim that the observed aperiodic bias is caused by the reconstruction objective requires that the synthetic inputs reproduce the joint spectral statistics of real EEG (non-stationarity, phase-amplitude coupling, channel correlations). The manuscript must supply the precise generation procedure and quantitative comparisons (e.g., PSD, coherence, PAC metrics) between synthetic and real data; absent these controls, the mismatch could originate in the synthetic construction itself rather than the pretext task."},{"response":"We acknowledge that detailed quantitative reporting is required to substantiate the relative encoding of subject identity versus task information. In the revision, we will update the linear probe evaluation paragraph to include the exact probe accuracies, the number of subjects and tasks, the cross-validation scheme, and statistical comparisons (e.g., paired t-tests or effect sizes) between the subject and task probes. These additions will quantify the bias and support the practical implications.","revision_made":"yes","referee_comment":"[Results / linear probe evaluation] Linear probe evaluation paragraph (Results section): the claim that embeddings encode subject identity more strongly than task-relevant information is load-bearing for the practical implication. The manuscript should report the exact probe accuracies, the number of subjects/tasks, cross-validation scheme, and statistical comparison (e.g., paired t-test or effect size) between subject and task probes; without these numbers the strength of the bias remains unquantified."}],"tokens_in":1403,"tokens_out":454,"duration_ms":19847,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing here is that reconstruction pretraining on EEG pushes embeddings to over-weight aperiodic and low-frequency content while under-representing higher-frequency oscillations, and that this shows up as stronger subject-identity encoding than task information in linear probes.\n\nWhat is new is the direct link from the reconstruction loss to that spectral mismatch, backed by controlled synthetic inputs plus real BCI probe results. The abstract frames this as a fixable limitation rather than a generic complaint about foundation models, and it points to auxiliary losses for oscillatory structure as a next step. That diagnosis is more specific than most prior work on EEG pretraining failures.\n\nThe experiments sound like a reasonable way to isolate the effect, and the motivation for better low-resource performance is practical. If the synthetic generation actually reproduces the joint statistics of real EEG, the causal attribution to the pretext task would hold up.\n\nThe soft spot is exactly the one in the stress-test note. Without details on how the synthetic signals were built—whether they preserve non-stationarity, phase-amplitude coupling, or channel correlations—the observed bias could partly reflect the input construction rather than the loss. The abstract does not report effect sizes or statistical tests, so it is hard to gauge how large or consistent the mismatch is. That makes the central claim provisional until the methods section is checked.\n\nThis is for people working on EEG foundation models or BCI transfer who need a handle on why current pretraining falls short. It is worth sending to referees because the question is well-posed and the suggested direction is concrete, even if the current evidence is thin on the synthetic controls.","headline":"The paper gives a concrete mechanistic story for why reconstruction-based EEG foundation models underperform in low-resource settings, but the synthetic data controls need more scrutiny to pin the bias on the objective itself.","tokens_in":2390,"tokens_out":410,"would_cite":false,"duration_ms":23841,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Reconstruction-based EEG foundation models capture aperiodic signal components while under-representing high-frequency oscillatory ones.","keywords":["EEG foundation models","spectral bias","aperiodic components","oscillatory components","reconstruction objectives","brain-computer interfaces","embeddings","pretext tasks"],"falsifier":"An experiment that measures the power spectrum of signals reconstructed from the model embeddings on synthetic inputs with isolated high-frequency oscillations and finds equal or stronger representation of those oscillations compared with aperiodic components.","tokens_in":2654,"feed_emoji":"","tokens_out":678,"duration_ms":37503,"temperature":0.7,"pith_summary":"The paper establishes that pretraining EEG foundation models via reconstruction on large unlabeled datasets creates a systematic bias: the learned embeddings favor the high-power aperiodic parts of the signal and under-represent the low-power oscillatory components, especially at higher frequencies. This mismatch arises because reconstruction objectives align more naturally with aperiodic structure than with the oscillatory elements that carry much task-relevant information in EEG. Linear probe tests on real BCI datasets show the embeddings encode subject identity more strongly than task labels, which explains the observed underperformance relative to smaller supervised models in low-resource settings. A reader would care because the bias points to a concrete, addressable limitation in current EEG foundation model design.","feed_headline":"EEG foundation models favor aperiodic over oscillatory signals","feed_subtitle":"Reconstruction pretraining creates a bias that favors low-frequency components and subject identity over task information in BCI settings.","key_machinery":"The reconstruction pretext task, which aligns embeddings with high-power aperiodic EEG components at the expense of low-power oscillatory ones.","core_discovery":"Using controlled synthetic EEG inputs that separate aperiodic and oscillatory components, reconstruction-based EEG foundation model embeddings are shown to preferentially encode aperiodic structure while under-representing oscillatory activity, with the effect strongest at higher frequencies. On real-world BCI datasets, linear probes confirm that these embeddings represent subject identity more strongly than task-relevant features, thereby reinforcing the low-frequency and aperiodic bias induced by the reconstruction objective.","pith_inferences":["The same reconstruction-induced bias could appear in foundation models trained on other biosignals that exhibit strong aperiodic backgrounds.","Pretraining objectives that directly penalize loss of oscillatory power, such as frequency-specific contrastive terms, offer a testable route to more balanced representations.","Subject-identity dominance in embeddings suggests that current models may require explicit disentanglement steps before they can generalize across individuals."],"forward_implications":["Embeddings will show weaker performance on downstream tasks that depend on high-frequency oscillatory content.","Linear probes will continue to recover subject identity more readily than task labels across multiple BCI datasets.","The performance gap versus fully supervised models will remain largest in low-resource regimes where fine-tuning cannot overcome the spectral bias.","Adding explicit auxiliary losses for high-frequency oscillatory structure during pretraining would reduce the mismatch."],"fun_headline_variants":["Reconstruction pretraining biases EEG models to aperiodic signals","EEG foundation models overlook high-frequency oscillatory components","Synthetic EEG reveals aperiodic bias in reconstruction-based embeddings","EEG embeddings encode subject identity over task-relevant features"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Synthetic EEG signals accurately reproduce the spectral decomposition and statistical properties of real EEG without introducing artifacts that exaggerate the observed bias.","fun_headline_variants_meta":{"raw":{"variants":["Reconstruction pretraining biases EEG models to aperiodic signals","EEG foundation models overlook high-frequency oscillatory components","Synthetic EEG reveals aperiodic bias in reconstruction-based embeddings","EEG embeddings encode subject identity over task-relevant features"]},"model":"grok-4.3","cost_usd":0.00504,"raw_usage":{"total_tokens":2374,"prompt_tokens":662,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":50403000,"prompt_tokens_details":{"text_tokens":662,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1651,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":662,"tokens_out":61,"duration_ms":19701,"temperature":1.0,"reasoning_tokens":1651,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T18:55:35.251577+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An experiment that measures the power spectrum of signals reconstructed from the model embeddings on synthetic inputs with isolated high-frequency oscillations and finds equal or stronger representation of those oscillations compared with aperiodic components.","supporting_citations":[],"review_version":1}