{"id":"d9680c1a-bc5c-4fd1-8047-2763af511b30","arxiv_id":"2507.20254","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An EEG foundation model pretrained exclusively on motor-imagery data, using masked-reconstruction plus classification, reports state-of-the-art few-shot decoding on five MI benchmarks.","lead":"A new EEG model family, MIRepNet, is pretrained only on motor-imagery brain data and then fine-tuned on a new user's short recording (under 30 trials per class) to decode imagined movements. The paper reports that this paradigm-specific approach beats both task-specific models and general-purpose EEG foundation models on five public datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MIRepNet's SOTA claim is confounded: baselines are evaluated on raw montages while MIRepNet uses a denser interpolated template plus EA, so the 3–8 point margins may reflect preprocessing rather than the learned foundation model.","rationale":"The reader's weakest_assumption focuses on the inverse-distance interpolation's validity for low-density montages (Section III.B.3, Eqs. 2–4). I agree that interpolation is a critical and under-specified step, but the more load-bearing issue is the asymmetry between MIRepNet and the baselines: MIRepNet receives a denser interpolated template plus Euclidean alignment, while the baselines are evaluated on their native channel layouts. This confound directly undermines the comparative SOTA claim, independent of whether the interpolation itself smears signals. The paper deserves credit for including ablations (w/o Pre-training and w/o Self-supervised, Table VIII), which show that pretraining contributes beyond the pipeline; those ablations keep the pipeline fixed and thereby isolate the pretraining component. However, they do not isolate the channel-template and EA contributions against the baselines, so the external comparison remains uncontrolled. A reviewer can resolve this by re-running baselines on MIRepNet's preprocessed inputs, as proposed in the concrete test. The concern is not about author integrity; it is about the absence of a controlled comparison, and it reinforces the reader's CONDITIONAL verdict rather than changing it. The missing code/weights and unspecified template enumeration are verification gaps, but the preprocessing confound is the single most load-bearing technical issue for the paper's central empirical claim.","tokens_in":23906,"tokens_out":4663,"duration_ms":59163,"concrete_test":"Re-run the strongest baselines (e.g., FBCNet, Conformer, and one generalist such as BIOT) on exactly the same preprocessed inputs as MIRepNet: after the 8–30 Hz filter, resampling to 250 Hz, inverse-distance interpolation to the exact template electrode set used for MIRepNet, and Euclidean alignment computed on the target subject's 30% training trials. Use the same 30% fine-tuning protocol, the same number of seeds, and the same data splits. If any baseline's accuracy rises to within ~1 point of MIRepNet's reported average, the SOTA advantage is attributable to preprocessing rather than the learned representation; if baselines remain 3+ points lower, the foundation model carries the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MIRepNet 'significantly outperforms' specialists and generalists rests on a comparison that is not apples-to-apples. MIRepNet is evaluated after its full pipeline: 8–30 Hz filtering, resampling, inverse-distance interpolation to a fixed channel template (Section III.B.3, Eqs. 2–4), and Euclidean alignment (Section III.B.4). Specialist baselines are described as '30% Trained' from scratch, presumably on each dataset's native channel montage; for BNCI2014004 that is only 3 bipolar channels (C3, Cz, C4). Generalist baselines are fine-tuned with 80% of the target session, but the text does not indicate that they receive the same template interpolation or EA. Thus MIRepNet's input is a dense, whitened montage (e.g., potentially 20+ template sites reconstructed from 3 channels), while baselines see raw, low-density signals. The reported 3–8 point margins could therefore arise from the preprocessing itself—especially since no ablation removes the template or feeds identical interpolated inputs to the baselines. The template's electrode set is never enumerated, so the interpolation's effect cannot be assessed. If template expansion or EA alone explains most of the gain, the 'foundation model' contribution is overstated and the few-shot SOTA claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MIRepNet, described as the first EEG foundation model tailored specifically to the motor imagery (MI) paradigm. The method consists of a preprocessing pipeline (8-30 Hz band-pass filtering, resampling to 250 Hz, subject screening, inverse-distance interpolation to a fixed FC/C/CP/T channel template, and Euclidean alignment) and a hybrid pretraining scheme combining masked token reconstruction with supervised MI classification. The authors pretrain on seven MI datasets and evaluate on five downstream datasets (47 subjects), comparing with nine specialist models and five generalist EEG foundation models. They report consistent improvements, e.g., 81.77% vs. 73.67% on BNCI2014001, with only 30% of a single subject's trials used for fine-tuning.","tokens_in":24030,"tokens_out":8079,"duration_ms":96116,"significance":"If the reported results withstand scrutiny, a paradigm-specific MI foundation model requiring fewer than 30 trials per class for a new subject would be a practically valuable contribution to BCI calibration. The paper has notable strengths: the evaluation covers five downstream datasets from MOABB, uses standard public baselines, reports standard deviations across three runs, and includes an ablation of pretraining and self-supervision. The authors also state that code will be released. However, the central claim is currently conditional because MIRepNet is inseparable from its preprocessing pipeline, and the baselines are not demonstrably given the same inputs. The significance therefore depends on the outcome of a controlled comparison that isolates the learned representation from the preprocessing.","major_comments":[{"comment":"The SOTA claim is confounded by an asymmetric preprocessing comparison. MIRepNet receives band-pass filtering, inverse-distance interpolation to the channel template, and Euclidean alignment before being fed to the model (Eqs. 1–6), while the specialist baselines in Tables II–VII are described only as '30% Trained' and the generalist baselines as '30%/80% Finetuned.' Section IV.B states only that all signals were resampled to 250 Hz; it does not state that baselines received the same template interpolation or EA. On BNCI2014004, which has only C3, Cz, and C4, MIRepNet reconstructs the full template while specialist baselines see the raw three-channel montage. The reported 3–8 point margins can therefore be explained by the preprocessing itself. The authors should rerun all baselines on identical template+EA inputs and add MIRepNet variants without template interpolation and without EA.","section":"§IV.C, Tables II–VII"},{"comment":"The channel template is never enumerated. The text defines 'a fixed set of C electrodes' over FC/C/CP/T regions but does not give the electrode names, their coordinates φ(t_i), or the value of C. This makes it impossible to check the inverse-distance interpolation, especially for low-density recordings such as the three-channel BNCI2014004, where most template sites lie far from any recorded electrode. The authors should specify the full template and, if possible, report reconstruction error or a sensitivity analysis for sparse montages.","section":"§III.B.3, Eqs. (2)–(4)"},{"comment":"The ablation study does not isolate the preprocessing components. It compares only 'w/o Pre-training' and 'w/o Self-supervised' variants against the full MIRepNet. Since the headline claim is that the learned foundation-model representation outperforms specialists, the absence of ablations that remove the channel template or EA means the contribution of those preprocessing steps to the reported gains is unknown. Add ablations with raw montage inputs and without EA to substantiate the attribution.","section":"Table VIII"},{"comment":"The dataset description is internally inconsistent. Section V.B.8 states that BNCI2014001 includes two sessions of 288 trials per subject, which would be 5,184 trials for 9 subjects, but Table I reports 1,296 total trials. The relationship between BNCI2014001 and BNCI2014001-4, and the exact 30% calibration split, are not defined. In addition, the abstract's 'fewer than 30 trials per class' claim appears not to hold for BNCI2015001, where 30% of 200 trials per subject with two classes is exactly 30 trials per class. Clarify the trial counts and the calibration protocol.","section":"Table I and §V.B.8"},{"comment":"The manuscript claims 'the first EEG foundation model tailored for the MI paradigm' but does not discuss the authors' own CLEAN-MI pipeline (ref. [13]), which appears to overlap with the preprocessing ideas in Section III.B (subject selection, channel template). The paper should state explicitly which components are inherited from CLEAN-MI, which are new to MIRepNet, and how the novelty claim is consistent with that prior work.","section":"§I and Ref. [13]"}],"minor_comments":[{"comment":"Equation (4) has an index inconsistency: the left-hand side uses X'[b,i,t], the sum runs over c, but the weight is written as W_ij instead of W_i,c (or W_i,k). Please correct the notation.","section":"Eq. (4)"},{"comment":"The symbol C is used both for the number of classes in the problem definition (Section III.A) and for the number of template electrodes (Section III.B.3). Use distinct symbols to avoid ambiguity.","section":"Section III.A and III.B.3"},{"comment":"The model is referred to as 'RepMI' in several places (e.g., 'A key benefit of RepMI' and 'this fast convergence demonstrates RepMI's practicality') although the paper's model is named MIRepNet. Unify the name throughout.","section":"Section V.D"},{"comment":"BNCI2014002 is listed with 14 subjects in Table I but Section V.B.1 says 13 participants. Please verify the correct count.","section":"Table I and Section V.B.1"},{"comment":"Figure 5's caption says the mask-ratio sensitivity is shown on 'three downstream MI datasets,' but Tables IX–XIII report results on five downstream datasets. Update the caption.","section":"Fig. 5 and Tables IX–XIII"},{"comment":"The 'significantly outperforming' claim in the abstract is not supported by any statistical test. Given that standard deviations are reported, paired significance tests or confidence intervals across subjects should be added for the main comparisons.","section":"Section IV.B"},{"comment":"'datatset' is a typo for 'dataset'.","section":"Section V.B.3"}],"recommendation":"major_revision","confidential_remarks":"The stressed concern about the preprocessing confound is well-founded and should be treated as the primary required revision: the authors need to run baselines on the same interpolated/aligned inputs and ablate the template and EA components. I also recommend asking the authors to clarify the relationship to their own CLEAN-MI preprint (ref. [13]), since the novelty claim 'first MI foundation model' cannot currently be assessed without that discussion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does something genuinely worth looking at—pretraining an MI-specific EEG model on seven datasets and fine-tuning on under 30 trials per class of a new subject, with reported gains of 3–8 points over strong baselines across five held-out datasets. The ablation shows pretraining and the self-supervised loss each help, and the mask-ratio analysis suggests the method is not hyper-sensitive. If the results hold, this is directly useful for low-calibration BCI applications. So credit where earned: the evaluation is broad (47 downstream subjects), internally consistent, and the hybrid pretraining combination is not something I have seen reported before.\n\nThe soft spot is the comparison. MIRepNet gets the full pipeline: 8–30 Hz filtering, resampling, inverse-distance interpolation to a channel template, and Euclidean alignment. The specialist and generalist baselines are described as trained or fine-tuned on the target session, but nothing says they receive the same preprocessing. For BNCI2014004, which has three bipolar channels, interpolating to a dense FC/C/CP/T template could plausibly account for a large chunk of the margin all by itself. There is no ablation that feeds identical interpolated inputs to the baselines or removes the template from MIRepNet. That is a load-bearing omission, and the reader's stress-test lands.\n\nOther issues are real but less severe. The channel template electrode set is never enumerated, so the interpolation cannot be checked. The subject-screening threshold is unspecified, which matters because selecting only good subjects inflates apparent pretraining quality. The text also has an internal inconsistency on whether the generalist 80%-finetuned baseline used only the target subject's session or other subjects' data—that needs to be clarified. The 'first MI foundation model' claim is weakened by the authors' own CLEAN-MI paper and by the table rows labeled 'MI-FM' and Section V.D calling the model 'RepMI'; either the naming history should be explained or the claim softened. No significance tests back 'significantly outperforming', and the three-seed variances (0.10–0.42) look smaller than one would expect, though that alone is not damning.\n\nBottom line: this deserves a serious referee, but not acceptance as-is. I would send it out, ask for code and pretrained weights, an enumerated template, a controlled preprocessing comparison or an ablation isolating template/EA, and a cleaned-up narrative about what is actually new relative to CLEAN-MI/RepMI. If those come back, this could be a useful paper. If the authors cannot or will not release the code and clarify the baseline protocol, the central claim stays unverified.","headline":"Potentially useful few-shot MI decoding results, but the main comparison is confounded by unequal preprocessing and the 'first' claim needs cleaning up before this is citable.","tokens_in":800,"tokens_out":815,"would_cite":false,"duration_ms":29368,"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 motor-imagery-specific EEG foundation model, fed through a fixed motor-cortex channel template and pretrained with masked reconstruction plus classification, can decode a new user from fewer than 30 trials per class and beat both…","keywords":["motor imagery","EEG foundation model","brain-computer interface","channel template","masked token reconstruction","few-shot calibration","Euclidean alignment","cross-dataset transfer"],"falsifier":"Take a high-density recording, hold out the exact template channels, interpolate them from the remaining channels with inverse-distance weighting, and compare the interpolated to the recorded signals; if the correlation is low or the reconstructed channels are spatially smeared, the cross-headset transfer mechanism is not doing what the paper claims. A direct experiment on BNCI2014004 comparing MIRepNet with and without template interpolation would also separate preprocessing gains from learned-representation gains.","tokens_in":23487,"feed_emoji":"🧠","tokens_out":6964,"duration_ms":72673,"temperature":0.7,"pith_summary":"The paper argues that EEG foundation models should be built for one brain-computer-interface paradigm at a time, rather than trained on many paradigms at once. It introduces MIRepNet, a motor-imagery (MI) foundation model whose preprocessing maps any headset's electrodes onto a fixed template over motor-cortex regions, and whose pretraining combines masked token reconstruction with supervised MI classification. The central claim is that after pretraining, a new subject's imagined movements can be decoded accurately from fewer than 30 trials per class and about 10 fine-tuning epochs, across five public datasets and headsets with 3 to 22 channels. If true, it would mean that short calibration sessions suffice to deploy MI-driven devices such as stroke-rehabilitation exoskeletons, and that paradigm-specific foundation models outperform both paradigm-agnostic models and specialist models trained from scratch.","feed_headline":"Motor-imagery EEG model decodes from under 30 trials","feed_subtitle":"MI-only pretraining beats specialist and generalist baselines on 47 subjects across five datasets","key_machinery":"The central object is the neurophysiologically informed channel template: a fixed list of electrode sites over the frontal-central, central, centro-parietal, and temporal scalp regions, reconstructed from any headset by inverse-distance weighting. It carries cross-headset transfer by encoding the MI-specific prior that discriminative sensorimotor rhythms live over C3, C4, and nearby sites. The second mechanism is the hybrid pretraining objective, L = L_rec + L_cls, which jointly minimizes masked temporal-spatial token reconstruction error and cross-entropy MI classification loss. The template standardizes arbitrary electrode layouts, and the hybrid objective supplies the few-shot adaptation that the paper demonstrates.","core_discovery":"The central discovery is that aligning heterogeneous MI recordings to a neurophysiologically informed channel template, then pretraining with a hybrid objective, yields representations that transfer to new subjects with very little data. MIRepNet's pipeline first bandpass-filters 8-30 Hz, resamples to 250 Hz, maps each headset's channels onto a fixed FC/C/CP/T template by inverse-distance weighting, and whitens each subject's trials by Euclidean alignment. A Transformer encoder then processes temporal-spatial tokens; 50% are masked and reconstructed while the same encoder drives a supervised MI classifier. On BNCI2014001, BNCI2015001, BNCI2014004, AlexMI, and BNCI2014001-4, covering 47 subjects, MIRepNet fine-tuned on 30% of one session reports average accuracies of 81.77, 81.67, 82.36, 69.79, and 64.14 percent, above both specialist models trained on the same 30% and generalist models fine-tuned on 80% of a session.","pith_inferences":["A direct test the paper does not run: replace the learned Transformer with a shallow classifier on the same template-plus-Euclidean-alignment features; if accuracy holds, the gains come from preprocessing, not representation learning.","The subject-screening step keeps only subjects whose within-subject classifier passes a threshold, which may bias pretraining toward easy-to-decode users and could overstate few-shot gains for the harder users assistive BCIs most need.","The same logic suggests building separate foundation models for SSVEP, ERP, and sleep staging, each with its own cortical priors, instead of one mixed-paradigm model.","Because the template sites are never enumerated, auditing whether 3-channel recordings are actually reconstructed from nearby electrodes is necessary before clinical deployment; if they are not, low-density headsets may need a different alignment."],"forward_implications":["If MIRepNet is right, a stroke-rehabilitation BCI can be calibrated from a short session of under 30 trials per class instead of requiring large amounts of the new user's data.","The same pretrained weights transfer across headsets with different electrode layouts, including a 3-channel cap, as long as the channel-template interpolation covers the motor region.","The hybrid objective is the reason for the gain: ablations show that removing pretraining or removing masked reconstruction lowers accuracy on all five downstream tasks.","Mask ratio is not a sensitive hyperparameter; 10% to 90% masking all land within roughly two points of the best average, so the model does not need per-dataset tuning.","Paradigm-specific pretraining is preferable whenever the deployment paradigm is known in advance, which the paper argues is the typical real-world setting."],"supporting_citations":[{"why":"Supplies the neurophysiological basis of alpha/beta sensorimotor rhythms and ERD/ERS that motivates the 8-30 Hz filter and the FC/C/CP/T channel template.","marker":"[17]"},{"why":"Supplies the Euclidean alignment whitening method used to match second-order statistics across subjects in the pipeline.","marker":"[18]"},{"why":"EEGNet is one of the specialist baselines that MIRepNet's few-shot decoding must beat.","marker":"[21]"},{"why":"BIOT is one of the generalist EEG foundation model baselines, compared under both 30% and 80% tuning.","marker":"[7]"},{"why":"LaBraM is the masked-modeling generalist baseline whose paradigm-mixed pretraining MIRepNet argues is inferior.","marker":"[10]"},{"why":"EEGPT is the generalist baseline whose universal-embedding claim MIRepNet directly contests.","marker":"[11]"},{"why":"Supplies the standardized benchmarking interface and dataset access used for the five downstream evaluations.","marker":"[31]"},{"why":"Provides the BCI Competition IV benchmark data (BNCI2014001 and BNCI2014001-4) used as downstream tasks.","marker":"[35]"},{"why":"Provides the BNCI2014004 three-channel dataset that tests the channel template under extreme electrode sparsity.","marker":"[37]"}],"fun_headline_variants":["Motor-imagery EEG model needs under 30 trials","First motor-imagery EEG foundation model beats baselines","MIRepNet: EEG foundation model decodes MI with 30 trials","Under 30 trials, MIRepNet tops specialist and general EEG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that a fixed set of scalp positions over the motor cortex can be reconstructed by distance-weighted interpolation from whatever electrodes a headset actually has, even a three-channel headset, and the text never lists those positions or checks the reconstruction error.","fun_headline_variants_meta":{"raw":{"variants":["Motor-imagery EEG model needs under 30 trials","First motor-imagery EEG foundation model beats baselines","MIRepNet: EEG foundation model decodes MI with 30 trials","Under 30 trials, MIRepNet tops specialist and general EEG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1552,"prompt_tokens":974,"completion_tokens":578,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":504}},"tokens_in":590,"tokens_out":578,"duration_ms":7292,"temperature":1.0,"reasoning_tokens":504,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:43:53.445908+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a high-density recording, hold out the exact template channels, interpolate them from the remaining channels with inverse-distance weighting, and compare the interpolated to the recorded signals; if the correlation is low or the reconstructed channels are spatially smeared, the cross-headset transfer mechanism is not doing what the paper claims. A direct experiment on BNCI2014004 comparing MIRepNet with and without template interpolation would also separate preprocessing gains from learned-representation gains.","supporting_citations":[{"cited_title":"Motor imagery and EEG-based control of spelling devices and neuroprosthe- ses,","cited_arxiv_id":null,"evidence_quote":"Supplies the neurophysiological basis of alpha/beta sensorimotor rhythms and ERD/ERS that motivates the 8-30 Hz filter and the FC/C/CP/T channel template."},{"cited_title":"Transfer learning for brain-computer interfaces: A euclidean space data alignment approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the Euclidean alignment whitening method used to match second-order statistics across subjects in the pipeline."},{"cited_title":"EEGNet: a compact convolutional neural network for EEG–based brain–computer interfaces,","cited_arxiv_id":null,"evidence_quote":"EEGNet is one of the specialist baselines that MIRepNet's few-shot decoding must beat."},{"cited_title":"BIOT: Biosignal transformer for cross-data learning in the wild,","cited_arxiv_id":null,"evidence_quote":"BIOT is one of the generalist EEG foundation model baselines, compared under both 30% and 80% tuning."},{"cited_title":"Large brain model for learning generic representations with tremendous EEG data in BCI,","cited_arxiv_id":null,"evidence_quote":"LaBraM is the masked-modeling generalist baseline whose paradigm-mixed pretraining MIRepNet argues is inferior."},{"cited_title":"EEGPT: Pretrained transformer for universal and reliable representation of EEG signals,","cited_arxiv_id":null,"evidence_quote":"EEGPT is the generalist baseline whose universal-embedding claim MIRepNet directly contests."},{"cited_title":"MOABB: trustworthy algorithm bench- marking for BCIs,","cited_arxiv_id":null,"evidence_quote":"Supplies the standardized benchmarking interface and dataset access used for the five downstream evaluations."},{"cited_title":"Review of the BCI competition IV,","cited_arxiv_id":null,"evidence_quote":"Provides the BCI Competition IV benchmark data (BNCI2014001 and BNCI2014001-4) used as downstream tasks."},{"cited_title":"Brain–computer communication: motivation, aim, and impact of exploring a virtual apartment,","cited_arxiv_id":null,"evidence_quote":"Provides the BNCI2014004 three-channel dataset that tests the channel template under extreme electrode sparsity."}],"review_version":1}