{"id":"749617a7-1a42-4df3-bf10-d8ae2bd68ba1","arxiv_id":"2509.00314","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CoMET, a contrastive-masked EEG transformer pre-trained on 1.9 million samples, achieves state-of-the-art linear-probe results on ten BCI and clinical EEG datasets.","lead":"CoMET is a 151-million-parameter self-supervised EEG model trained on more than a million samples from over 3,000 subjects, combining masked reconstruction with a contrastive mirror-scale augmentation. On ten downstream BCI and clinical EEG datasets, it reports state-of-the-art balanced accuracies, with gains up to about 7 percentage points over prior EEG foundation models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mirror-scale momentum views may be temporally mismatched: after 2x upsampling, the N-patch sliding window covers only half the trial, so InfoNCE aligns full-trial MEM views with half-trial disjoint-channel views; this unvalidated assumption underlies the SOTA claim.","rationale":"I read the paper's central claim as: CoMET-Large's frozen encoder yields SOTA linear-probe representations on ten downstream datasets, with the mechanism being the combination of masked reconstruction and a contrastive loss over mirror-scale augmented views. The strongest empirical result is the 62.75% vs 55.85% on BCIC IV 2A. The load-bearing premise is that the InfoNCE positives are semantically valid views of the same brain state. The reader identified this premise as the weakest assumption; I agree, and I find a more specific, checkable technical problem within it: as written, the momentum branch's sliding window covers only half the trial duration after upsampling (N patches at 2f cover T/2 seconds), while the MEM branch sees the full T-second trial. This is not a minor implementation detail because the contrastive loss is supposed to align global brain-state representations; if the momentum view is temporally truncated, the model may instead learn invariance to temporal position or exploit interpolation artifacts. The paper's ablation does not isolate this because it changes the augmentation family rather than the window duration. I do not see evidence of fraud or intentional misreporting, and the broad empirical pattern (including Tiny/Base/Large monotonic gains and several datasets where even CoMET-Tiny beats baselines) suggests the overall finding is not an artifact of one table row. However, the mechanism claimed in the title and abstract is not validated until the temporal alignment of the mirror-scale views is pinned down. Secondary issues also support a conditional stance: no code/weights are released, appendix tables 6-10 are mislabeled (e.g., Table 6 labeled BCIC IV 2B contains PhysioP300 values), and the scaling-law fits in Section 5.2 disagree between the text and the Figure 8 caption. None of these by itself refutes the central claim, so I would keep the reader's CONDITIONAL verdict rather than upgrade to ACCEPT or downgrade to REJECT.","tokens_in":20992,"tokens_out":11449,"duration_ms":152487,"concrete_test":"Re-pretrain CoMET-Base under the same settings but use a sliding window of length 2N (full T=4s) in the momentum branch, while retaining the complement-channel split and 2x upsampling; evaluate the frozen encoder by linear probing on BCIC IV 2A and BCIC2020-3 with the Table 4 protocol. Compare balanced accuracy to CoMET-Base in Table 1 (61.66% and 54.43%). If the 2N-window variant is not at least as good, the temporal mismatch is load-bearing for the current result rather than a harmless implementation detail. If it is clearly better, the published mirror-scale view is temporally misaligned and the SOTA claim should be re-derived under the corrected view. Additionally, if code/weights are released, verify the window length actually used in the released implementation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim rests on the InfoNCE branch (Eq. 14) teaching global discriminative features by aligning MEM and momentum views. In the Mirror-Scale Augmentation section (Eqs. 10-11), after upsampling to 2f the number of temporal patches doubles from N to 2N, then 'a sliding window of length N with stride 1' is applied. With patch length l fixed, N upsampled patches cover N*(l/2f) = T/2 seconds, whereas the MEM view covers T seconds (Eq. 9). The momentum 'positive' is therefore a half-length, channel-disjoint view, not the same brain state as the anchor. For non-stationary EEG tasks (motor imagery, P300, imagined speech), discriminative structure is time-localized, so the loss can be minimized by ignoring temporal position rather than by learning shared global semantics. The paper never specifies which sliding window is used, how multiple windows are pooled into the single e2_g in Eq. 14, or why window length N was chosen. The ablation 'w/o Mirror-Scale' only swaps the augmentation strategy; it does not separate the channel-complement operation from the temporal-window mismatch, so it cannot validate the semantic-alignment assumption. If this assumption fails, the contrastive branch adds noise, and the reported gains would be attributable to scale, reconstruction, or hyperparameters rather than the proposed global-contrast mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoMET, a self-supervised EEG foundation model that combines masked autoencoding with an InfoNCE contrastive loss. A novel 'mirror-scale augmentation' generates a second view by taking the complement of the masked channels, upsampling the signal to 2f, and applying a sliding window over the doubled patch sequence. The model is pretrained on mixed EEG data (>3000 subjects, ~1.91M samples) at three scales (5M, 19M, 151M parameters) and evaluated on ten downstream BCI/clinical datasets with a frozen encoder and linear probing. The central claim is that CoMET-Large achieves state-of-the-art balanced accuracy on most of these datasets, with gains up to 6.9% absolute over CBraMod on BCIC IV 2A, while scaling-law fits and attention analyses support the design.","tokens_in":21462,"tokens_out":4721,"duration_ms":53465,"significance":"If the claims hold, CoMET would be a substantial advance: it is one of the largest publicly described EEG foundation models, uses a plausible combination of local (masked reconstruction) and global (contrastive) objectives, and is tested across a broad set of downstream tasks with statistical tests and ablations. The mirror-scale augmentation is a creative way to build contrastive views from EEG's channel structure. However, the validity of the SOTA comparison and of the proposed contrastive mechanism is not yet established, because the evaluation protocol is not uniform across baselines and because the momentum view may not be semantically aligned with the anchor view. These issues are load-bearing for the main claims.","major_comments":[{"comment":"The main comparison follows 'the original baseline’s downstream strategy', so CBraMod and LaBraM are fine-tuned, while EEGPT and CoMET are linear-probed. The claimed SOTA margins therefore conflate model quality with adaptation protocol. Appendix Tables 5-10 attempt to remedy this but are incomplete and internally inconsistent: Table 6 is labeled 'BCIC IV 2B' yet contains the PhysioP300 rows from Table 1, and Table 8 (FACED) duplicates the first four KaggleERN rows. A complete same-protocol comparison for all ten datasets is required; the current tables do not establish SOTA under a fixed evaluation protocol.","section":"Table 1 / Appendix 9"},{"comment":"The momentum positive view is not obviously the same brain state as the MEM anchor. After upsampling from f to 2f with fixed patch length, N patches cover only T/2 seconds, and the sliding window of length N then selects a half-trial segment on the complement channels. The text never specifies how the 'group of different augmentation views' is pooled into the single e2_g used in Eq. (14), nor why window length N was chosen. For non-stationary tasks (motor imagery, P300, imagined speech), the InfoNCE loss can be minimized by ignoring temporal position rather than by learning shared global semantics. The w/o Mirror-Scale ablation changes both the channel complement and the temporal window, so it cannot validate the alignment assumption.","section":"Methodology, Mirror-Scale Augmentation (Eqs. 10-11)"},{"comment":"The ablation study is reported only graphically, and the claimed numerical reductions ('up to 6.98% on TUEV, 5.54% on BCIC IV 2A') are not tied to specific conditions or repeated-seed standard deviations in the text. More importantly, the w/o Mirror-Scale condition replaces the proposed augmentation with random-crop masking, so it jointly changes two factors: the channel-complement operation and the temporal-window length. A controlled ablation is needed—e.g., full-trial complement channels with no window, or half-trial non-complement channels—to attribute the downstream gains to the mirror-scale mechanism rather than to longer context or to channel-disjoint positives.","section":"Ablation Study / Figure 3"}],"minor_comments":[{"comment":"BCIC2020-3 is listed as 20 subjects in Table 2 but as 15 subjects in the dataset description; the number of samples (8000) should also be reconciled with the per-class trial counts.","section":"Appendix 2, Table 2"},{"comment":"The downstream output dimensions read '16, classes', '32, classes', '48, classes', which appears to be a formatting error; clarify the intended linear-head dimensions.","section":"Table 4"},{"comment":"The scaling-law constants are fitted to the model's own accuracy points. This is fine as a descriptive summary, but the text should avoid calling it a predictive scaling law without evaluation on held-out model sizes or data fractions.","section":"Appendix 5"},{"comment":"Typos and inconsistent naming: 'LabraM' vs 'LaBraM', 'CbraMod' vs 'CBraMod', 'banlanced', 'schedULered'. Also, Figure 3 is referenced without a table of numerical values, making the ablation hard to audit.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The contribution is potentially strong, but the SOTA claim rests on a non-uniform evaluation protocol, and the mirror-scale view has a temporal mismatch that could undermine the proposed mechanism. I recommend asking the authors for a complete same-protocol benchmark and a controlled validation of the augmentation before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the most serious EEG foundation model claim I've seen this year, and the broad empirical base is real, but I wouldn't trust the exact numbers until the code is out and the mirror-scale augmentation is actually validated.\n\nWhat's new: the channel-complement 'mirror-scale' augmentation for contrastive learning is genuinely EEG-specific, and pairing it with MAE-style reconstruction is a reasonable response to the known failure modes of both CL (attention collapse) and MEM (local similarity overfit). The pretraining corpus (over 1M samples) and ten downstream datasets with ablations and Wilcoxon tests are a solid effort. The scaling-law analysis is a nice addition, though the fits in Section 5 have an apparent text/figure inconsistency that needs fixing.\n\nWhere I get off the bus: (1) No code or weights. For a 151M-parameter foundation model, that's not optional. (2) The main table mixes each baseline's preferred adaptation strategy. The appendix gives both linear probing and fine-tuning for some datasets, but the labels are scrambled (e.g., the 'BCIC IV 2B' table in the appendix looks like PhysioP300). That undermines trust in the headline comparison. (3) The mirror-scale augmentation itself: after 2x upsampling, the sliding window of length N covers half the trial. So the momentum 'positive' is a half-length, channel-disjoint view. The paper doesn't say how the multiple windows are pooled into the single global token for the InfoNCE loss, or why window length N is right, or what happens to temporal position. For non-stationary tasks like motor imagery, this could be aligning apples to oranges. The ablation 'w/o mirror-scale' only swaps the augmentation strategy; it doesn't isolate the temporal windowing from the channel complement, so it cannot validate the alignment assumption. This is the load-bearing part of the claimed mechanism, and it deserves a dedicated experiment (e.g., full-trial upsampled views, different window lengths).\n\nNone of these are fatal in the sense that the central strength—a large, well-evaluated encoder—would likely survive. The self-supervised training is not circular, the baselines are re-implemented with official code, and the reported gains are large and often consistent across model sizes. But a reader should treat the exact numbers as provisional. This paper deserves a serious referee; I'd want revision before acceptance, with code/weights plus the mirror-scale validation.","headline":"Serious, large-scale EEG foundation model with a plausible SOTA claim, but the signature augmentation is under-specified and the paper isn't reproducible until code and weights are released.","tokens_in":21912,"tokens_out":5851,"would_cite":true,"duration_ms":56325,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A self-supervised EEG model, CoMET, pairs masked reconstruction with a mirror-scale contrastive view and reports the best results on eight of ten downstream brain-signal datasets.","keywords":["EEG foundation model","self-supervised learning","masked autoencoder","contrastive learning","brain-computer interface","mirror-scale augmentation","linear probing","universal representation"],"falsifier":"Pre-train CoMET-Base twice, once with the proposed complement-channel mirror-scale view and once with a random channel subset of the same size, keeping all other settings identical; if balanced accuracy on BCIC IV 2A changes by less than the reported gap, the complement-channel mechanism is not what carries the contrastive gain.","tokens_in":20923,"feed_emoji":"🧠","tokens_out":8380,"duration_ms":88689,"temperature":0.7,"texified_at":"2026-08-05T20:19:09.873602+00:00","pith_summary":"CoMET is a self-supervised EEG foundation model whose central claim is that masked reconstruction and contrastive learning solve complementary failure modes and are best trained together. Masked-only models exploit the high correlation of neighboring EEG channels and settle for local, low-dimensional similarity; contrastive-only models collapse their attention onto a few tokens. CoMET pairs an MAE-style reconstruction branch with a momentum-encoder contrastive branch whose second view is produced by mirror-scale augmentation: the complementary channels, upsampled and windowed, are treated as the same brain state. Pre-trained on more than one million samples from over 3,000 subjects and frozen for downstream evaluation, its 151-million-parameter version outperforms previous brain foundation models on eight of ten BCI and clinical datasets, including a 6.9-point balanced-accuracy gain on four-class motor imagery. If the claim holds, a single encoder can extract both local context and global discriminative patterns across heterogeneous devices without retraining.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":5307,"prompt_tokens":723,"completion_tokens":4584,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":723,"completion_tokens_details":{"reasoning_tokens":3950}},"feed_headline":"CoMET EEG model beats prior best on 8 of 10 downstream tasks","feed_subtitle":"A frozen 151M-parameter encoder transfers across BCI and clinical EEG tasks with linear probing alone.","key_machinery":"The load-bearing object is the mirror-scale augmentation that generates the second contrastive view. Given a random half of channels $V_a$ visible to the masked reconstruction branch, the momentum branch sees the complementary channels $V_b = \\text{all channels} \\setminus V_a$; the sample is then upsampled from $f$ to $2f$, doubling temporal patches while keeping patch length and the visible-channel list, and a length-$N$ sliding window with stride $1$ creates several views per sample. Positive pairs are views of the same trial; negative pairs are views of other trials in the batch. Supporting this are the learnable global token, which aggregates whole-brain discriminative information through every self-attention laye","core_discovery":"The paper's discovery claim is that contrastive learning and masked EEG modeling need not be alternatives: fused in one model, each corrects the other's bias. The authors trace the masked-only failure to volume conduction, which makes neighboring electrodes record near-identical signals, so reconstruction can be solved with local similarity and never learns cross-region discriminative structure; the contrastive-only failure is attention collapse, where tokens become homogeneous and the model latches onto task-irrelevant sample-level cues. CoMET keeps an MAE-style asymmetric encoder-decoder with channel-time decoupling embeddings, a 50% mask ratio, and a learnable global token, and adds a mom","pith_inferences":["An unstated corollary: any multi-channel biosignal with fixed electrode geometry and strong inter-channel correlation—ECoG, high-density EMG arrays, possibly MEG—could benefit from the same complement-channel contrastive view, though the paper only tests scalp EEG.","A decisive test the authors leave implicit: replace the complement-channel view with a random channel subset of equal size while keeping everything else fixed. If linear-probe accuracy stays the same, the complementarity of channels is not what drives the gain.","The sliding-window upsample multiplies the number of positive pairs per sample, so part of the contrastive benefit may come from temporal multi-crop augmentation rather than channel complementarity; isolating those factors would sharpen the architectural claim."],"forward_implications":["A frozen CoMET encoder can be adapted to unseen BCI and clinical tasks with a linear head, without retraining on each dataset's channels or sampling rate; differences in montage and duration are absorbed by the learned channel-time embeddings.","Downstream performance grows log-linearly with both model size (R^2 ≈ 0.95–0.99) and pre-training data size (R^2 ≈ 0.92–0.98), so extending the same pre-training recipe to more data and larger encoders should continue to lift accuracy.","Tasks that require coordinated activity across multiple brain regions—motor imagery, imagined speech, emotion—gain the most, while frequency-local tasks like SSVEP are less sensitive to the contrastive branch.","The two losses contribute different information: removing the reconstruction loss hurts by up to about 7 points and removing the contrastive loss by up to about 5.5 points on the tested sets, so the gains come from combining both, not from either alone."],"supporting_citations":[{"why":"Supplies the MAE-style asymmetric encoder-decoder and masked-token reconstruction design that CoMET adapts to EEG patches.","marker":"He et al. 2022"},{"why":"Supplies the momentum encoder and exponential-moving-average update used to generate stable contrastive targets.","marker":"He et al. 2020"},{"why":"Supplies the InfoNCE objective that aligns the two mirror-scale views and separates negative samples.","marker":"Oord, Li, and Vinyals 2018"},{"why":"Provides the volume-conduction evidence that neighboring EEG channels are highly correlated, motivating the complement-channel second view.","marker":"Brunner et al. 2016"},{"why":"Establishes the masked-reconstruction baseline (LaBraM) and the preprocessing conventions CoMET follows for pre-training.","marker":"Jiang, Zhao, and Lu 2024"},{"why":"Supplies the strongest masked baseline (CBraMod), the 50% mask ratio, and the comparison CoMET must beat.","marker":"Wang et al. 2025"},{"why":"Provides the earlier contrastive EEG pre-training result that motivates the attention-collapse limitation CoMET addresses.","marker":"Kostas, Aroca-Ouellette, and Rudzicz 2021"}],"fun_headline_variants":["CoMET fuses masked and contrastive learning for universal EEG","EEG model fixes masking and contrastive pitfalls with fusion","CoMET ends volume conduction bias in EEG foundation models","Masked-only and contrastive-only EEG models fail; CoMET fuses them","CoMET: one universal EEG model beats prior best on 8/10 tasks"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Mirror-scale augmentation assumes the upsampled complementary-channel view represents the same brain state as the masked-reconstruction view; if the channel split drops task-relevant information, the contrastive branch adds noise instead of global structure.","fun_headline_variants_meta":{"raw":{"variants":["CoMET fuses masked and contrastive learning for universal EEG","EEG model fixes masking and contrastive pitfalls with fusion","CoMET ends volume conduction bias in EEG foundation models","Masked-only and contrastive-only EEG models fail; CoMET fuses them","CoMET: one universal EEG model beats prior best on 8/10 tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001333,"raw_usage":{"total_tokens":5248,"prompt_tokens":726,"completion_tokens":4522,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":4431}},"tokens_in":470,"tokens_out":4522,"duration_ms":36877,"temperature":1.0,"reasoning_tokens":4431,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:44:09.112603+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pre-train CoMET-Base twice, once with the proposed complement-channel mirror-scale view and once with a random channel subset of the same size, keeping all other settings identical; if balanced accuracy on BCIC IV 2A changes by less than the reported gap, the complement-channel mechanism is not what carries the contrastive gain.","supporting_citations":[],"review_version":1}