{"id":"d5f9511e-aa14-4270-9db5-348031138956","arxiv_id":"2608.00114","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Structured latent prediction—EMA contextual targets, neurotopology-aware masking, and multi-depth supervision—raises frozen 14-task EEG balanced accuracy from 40.5% to 50.4% (52.9% with multi-source data).","lead":"EEG-JEPA trains EEG models to predict abstract summaries of masked brain-signal regions produced by a teacher network, instead of reconstructing the missing raw waveforms. On a 14-task EEG benchmark, the same backbone improves average balanced accuracy from 40.5% to 50.4% (52.9% with extra pretraining data), a step toward reusable EEG representations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No subject-level split reported between TUEG pretraining and TUAB/TUEV benchmark evaluation; the 10-point frozen gap may reflect record overlap rather than transferable neural structure.","rationale":"The reader's weakest assumption concerned the general validity of the frozen-transfer benchmark as a proxy for representation transfer. I am going one step further and identifying a concrete, fixable flaw in that benchmark's construction: potential subject/record leakage between the pretraining corpora (TUEG; TDBRAIN/HBN in Stage 2) and the downstream evaluation tasks (TUAB, TUEV, and possibly seizure/sleep tasks). This is more load-bearing than a generic 'benchmark validity' concern because it threatens the internal validity of the controlled comparison itself, not just its external generalizability. The paper provides no mention of subject-level splitting or de-duplication, which is a missing control. If overlap exists, the large frozen accuracy gap could be due to memorizing subject/record-specific statistics, which would contradict the paper's core claim about learning transferable neural structure. The controlled setup (same backbone, corpus, and training duration) is a genuine strength, and the ablations are well structured, which is why I do not recommend rejection outright. However, the leakage test is a necessary condition for accepting the causal claim. The reader's verdict of CONDITIONAL is therefore appropriate; my concern does not change the verdict but sharpens the condition that must be met.","tokens_in":13592,"tokens_out":15716,"duration_ms":169115,"concrete_test":"Compute the exact subject/record overlap between TUEG pretraining crops and the TUAB and TUEV evaluation sets, and between TDBRAIN/HBN and the benchmark's seizure/sleep tasks. Then re-run the frozen 14-task evaluation with all overlapping subjects or records removed from pretraining (or from downstream evaluation), and recompute macro BA for the CBraMod MAE control and EEG-JEPA. If the 40.49 vs 50.42 gap collapses below, say, 5 points, the central claim that objective design causes general transfer is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim attributes the 40.49% to 50.42% frozen macro-BA gain to objective design (EMA latent targets, N-MET, hierarchical depth supervision). The evidence rests on the EEG-FM-Bench 14-task frozen protocol. However, Stage-1 pretraining uses TUEG (39,758 records), while two benchmark tasks, TUAB and TUEV, derive from the Temple University Hospital EEG corpus. The manuscript never states that subjects or records were de-duplicated between pretraining and downstream evaluation. If TUEG crop-level pretraining overlaps with TUAB/TUEV test records, the frozen MLP heads can exploit subject- or record-specific signatures (acquisition effects, artifacts, subject identity) rather than task-relevant neural structure. This directly undercuts the paper's stated motivation: that latent prediction avoids 'subject- or device-specific patterns' and learns 'reusable neural structure.' Even though CBraMod is pretrained on the same TUEG data and thus matched, the relative 10-point gap could still be inflated if EEG-JEPA's objective better preserves the non-neural statistics that the leaked benchmark rewards. A second, related leakage risk is Stage 2: TDBRAIN and HBN are added, and these corpora may overlap with seizure/sleep benchmark tasks, threatening the 52.94% 'highest average' claim. In the absence of any subject-level split description, the controlled comparison is not yet clean evidence for the causal claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents EEG-JEPA, a latent-prediction objective for EEG foundation models. The central design is a masked context encoder plus predictor that infers EMA-derived contextual targets at selected electrode–time locations (N-MET), with supervision at four encoder depths. Under a controlled Stage-1 setting (same CBraMod backbone, TUEG corpus, and 100 epochs), it reports raising the 14-task frozen macro balanced accuracy from 40.49±0.32 (masked waveform reconstruction) to 50.42±0.58, and after two-stage continuation to 52.94±0.30, which it claims is the highest among EEG foundation models on EEG-FM-Bench. Protocol-matched full fine-tuning on nine tasks improves the average BA from 68.98 to 70.65. The paper also provides layerwise probe and CKA analyses to support the claim of selective deep organization of interaction and cross-channel information.","tokens_in":13942,"tokens_out":7086,"duration_ms":66003,"significance":"If the controlled improvement is causal, the paper makes a strong contribution: it cleanly isolates objective design from architecture, corpus, and training duration, provides a systematic ablation over target content, support, and depth, and releases code. The progressive ablations and the parameter-efficient depth-prefix results are useful for practitioners. However, two load-bearing gaps—possible subject/record overlap between pretraining and benchmark evaluation, and the undefined context-consistency loss—prevent the central causal claim from being accepted at face value.","major_comments":[{"comment":"The manuscript does not state whether subjects or records are disjoint between the TUEG pretraining corpus and the TUAB/TUEV tasks in EEG-FM-Bench, even though all derive from the Temple University Hospital corpus. Stage 1 trains on all 39,758 TUEG records; Stage 2 additionally uses TDBRAIN and HBN, which may overlap with seizure/sleep tasks. If TUAB/TUEV test records or subjects are included in pretraining, the frozen MLP heads can exploit subject- or record-specific signatures rather than transferable neural structure, inflating the 10-point controlled gain and the 52.94% 'highest average' claim. Please report the exact de-duplication procedure (subject-level split, record exclusion, or overlap statistics) for both stages, and rerun on disjoint data if overlaps exist.","section":"Pretraining Setup / Data Preprocessing (Tables 2 and 4)"},{"comment":"Table 6 and Figure 3 attribute a 'context consistency' loss with part of the improvement (50.13 → 50.42), but Eqs. (5)–(7) only define L_latent and L_VCReg. No equation or prose defines the context-consistency term L_ctx, its weighting λ_ctx, or its inputs. Without a precise definition, the ablation component is unreproducible and its contribution cannot be verified. Please add the full expression and clarify how it is combined with L_latent.","section":"Training Objective (Eqs. 5–7), Figure 3"},{"comment":"Table 3 reports frozen macro BAs without error bars or statistical significance, including the margins over EEGPT (52.94 vs 52.15) and REVE (52.94 vs 51.50) that are smaller than the reported run-to-run variation of the full model (±0.30 in Table 6). It is also not stated whether the external baselines were re-run under the identical frozen protocol or taken from earlier papers. The 'highest average' claim should be supported by per-task uncertainties and a clear statement of which numbers were recomputed in this paper.","section":"Main Results, Table 3"}],"minor_comments":[{"comment":"The text repeatedly refers to an appendix/supplementary for preprocessing details and task-level results, but no such material is included in the submitted manuscript. Please provide it or indicate where it is available.","section":"Appendix / Supplementary"},{"comment":"Typo: 'trainnig' should be 'training'.","section":"Table 2 caption"},{"comment":"It is ambiguous which baseline numbers in Table 5 were re-run under the same protocol versus taken from the literature. The sentence 'External baseline results are taken from REVE' should be expanded to specify the source for each baseline and confirm protocol matching.","section":"Full fine-tuning comparison"},{"comment":"The discussion of L3/L6/L9 prefixes compares them against the full CBraMod encoder, which is parameter-matched only to the full EEG-JEPA model. Clarify that the prefixes are not parameter-matched to CBraMod and that the comparison is about efficiency of the same pretrained encoder.","section":"Parameter-efficient transfer"},{"comment":"All transfer conclusions rest on the frozen mean-pooled linear-head protocol on 14 selected tasks. The paper could usefully acknowledge that this protocol is one particular yardstick and discuss how the conclusions might change under alternative readouts or task distributions.","section":"Evaluation Protocol"}],"recommendation":"major_revision","confidential_remarks":"The Stage-1 comparison is carefully matched and is the strongest part of the paper. If the authors can provide a subject-level disjointness guarantee between TUEG and TUAB/TUEV (and between the Stage-2 corpora and the benchmark tasks) and define the context-consistency loss explicitly, I would expect the paper to be publishable after verification. The missing appendix/supplementary should also be supplied."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a legitimate attempt to isolate what an EEG foundation model should predict, and the controlled Stage-1 comparisons are the right way to test it. The 40.49% to 50.42% frozen macro-BA gain over CBraMod-style masked reconstruction is the result to scrutinize, not dismiss. If it survives cleanup, it's an important result for the subfield.\n\nWhat is actually new: JEPA-style latent prediction is not new, and the paper cites data2vec, I-JEPA, and EEG2Rep. The contribution is the joint design of target content, structured electrode–time support (N-MET), and hierarchical depth supervision for EEG, plus a progressive ablation that shows each piece adds something. The stage-1 controlled setup—same backbone, corpus, and training duration—is credible. The layerwise probe and CKA analysis is a useful extra that supports the claim that the objective changes how information is organized across depth. Code is released, which matters.\n\nThe soft spots are real but mostly addressable. The stress-test concern is the one I'd want resolved first: TUEG is used for pretraining, and TUAB and TUEV come from the same Temple University Hospital corpus, yet the paper never describes a subject-level or record-level split between pretraining and evaluation. That gap could inflate the frozen transfer numbers, and it directly undercuts the stated motivation of learning 'reusable neural structure' rather than subject- or device-specific patterns. The fact that CBraMod was pretrained on the same TUEG data makes the comparison fairer, but it does not fix the attribution problem: EEG-JEPA's objective might better preserve the record-specific statistics that a leaked benchmark rewards. This needs to be checked, not assumed away.\n\nMinor mechanical issues: the context-consistency loss is named in Figure 3 and Table 6 but never defined in the objective equations—Equation (7) only includes the latent prediction loss and VCReg. Table 3 lacks error bars, unlike Table 6. The 52.94% headline uses a second-stage corpus mixture that most baselines were not trained on, so it should not be presented as a clean comparison; to the paper's credit, the ablation table does separate Stage-2 from the controlled Stage-1 results. And the design hyperparameters were selected on the same EEG-FM-Bench where the model is then declared best—common in this literature, but worth stating plainly.\n\nWho is this for: researchers working on EEG foundation models, especially those designing self-supervised objectives. The paper deserves a serious referee: the central claim is important, the experiments are mostly well-controlled, and the flaws are fixable rather than fatal. I would send it to review and ask for the subject-overlap analysis and a defined context-consistency loss before acceptance.","headline":"Worth refereeing; the controlled objective comparisons are solid, but the headline frozen gain needs a subject-level de-duplication check before the causal story holds.","tokens_in":14490,"tokens_out":1736,"would_cite":true,"duration_ms":19132,"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":"EEG models trained to predict latent neural states from structured context transfer better than models trained to reconstruct raw waveforms, raising frozen-task accuracy from 40.5% to 50.4% on the same backbone, corpus, and training budget.","keywords":["EEG foundation models","latent prediction","masked waveform reconstruction","self-supervised pretraining","transfer learning","electrode–time masking","hierarchical representation","EMA target encoder"],"falsifier":"Train the exact same CBraMod-style masked-reconstruction objective with N-MET structured masking and multi-depth supervision (predict raw waveform patches at L3, L6, L9, and L12 with the same predictor). If its 14-task frozen balanced accuracy under the matched Stage-1 protocol reaches about 50%, the paper's attribution of the gains to latent content is wrong; if it stays near 40.5%, the latent-target claim is confirmed. A second check: evaluate the released Stage-1 EEG-JEPA checkpoint on a task that demands precise transient timing (e.g., spike detection)—if reconstruction-trained models clea","tokens_in":13465,"feed_emoji":"🧠","tokens_out":6640,"duration_ms":54050,"temperature":0.7,"pith_summary":"This paper tries to settle what an EEG foundation model should predict during self-supervised pretraining. Its claim: rather than reconstructing masked voltage samples, the model should infer contextual latent states—representations of full-crop brain activity produced by an EMA teacher—at structured electrode–time locations and multiple encoder depths. If correct, the design choice itself, not architecture, corpus, or training length, explains a roughly 10-point jump in frozen 14-task balanced accuracy (40.49% to 50.42%), rising to 52.94% with multi-source continuation, and a 1.67-point gain in nine-task full fine-tuning. The paper also shows the resulting representations reorganize across layers, with cross-channel and interaction information shifting deeper, which offers a concrete signature for future EEG pretraining.","feed_headline":"Predicting latent states lifts EEG transfer accuracy 10 points","feed_subtitle":"Same backbone, data, and training time: structured latent prediction beats masked waveform replay, 40.5% to 50.4% frozen.","key_machinery":"Three coupled design choices carry the argument: (1) target content—an EMA target encoder observing the complete crop produces stop-gradient contextual targets, so prediction is of latent states, not voltage; (2) target support—N-MET, a mixture of six electrode–time masking primitives (temporal stripes at 35%, channel stripes 17.5%, local blocks 17.5%, topo-pair stripes 10%, full channel 15%, full topo-pair 5%) that forces inference across temporal continuity, focal activity, regional fields, bilateral correspondence, and sensor loss; (3) target depth—hierarchical prediction of L3, L6, L9, and final-layer targets from fused multi-depth context, with a shared predictor trunk and separate head","core_discovery":"The central claim is that latent prediction is a strictly better pretext task than masked waveform reconstruction for EEG transfer. EEG-JEPA masks structured electrode–time regions (N-MET: temporal stripes, channel stripes, local blocks, topo-pair stripes, full-channel and full-topo pairs), feeds the visible context through an encoder, and trains a predictor to recover the target encoder's full-input contextual representation at those masked locations—with the target encoder updated by exponential moving average and supervision applied at layers 3, 6, 9, and 12. Under controlled comparison (same backbone, corpus, and duration), this objective raises the 14-task frozen macro balanced accuracy","pith_inferences":["If the claim holds, the structured-masking and depth-supervision components can be tested in isolation on a raw-waveform predictor: if those two changes alone close most of the gap, the attribution to latent content is weaker than the paper's ablation implies.","The ridge-probe depth signature could be recycled as a cheap screening metric for future EEG pretraining runs: a model whose cross-channel descriptors peak in shallow layers likely transfers worse to multi-channel classifiers.","A 10-point frozen-transfer gain, if replicated, would make latent prediction the default objective for large EEG corpora and would let downstream applications use a single frozen encoder with task-specific heads rather than per-task fine-tuning.","The paper does not isolate the EMA stabilizer itself: the 0.33-point gain over an independently optimized target encoder is small, so most of the effect must come from content, support, and depth—but the independent-encoder control still updates on the same data; a fixed random target encoder would further clarify how much of the EMA design matters."],"forward_implications":["EEG pretraining objectives can be compared cleanly: the same backbone, corpus, and duration produce a 9.93-point swing in frozen transfer, so future pretraining efforts should treat the objective, not scale alone, as the main lever.","Intermediate encoder prefixes become usable as smaller models: L3/L6/L9 prefixes, without retraining, give 48.87/49.06/49.70 macro balanced accuracy vs 40.49 for the parameter-matched reconstruction encoder.","Representations from latent prediction organize by depth: cross-channel and within-signal interactions peak in deep layers (54.8% and 41.5% of late-layer peaks), a concrete signature that future benchmarks can probe.","Latent prediction with structured support transfers across clinical, sleep, motor-imagery, affective, workload, seizure, depression, and visual tasks under a frozen encoder, demonstrating that one pretrained model can cover heterogeneous EEG applications.","Multi-source continuation (TUEG plus TDBRAIN plus HBN) adds a further 2.52 points, suggesting corpus diversity and latent prediction compound."],"fun_headline_variants":["EEG foundation model: predict latents, not waveforms","Latent prediction lifts EEG transfer by 10 points","Same data, same time: latent beats waveform for EEG","Structured latent prediction wins over masked replay","EEG-JEPA: infer context, boost transfer accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"All conclusions rest on the frozen-transfer yardstick: balanced accuracy of mean-pooled token features fed to dataset-specific MLP heads on 14 EEG-FM-Bench tasks; if that protocol rewards dataset-specific statistics or the tasks are not representative, the 10-point gap does not establish general transferability.","fun_headline_variants_meta":{"raw":{"variants":["EEG foundation model: predict latents, not waveforms","Latent prediction lifts EEG transfer by 10 points","Same data, same time: latent beats waveform for EEG","Structured latent prediction wins over masked replay","EEG-JEPA: infer context, boost transfer accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1146,"prompt_tokens":841,"completion_tokens":305,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":585,"tokens_out":305,"duration_ms":3547,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T00:35:32.022410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the exact same CBraMod-style masked-reconstruction objective with N-MET structured masking and multi-depth supervision (predict raw waveform patches at L3, L6, L9, and L12 with the same predictor). If its 14-task frozen balanced accuracy under the matched Stage-1 protocol reaches about 50%, the paper's attribution of the gains to latent content is wrong; if it stays near 40.5%, the latent-target claim is confirmed. A second check: evaluate the released Stage-1 EEG-JEPA checkpoint on a task that demands precise transient timing (e.g., spike detection)—if reconstruction-trained models clea","supporting_citations":[],"review_version":1}