{"id":"5eff6934-7c25-4825-8ce8-b36a949fc02c","arxiv_id":"2608.13072","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EEG-PRIME aligns EEG signals with frozen text embeddings of class labels under task, dataset, and subject-invariance conditioning, achieving strong cross-dataset decoding and competitive zero-shot motor-imagery transfer.","lead":"EEG-PRIME is a two-stage EEG foundation model that pretrains an encoder on unlabeled brain signals, then uses language instructions and text class labels to decode across many datasets. On 16 benchmark datasets it exceeds prior EEG models on most tasks, and on two held-out datasets it reaches nearly the accuracy of per-person calibrated classifiers without using any target data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot protocol omits how dataset embedding e_d is supplied for unseen datasets; because Table IV shows e_d strongly affects inference, the zero-shot numbers are not fully interpretable.","rationale":"The reader's weakest assumption is the same load-bearing concern I would raise: the dataset embedding is undefined for unseen datasets, and the paper's own ablation (Table IV) shows this component materially changes frozen-model inference. The zero-shot claim is the strongest and most distinctive claim in the paper, so an ambiguous protocol for a conditioning input that directly enters Eq. (15) is a genuine correctness and reproducibility risk. I considered whether the missing NeuroLM baseline is more load-bearing; it weakens the 'consistent improvements over prior EEG foundation models' phrasing, but the zero-shot conclusion does not depend on it, and the paper includes several other strong foundation-model baselines. I also note the paper's limitation paragraph honestly narrows zero-shot claims to MI, which mitigates overreach. A conditional verdict is appropriate: the concern is resolvable with a small, well-defined experiment, and the rest of the empirical story is extensive and coherent.","tokens_in":22282,"tokens_out":5028,"duration_ms":51162,"concrete_test":"Evaluate both held-out datasets with three explicit e_d settings: (i) e_d = 0, (ii) e_d = mean of the learned dataset embeddings, (iii) Eq. (6) with alpha = 0 (task-only). Keep all other preprocessing, text prompts, and prototypes identical, and report per-subject balanced accuracy and mean +/- std over seeds for Dreyer2023A and Weibo2014. If the reported 63.0/64.2 values are reproduced only under one of these settings, amend the paper to state that setting and rerun the in-domain no-DS-emb condition from Table IV on the MI subset to check consistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central zero-shot claim is not fully specified. Eq. (6) defines the conditioning vector as e_cond = e_ins + alpha * e_d, with e_d a per-dataset learned embedding. For the two held-out datasets (Dreyer2023A, Weibo2014), no e_d exists, yet Section III-D and Eq. (15) do not state what is substituted at test time (zero vector, averaged embedding, task-only conditioning, or something else). This matters because Table IV shows the learned dataset embedding is not a minor detail: removing it lowers in-domain direct inference from 0.4848 to 0.4242 balanced accuracy (Kappa 0.2179 to 0.1111). If the zero-shot evaluation silently uses e_d = 0, the model is being run in a configuration that Table IV identifies as substantially degraded, and the reported 63.0%/64.2% balanced accuracy needs to be reconciled with that ablation. If instead a nonzero e_d is supplied (e.g., the mean of training dataset embeddings), that choice is a form of dataset-level prior information and must be disclosed and ablated. Without this specification, the zero-shot 'comparable to within-session calibration' conclusion is not interpretable or reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EEG-PRIME, a two-stage EEG foundation model for cross-dataset multi-task decoding. In the first stage, an EEG encoder is pretrained on roughly 1,153 hours of unlabeled EEG from nine datasets using masked reconstruction with frequency-cutoff spectral augmentation. In the second stage, the frozen encoder is paired with a Q-Former whose queries are modulated by a conditioning vector that combines a frozen SBERT task-instruction embedding and a learned per-dataset embedding; a gradient-reversal adversarial loss encourages subject invariance, and a query-diversity loss prevents query collapse. Classification is performed by cosine similarity between the Q-Former output and frozen text prototypes of class labels, removing the need for dataset-specific classifier heads. The authors evaluate on sixteen datasets for task-specific fine-tuning and two held-out motor-imagery datasets for zero-shot inference, reporting gains over EEGNet, TSception, ST-Transformer, Conformer, BIOT, EEGPT, LaBraM, CBraMod, and MIRepNet, with three-seed runs, Wilcoxon tests, Cohen's d, and ablations of the text encoder, instruction, and dataset embedding.","tokens_in":22551,"tokens_out":10127,"duration_ms":99579,"significance":"If the results hold, EEG-PRIME contributes a genuinely useful design in the prototype-aligned unified head and the layer-wise query modulation mechanism, and the breadth of the evaluation (18 datasets, five paradigms, three seeds, statistical testing, and an honest limitation section) is a clear strength. The release of code and pretrained models is also commendable. However, the central zero-shot claim is currently not fully interpretable because the paper does not specify how the learned dataset embedding is supplied for the two held-out datasets, and the fine-tuning comparison is potentially confounded by the unstated use of downstream training splits during instruction tuning. The absence of NeuroLM, the closest language-aligned EEG foundation model, further weakens the comparative claim. These issues are substantial but appear fixable within the scope of a major revision.","major_comments":[{"comment":"The zero-shot protocol does not specify what value the learned dataset embedding e_d takes for the held-out datasets Dreyer2023A and Weibo2014. Since e_cond = e_ins + alpha * e_d in Eq. (6), and Table IV shows that removing e_d degrades in-domain direct inference from 0.4848 to 0.4242 balanced accuracy, the zero-shot results in Section V.B cannot be interpreted without knowing the test-time e_d. If e_d is set to zero for unseen datasets, the model is evaluated in a configuration that Table IV identifies as substantially degraded; if a non-zero value is used (e.g., the mean of trained dataset embeddings), that choice constitutes a dataset-level prior and must be disclosed and ablated. Please state the exact procedure used at zero-shot inference and, if a non-zero default is used, ablate that choice.","section":"Section III-C1, Eq. (6); Section III-D, Eq. (15); Table IV"},{"comment":"The manuscript does not state which datasets are used in the instruction-tuning stage, but the sentence in Section IV.A that the two held-out datasets are 'not seen during instruction tuning or adaptation' strongly implies that the sixteen downstream datasets are used for instruction tuning. If instruction tuning is performed on the training splits of those same sixteen datasets, then the task-specific fine-tuning results in Table II compare EEG-PRIME, which has already received a supervised multi-dataset instruction-tuning pass on the target training data, against baselines that are fine-tuned directly from their pretrained checkpoints. This would confound the fine-tuning comparison. Please clarify the instruction-tuning data, and if the downstream training splits are used, add an equivalent multi-task instruction-tuning stage for all baselines or restrict instruction tuning to the pretraining corpus and re-evaluate.","section":"Section IV.A and Section III-D"},{"comment":"The baseline list omits NeuroLM (reference [5]), which is the closest prior work to the proposed method: it is a language-aligned EEG foundation model that performs instruction tuning and would be the most direct comparison for the paper's central claims. Without NeuroLM, the statement that EEG-PRIME achieves 'consistent improvements over ... prior EEG foundation models' is not fully supported. Please add NeuroLM to the fine-tuning and zero-shot comparisons where feasible, or provide a concrete justification for its exclusion (for example, unavailability of a pretrained checkpoint), and update Tables II and III accordingly.","section":"Section IV.B, Section II.B"}],"minor_comments":[{"comment":"The Dreyer2023A dataset is cited as reference [29] in the dataset list but as reference [41] in the zero-shot section; please cite the correct dataset release consistently.","section":"Section IV.A, Section V.B"},{"comment":"The rows of Table IV are ambiguous because the checkmark placement does not clearly indicate which variant corresponds to the 'default instruction' condition and which corresponds to the 'no dataset embedding' condition; please label the rows explicitly.","section":"Table IV"},{"comment":"The pretraining loss ell_mask(Encoder_theta, X_tilde, m) is not formally defined; please specify the exact mean squared error computation over masked token positions.","section":"Eq. (5)"},{"comment":"The abstract and Section V.A describe the evaluation as being under 'cross-subject settings,' but Table I uses cross-trial splits for SEED, SEED-IV, SEED-V, and SEED-VII; please qualify the claim to 'cross-subject and cross-trial settings' where appropriate.","section":"Abstract and Table I"},{"comment":"The problem formulation writes the model as f_theta(X; d, s), but the subject identity s is never provided as an input at inference; subject conditioning is implemented only through the gradient-reversal adversarial loss in Section III-C2. Please align the notation with the actual mechanism.","section":"Eq. (1) and Section III-C2"}],"recommendation":"major_revision","confidential_remarks":"The reader's stress-test concern about the zero-shot dataset embedding is valid and should be the primary revision point. I also found what appears to be a protocol confound: the instruction-tuning stage seems to use the same downstream datasets that are later fine-tuned, which would put the baselines at a disadvantage in Table II. The paper is otherwise well-executed, with broad experiments, statistical tests, and honest limitations; the requested fixes are feasible. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: EEG-PRIME is a serious, broad EEG foundation model paper, and the architecture is plausible, but the headline zero-shot result is under-specified because the paper never says what happens to the learned dataset embedding when evaluating on a dataset that wasn't in training. That needs to be fixed before the claim is interpretable.\n\nWhat's actually new: the combination of a Q-Former with layer-wise scale-shift modulation, per-dataset learnable embeddings, gradient-reversal subject invariance, and frozen text prototypes for unified classification across heterogeneous label spaces. Individual pieces exist, but I don't know of another EEG model that puts them together and evaluates on 18 datasets across five paradigms. The ablations of text encoder and conditioning are useful, three seeds with Wilcoxon and Cohen's d is about as good as this literature gets, and they promise code and models. The ERD-attention correlation is a nice sanity check even if it's not causal.\n\nWhere it wobbles: first, the zero-shot protocol. Eq. (6) defines e_cond = e_ins + alpha*e_d, but the held-out datasets have no e_d. The inference section (III-D) and Eq. (15) don't say what's used at test time. That matters because Table IV shows removing e_d drops in-domain direct inference from 0.4848 to 0.4242 balanced accuracy. If the zero-shot evaluation used e_d = 0, the model is running in a distinctly weaker configuration and still lands at 63-64% -- that would be impressive and should be stated. If they used some average or a heuristic, that's a dataset-level prior that must be disclosed and ablated. As written, the central zero-shot claim is not reproducible.\n\nSecond, NeuroLM is missing from the baselines. It's the closest language-aligned EEG foundation model and they cite it in related work, so leaving it out weakens the SOTA claim. This is probably fixable in revision.\n\nMinor: no ablation for the subject-invariance loss or frequency-cutoff augmentation, both of which are presented as components. And the Wilcoxon across datasets treats each dataset as one observation, which is loose but tolerable as a summary.\n\nWho it's for: anyone working on EEG foundation models or cross-dataset decoding. The broad benchmark and the text-prototype idea are worth engaging with. The issues are real but fixable; this deserves a serious referee, not a desk reject. I'd send it reviewing, with a request to pin down the zero-shot inference procedure.","headline":"Solid broad EEG foundation model with a serious evaluation, but the headline zero-shot result isn't interpretable until they specify what happens to the dataset embedding for unseen datasets.","tokens_in":23088,"tokens_out":3800,"would_cite":true,"duration_ms":36939,"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":"This paper claims that a single frozen EEG model, conditioned on task language and dataset identity, can decode across BCI paradigms and match calibration-based decoders on unseen datasets.","keywords":["EEG foundation model","instruction tuning","multi-level conditioning","prototype classification","zero-shot transfer","motor imagery","emotion recognition","brain-computer interface"],"falsifier":"Re-run the zero-shot evaluation on the two held-out motor-imagery datasets with the dataset embedding $e_d$ explicitly fixed to the zero vector, and compare to the reported 63.0% and 64.2% balanced accuracy; a large drop would show that the zero-shot results depend on a dataset signal not described in the protocol.","tokens_in":22098,"feed_emoji":"🧠","tokens_out":6931,"duration_ms":70421,"temperature":0.7,"pith_summary":"The paper proposes EEG-PRIME, a two-stage EEG foundation model that tries to make one frozen network decode many brain-computer interface tasks at once. Its central idea is that language should steer how the model reads brain signals, not rewrite the signals themselves: a task description and a learned dataset embedding modulate a Q-Former's queries at every layer, and predictions are made by matching the resulting EEG embedding to frozen text embeddings of class labels. If true, a practical decoder could be deployed on a new person or lab without collecting calibration trials or retraining task-specific heads. On sixteen datasets the model reports the best balanced accuracy on thirteen and top-three on all, and on two held-out motor-imagery datasets the frozen model reaches balanced accuracy comparable to per-subject calibrated decoders.","feed_headline":"Frozen EEG model matches within-session decoders on unseen users","feed_subtitle":"Text-anchored prototypes let one model decode across BCI paradigms without per-dataset retraining or calibration.","key_machinery":"The load-bearing mechanism is a modified Q-Former, a compact transformer whose learnable query vectors read out EEG token sequences by cross-attention. The conditioning vector $e_{\\mathrm{cond}} = e_{\\mathrm{ins}} + \\alpha e_d$, where $e_{\\mathrm{ins}}$ is a frozen sentence embedding of the task instruction and $e_d$ is a learned dataset embedding, enters every Q-Former sublayer through Layer-wise Query Modulation (LQM): each sublayer gets per-dimension scale $\\gamma(e_{\\mathrm{cond}})$ and shift $\\beta(e_{\\mathrm{cond}})$ after layer normalization. A gradient-reversal subject classifier pushes the summary embedding $z$ to be subject-invariant, a query-diversity penalty stops query collapse, and the final prediction is $\\arg\\max_y \\cos(z, p_y)$ against frozen text prototypes $p_y$ of class names. This replaces all task-specific heads with one semantic similarity comparison, supported by a masked-autoencoding pretraining stage with frequency-cutoff spectral augmentation.","core_discovery":"EEG-PRIME's central claim is that cross-dataset, cross-subject EEG decoding can be unified without dataset-specific classifier heads: a single prototype-based head over text class embeddings, steered by task-instruction and dataset conditioning, generalizes across motor imagery, emotion recognition, ADHD detection, covert speech, and mental workload. Concretely, the paper reports winning or top-three balanced accuracy on all sixteen fine-tuning benchmarks and a statistically significant advantage over every baseline, with large effect sizes. On the two held-out datasets, the frozen model obtains 63.0% mean balanced accuracy across 60 subjects versus 62.8% for within-session CSP+LDA, and 64.2% versus 66.2–68.1% for supervised leave-one-subject-out baselines. The paper reads these numbers as evidence that zero-shot EEG decoding is feasible for paradigms with clear neural correlates, while noting in its limitations that zero-shot transfer is not yet uniform across all paradigms.","pith_inferences":["A reader extending the method would need to pin down the zero-shot dataset-embedding protocol; the paper's ablation makes this non-optional, since dropping $e_d$ cuts frozen in-domain inference from 0.4848 to 0.4242 balanced accuracy.","The dominance of scale over shift in LQM may transfer to other instruction-conditioned transformer decoders beyond EEG; a cheap test is ablating $\\gamma$ and $\\beta$ in a language-conditioned audio or image decoder.","Because class prototypes are frozen text embeddings, the label vocabulary's semantics should matter: using synonymous or paraphrased class names should shift zero-shot accuracy, and this could be used as a deliberate probe of the EEG–language alignment.","The paper's own limitation note implies that zero-shot transfer will likely need per-paradigm calibration for emotion and other tasks with weaker shared neurophysiology, so a practical rollout would pair this model with lightweight domain adaptation for those paradigms."],"forward_implications":["If the result holds, new BCI datasets can be added without designing a new classifier head: only the class-name text prototypes and a task instruction are needed.","Zero-shot deployment becomes practical for motor imagery: a frozen model can be run on a new user's session with no calibration, at accuracy comparable to per-subject calibrated CSP+LDA.","Instruction conditioning in the Q-Former gives a principled place for language in EEG decoding: semantics modulate query geometry rather than being concatenated with EEG tokens, so attention is steered without distorting the neural signal.","The finding that scale modulation $\\gamma$ drives most of the gain suggests LQM's effectiveness comes from multiplicative reshaping of normalized query states, and that shift-only variants are insufficient.","The Q-Former's cross-attention aligns with event-related desynchronization (correlation $r=0.83$), indicating the learned reader focuses on the same temporal windows a neurophysiologist would, supporting interpretability."],"supporting_citations":[{"why":"Prior EEG foundation model based on transformer pretraining; serves as a comparison baseline for cross-dataset generalization.","marker":"[2]"},{"why":"Masked-autoencoding EEG foundation model over a large heterogeneous corpus; the main pretraining lineage and a strong baseline.","marker":"[3]"},{"why":"Criss-cross brain foundation model with strong cross-task generalization; the strongest prior foundation-model baseline in the comparison.","marker":"[4]"},{"why":"Text-aligned EEG foundation model that injects language tokens into an LLM; provides the contrasting approach to EEG-PRIME's query-modulation conditioning.","marker":"[5]"},{"why":"Frozen Sentence-BERT encoder that produces the task-instruction embeddings and text class prototypes, defining the semantic space for prototype classification.","marker":"[13]"},{"why":"Dreyer2023A dataset release supplying the within-session CSP+LDA calibration baseline used for the zero-shot comparison.","marker":"[41]"},{"why":"Supplies the supervised leave-one-subject-out baselines on Weibo2014 used for the zero-shot comparison.","marker":"[42]"}],"fun_headline_variants":["Zero-shot EEG decoding matches per-user calibration","One EEG model, zero calibration: decodes unseen users","Cross-subject EEG decoding without per-user training","Prototype-tuned EEG model transfers across datasets","Frozen EEG net rivals per-session calibration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a meaningful dataset-conditioning vector can be supplied for a dataset never seen during training, but the paper does not specify what $e_d$ is set to at zero-shot test time, and the ablation shows that dropping it substantially hurts frozen inference.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot EEG decoding matches per-user calibration","One EEG model, zero calibration: decodes unseen users","Cross-subject EEG decoding without per-user training","Prototype-tuned EEG model transfers across datasets","Frozen EEG net rivals per-session calibration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000676,"raw_usage":{"total_tokens":3082,"prompt_tokens":960,"completion_tokens":2122,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":2050}},"tokens_in":576,"tokens_out":2122,"duration_ms":15204,"temperature":1.0,"reasoning_tokens":2050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:22:29.812754+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the zero-shot evaluation on the two held-out motor-imagery datasets with the dataset embedding $e_d$ explicitly fixed to the zero vector, and compare to the reported 63.0% and 64.2% balanced accuracy; a large drop would show that the zero-shot results depend on a dataset signal not described in the protocol.","supporting_citations":[{"cited_title":"Eegpt: Pretrained transformer for universal and reliable representation of eeg signals,","cited_arxiv_id":null,"evidence_quote":"Prior EEG foundation model based on transformer pretraining; serves as a comparison baseline for cross-dataset generalization."},{"cited_title":"A large eeg database with users’ profile information for motor imagery brain- computer interface research,","cited_arxiv_id":null,"evidence_quote":"Dreyer2023A dataset release supplying the within-session CSP+LDA calibration baseline used for the zero-shot comparison."},{"cited_title":"Svm-enhanced attention mechanisms for motor imagery eeg classification in brain-computer interfaces,","cited_arxiv_id":null,"evidence_quote":"Supplies the supervised leave-one-subject-out baselines on Weibo2014 used for the zero-shot comparison."}],"review_version":1}