{"id":"48ee9233-e812-47d6-a515-93202122268f","arxiv_id":"2608.02070","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"STEAM, a dual-branch EEG encoder with soft mixture-of-experts interaction and two-stage pre-training, ranks first on average across seven EEG decoding datasets.","lead":"Researchers built STEAM, an EEG decoding model with two parallel branches for spatial and temporal signal structure and a shared expert module that lets the branches exchange information. Evaluated on seven EEG datasets and fourteen settings, it achieves the best average rank among general EEG foundation models at relatively low inference cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline fairness is the load-bearing assumption: official checkpoints may be fed interpolated 60-channel inputs outside their native formats, with no per-method tuning reported; a validation-tuned re-run could shrink the 9.12-rank margin.","rationale":"The paper is a carefully executed empirical study with clean pre-training/evaluation data separation, a transparent rank-calculation appendix, and extensive ablations (Tables 2-3, 5, 11-12) that support the architecture's internal consistency. The central claim, however, is the benchmark headline, and that claim depends on the baseline comparison being fair. The absence of any reported downstream tuning protocol for baselines, combined with the standardized 60-channel interpolation applied to all inputs, makes it plausible that some foundation models were evaluated outside their native formats. The reader's weakest assumption identifies exactly this point; I agree it is the most plausible invalidation of the headline. The expert-count selection on evaluation datasets (Figure 6) is a separate methodological concern that also warrants disclosure, but it is secondary because the sensitivity analysis shows modest performance differences between 8 and 16 experts. The lack of significance tests and missing code/checkpoints further support the CONDITIONAL verdict. A re-run with per-baseline validation tuning would settle whether the rank margin is real; until then, the headline should be scoped as 'best average rank among methods evaluated with their official checkpoints under a shared 60-channel interface,' and code plus baseline configurations should be released.","tokens_in":22228,"tokens_out":16545,"duration_ms":142131,"concrete_test":"Re-run the fourteen benchmark settings (Table 1) with each of the ten foundation-model baselines using its own native preprocessing and a small validation-based hyperparameter search: for each baseline, hold out 10% of each dataset's training partition as validation; try learning rates {1e-4, 3e-4, 1e-3} and epochs {20, 50} with early stopping; select the best checkpoint by validation BAC/RMSE. Recompute the combined ranks exactly as in Table 13. If STEAM's combined rank is no longer 1.43, or the 9.12-position margin over the mean FM baseline narrows by more than half, the headline 'best average rank' claim is materially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix C ('Aggregate-rank calculation') reports the headline combined rank of 1.43 for Stage-I STEAM versus a mean of 10.55 for ten foundation-model baselines, a 9.12-position margin. This margin is load-bearing because the paper's central claim is 'best average rank' across fourteen settings. The fairness of the comparison is not substantiated: the paper only states that baselines use 'their official implementations and publicly released pre-trained checkpoints' (Appendix C, 'Subject-Level Results'), and reports no per-method downstream hyperparameters, epochs, or learning rates. More concretely, all data are mapped to a standardized 60-channel montage via inverse-distance interpolation (Appendix B) before being fed to models pretrained on their native channel layouts (e.g., LaBraM's channel-wise patching, SingLEM/TFM single-channel inputs, BIOT's multi-signal tokenization). Feeding interpolated 60-channel inputs into checkpoints expecting native montages can degrade performance independently of model quality. The paper does not state whether each baseline's own preprocessing was used or whether validation-based hyperparameter selection was performed. If the baselines were effectively run with default settings on out-of-native-format inputs, the rank margin could partly reflect under-tuning rather than STEAM's superiority. The paper's internal ablations for STEAM are thorough, but they do not validate the external comparison. A secondary weakness is that the expert count (8) was selected using Figure 6 on four of the evaluation datasets, a test-set selection issue; however, the primary load-bearing concern is baseline evaluation fairness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STEAM, a dual-branch spatio-temporal encoder with a shared soft mixture-of-experts (SSMoE) module, frequency-aware gated attention, and a two-stage hierarchical pre-training strategy (Stage-I on TUH, Stage-II on paradigm-specific corpora). It claims the best average rank across seven downstream datasets and fourteen evaluation settings at competitive inference cost, and reports consistent gains from paradigm-specific specialization, alongside a lightweight adaptation method that updates about 5.1% of parameters. The empirical evaluation uses cross-subject LOSO and within-subject few-shot protocols, with internal ablations isolating the contributions of the architecture and pre-training objectives.","tokens_in":22538,"tokens_out":5365,"duration_ms":47277,"significance":"If the empirical claims are substantiated, STEAM is a meaningful contribution to EEG foundation models: it proposes a plausible mechanism for reconciling broad transfer with paradigm specialization, and the architecture details are presented with enough specificity to permit reimplementation. The paper's internal ablations are unusually thorough—Tables 2, 3, 11, and 12 isolate SSMoE, frequency gating, pre-training objectives, and fusion mechanisms, including a parameter-matched control for the shared expert pool—and the subject-level figures and explicit corpus partitioning strengthen the presentation. The main weakness is that the headline rank claim rests on a baseline-fairness assumption that is not yet documented, and the configuration of the model itself is selected on the same benchmark datasets, which undermines the comparison.","major_comments":[{"comment":"The headline claim of a 1.43 average rank versus a 10.55 mean for ten foundation-model baselines rests entirely on the fairness of the external comparison. The paper states that all baselines use 'their official implementations and publicly released pre-trained checkpoints,' but it does not report per-method preprocessing or hyperparameter tuning. All downstream data are mapped to a standardized 60-channel montage by inverse-distance interpolation (Appendix B), and these interpolated signals are then fed to models pretrained on their native channel layouts (e.g., LaBraM, SingLEM, BIOT). If the official checkpoints receive out-of-native-format inputs and default hyperparameters, the 9.12-position rank margin could partly reflect under-tuning. Please provide, for each baseline, the exact preprocessing pipeline used, the validation-based hyperparameter selection protocol, and the number of epochs/learning rates, or re-run the comparison with the native preprocessing of each baseline.","section":"Appendix C (Subject-Level Results / Aggregate-rank calculation), Table 13"},{"comment":"The default configuration of eight SSMoE experts was selected from a sensitivity sweep on the same downstream datasets that appear in the main benchmark (BNCI2015001, CHB-MIT, SEED, BNCI2014009). This constitutes selection on the evaluation set: the reported STEAM numbers in Table 1 are produced by the configuration that performs best on those datasets, while the ten baselines did not receive an equivalent configuration-selection step. This asymmetry can inflate the rank margin. Please either select the expert count on a held-out subset or show that the rank conclusion is stable across expert counts (e.g., that STEAM's combined rank remains below the baselines' mean for M = 2, 4, 8, and 16).","section":"Appendix C (SSMoE Expert-Count Sensitivity), Figure 6, Table 1"},{"comment":"The paper reports mean ± std over three runs but provides no significance tests. Several setting-level differences that contribute to the rank comparison are smaller than one standard deviation; for example, on BNCI2014009 LOSO, STEAM achieves 78.63±0.54 versus EEGNet's 78.39±0.44, and on SEED-VIG LOSO, STEAM's RMSE 0.2281±0.0036 is statistically indistinguishable from LaBraM's 0.2281±0.0035. The 'best average rank' claim should be supported by paired tests at the subject level (e.g., Wilcoxon signed-rank tests for LOSO and few-shot per dataset) or by reporting confidence intervals for the combined rank.","section":"Tables 1 and 13"},{"comment":"The manuscript provides architecture tables and hyperparameters but does not release code, pre-trained checkpoints, or a link to any repository. Given the complexity of the two-stage pre-training pipeline and the many baseline evaluations, independent verification is not currently possible. A public release of the code and the Stage-I/Stage-II checkpoints (or at least a detailed, step-by-step reproduction protocol) is necessary to support the empirical claims.","section":"Appendix A (Reproducibility)"}],"minor_comments":[{"comment":"The diversity loss L_div is referenced in the pre-training objective but never defined in the main text; please provide its exact form.","section":"Method, Eq. (15)"},{"comment":"Entries such as '–38.02 ±0.03' and '–60.83 ±0.30' appear to be rendering artifacts of missing values; please use a clear dash with a footnote explaining that the result is not applicable.","section":"Table 1"},{"comment":"The caption should describe the axes (presumably average rank and inference FLOPs) and explain the bubble-size parameter count scale more explicitly.","section":"Figure 1 caption"},{"comment":"The input duration is 0.80 s while the model input is 1.00 s; please state whether zero-padding or resampling is used to reconcile this mismatch.","section":"Appendix B, 'BNCI2014009'"},{"comment":"The statement that 'all models use full fine-tuning' should also explicitly state that the same full-fine-tuning protocol is used for the rank comparison in Tables 1 and 13, so that the reader does not conflate the lightweight adaptation results with the main benchmark.","section":"Experiment Settings"}],"recommendation":"major_revision","confidential_remarks":"The paper is a strong empirical study of an architecture, but the two central claims—best average rank and the benefit of hierarchical pre-training—need additional validation. The baseline fairness issue is the most serious: without per-method native preprocessing and tuning, the rank margin cannot be interpreted. The expert-count selection on the evaluation sets is a related concern that the authors can address by reporting sensitivity across counts. The internal ablations are excellent and suggest the core method is worth publishing once the external comparison is either fixed or appropriately qualified. I would recommend a major revision rather than rejection, because the methodological core appears sound and the missing evidence is obtainable with additional experiments and reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Best to think of this as a strong architecture paper with an ambitious benchmark claim. What's genuinely new is the specific integration: a dual-branch spatio-temporal encoder with a shared soft-MoE module that lets the two branches exchange information through a small set of slots, plus frequency-aware gating and a two-stage pre-training recipe. The authors do a good job of isolating the contribution of the shared routing: Table 12 compares against late fusion, cross-attention, MLP fusion, and a parameter-matched separate-Soft-MoE, and shared routing wins on both datasets. That is the most convincing evidence in the paper. The pre-training data is cleanly separated from the evaluation data, and the Stage-II gains on held-out BNCI2014001, BNCI2015001, and SEED are real.\n\nThe soft spot is the external benchmark. The headline 'best average rank' (1.43 vs 10.55) depends on how the ten foundation-model baselines were evaluated. The paper says official implementations and checkpoints were used, but doesn't report per-method preprocessing, tuning, or adaptation. All data is mapped to a standardized 60-channel montage before being fed to models trained on their native channel layouts. That can hurt baselines independent of their quality. The margin could shrink under a more careful per-method protocol. This is the load-bearing assumption, and the stress-test note is right to flag it. It doesn't invalidate the architecture results, but it does mean the 'best average rank' claim is scoped more narrowly than the abstract suggests.\n\nA second, smaller issue: the default 8-expert configuration was chosen from Figure 6, which uses four datasets that also appear in the main evaluation table. That's a mild test-set selection problem. It probably doesn't change the qualitative conclusions, but it should be acknowledged or the count should be validated on a separate split.\n\nAlso missing: significance tests across subjects, and code/checkpoints. Means over three runs without paired tests make it hard to judge whether the small margins on some datasets are real.\n\nOverall, this paper deserves a serious referee. The architecture is sensible, the ablations are thorough, and the two-stage pre-training story is useful for practitioners in the EEG foundation model community. The reviewer should push for baseline tuning details, code release, and significance tests. If those come through, the benchmark claim becomes credible.","headline":"Solid empirical architecture study with thorough ablations; the benchmark win is plausible but the headline rank margin depends on baseline fairness that isn't fully documented.","tokens_in":23088,"tokens_out":2493,"would_cite":true,"duration_ms":23759,"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":"STEAM claims that one EEG foundation model can both transfer broadly across paradigms and specialize cheaply: across seven datasets and fourteen settings it reports the best average rank among compared models.","keywords":["EEG decoding","foundation models","mixture-of-experts","spatio-temporal representation learning","pre-training","brain-computer interface","transfer learning","parameter-efficient fine-tuning"],"falsifier":"Re-run the fourteen-setting evaluation giving every baseline foundation model the same per-method validation tuning and the same adaptation budget on the standardized 60-channel preprocessing, then check whether STEAM's 1.43 mean rank and 9.12-position margin over the ten baseline foundation models persist; if the baselines close the gap once tuned equally, the best-rank claim would not survive.","tokens_in":22009,"feed_emoji":"🧠","tokens_out":6498,"duration_ms":53960,"temperature":0.7,"pith_summary":"The paper is trying to establish that broad EEG transferability and paradigm-specific accuracy do not have to be traded off against each other. It presents STEAM, a dual-branch spatio-temporal encoder in which a shared soft mixture-of-experts (SSMoE) module lets channel-centric and temporal-segment tokens exchange information, plus a two-stage general-to-paradigm pre-training scheme. On seven downstream datasets spanning six EEG applications, STEAM reports the best average rank among the compared specialist and foundation models at a competitive inference cost measured in FLOPs. Stage-II specialization is reported to give consistent paradigm-specific gains, and lightweight adaptation that updates about 5.1% of parameters stays competitive with full fine-tuning on most classification tasks.","feed_headline":"One EEG backbone tops seven decoding benchmarks","feed_subtitle":"A two-stage pre-training regime adds cheap paradigm-specific gains while fine-tuning only about 5% of the model.","key_machinery":"The central mechanism is the shared soft mixture-of-experts (SSMoE) module, a cross-branch communication interface placed at selected encoder layers: spatial and temporal tokens are concatenated, softly routed through a small set of expert slots (eight by default) using two softmax normalizations, and the transformed slots are redistributed back to each branch through token-specific combination weights. This lets the two streams exchange global information while preserving their separate local structures. Two supporting mechanisms carry the rest of the argument: frequency-aware gated attention, which computes log band-power descriptors per token and uses them to reweight attention outputs inside every layer, injecting a spectral prior useful under low signal-to-noise ratios; and the two-stage general-to-paradigm pre-training, which first optimizes masked reconstruction with cross-view contrastive alignment and then adds a supervised classification objective on multi-dataset paradigm corpora to specialize the backbone.","core_discovery":"The paper's central claim is that a single EEG foundation model can simultaneously achieve three things that prior models treat as a trade-off: broad cross-paradigm transfer, accurate paradigm-specific decoding, and cheap downstream adaptation. The proposed STEAM architecture keeps spatial and temporal representations in separate branches, couples them through a shared soft mixture-of-experts bottleneck with compact slots, and injects band-power-based frequency gates into every encoder layer. A two-stage pre-training regime first learns general EEG structure from a large unlabeled clinical corpus, then continues training on multiple datasets within a target paradigm with supervised objectives to produce a specialized variant. The paper reports that this recipe yields the best average rank across seven datasets and fourteen evaluation settings, with the specialized variants improving on the general model for motor imagery and emotion decoding.","pith_inferences":["The same general-to-paradigm recipe would likely carry over to other physiological signals with spatial and temporal tokenization, such as MEG or multimodal biosignals, though the paper does not test this.","The gains could partly come from the standardized 60-channel interpolation rather than the architecture; ablating montage normalization against native-channel adaptation would isolate the source.","Stage-II specialization helps only within the target paradigm, so deployment choices depend on the task: general Stage-I for broad transfer, specialized variants when the paradigm is known in advance.","The SSMoE visualizations hint at interpretable spatio-temporal explanations for predictions, but the paper offers only qualitative evidence; measuring slot selectivity quantitatively would test that promise."],"forward_implications":["A single pre-trained EEG backbone can be reused across tasks and subjects without per-task training from scratch, then specialized to a target paradigm by continued pre-training instead of a fresh model.","Adapting only the spatial and temporal embedding modules plus a task head, about 5.1% of parameters, can match full fine-tuning on most classification tasks and shrink optimizer-state storage.","Stage-I initialization yields stronger downstream accuracy throughout Stage-II training and reaches a given accuracy in about 2.8x fewer epochs, so paradigm-specific variants are cheaper to produce.","The SSMoE slots visually concentrate on distinct temporal intervals and channel subsets, indicating that the model learns decomposable spatio-temporal responses rather than one global routing pattern."],"supporting_citations":[{"why":"Supplies the TUH EEG corpus used for Stage-I general pre-training.","marker":"Obeid and Picone 2016"},{"why":"Provides the soft mixture-of-experts routing principle that the shared SSMoE module adapts.","marker":"Puigcerver et al. 2024"},{"why":"LaBraM is a general EEG foundation model baseline that STEAM must beat in the benchmark.","marker":"Jiang, Zhao, and Lu 2024"},{"why":"EEGPT is another general foundation-model baseline in the fourteen-setting comparison.","marker":"Wang et al. 2024"},{"why":"MIRepNet is the paradigm-specific motor-imagery baseline whose results STEAM-Spec is compared with.","marker":"Liu et al. 2026b"},{"why":"Provides BNCI2014001, one of the downstream motor-imagery evaluation datasets.","marker":"Tangermann et al. 2012"},{"why":"Provides the SEED emotion dataset used both as a downstream benchmark and in the Stage-II emotion corpus.","marker":"Zheng and Lu 2015"},{"why":"Motivates parameter-efficient transfer learning, the basis for the lightweight adaptation strategy.","marker":"Houlsby et al. 2019"}],"fun_headline_variants":["EEG foundation model tops seven benchmark datasets","One model, two-stage training, best EEG rank","Hierarchical pre-training boosts EEG decoding across paradigms","SSMoE-coupled encoder achieves top average rank","Unified EEG backbone delivers top rank in 14 settings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline rank result depends on the ten baseline foundation models having been evaluated as fairly and carefully as STEAM: the paper says they used official implementations and released checkpoints, but it does not report per-method hyperparameter tuning or adaptation to the standardized 60-channel inputs.","fun_headline_variants_meta":{"raw":{"variants":["EEG foundation model tops seven benchmark datasets","One model, two-stage training, best EEG rank","Hierarchical pre-training boosts EEG decoding across paradigms","SSMoE-coupled encoder achieves top average rank","Unified EEG backbone delivers top rank in 14 settings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000235,"raw_usage":{"total_tokens":1472,"prompt_tokens":892,"completion_tokens":580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":507}},"tokens_in":508,"tokens_out":580,"duration_ms":6514,"temperature":1.0,"reasoning_tokens":507,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:09:57.486761+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the fourteen-setting evaluation giving every baseline foundation model the same per-method validation tuning and the same adaptation budget on the standardized 60-channel preprocessing, then check whether STEAM's 1.43 mean rank and 9.12-position margin over the ten baseline foundation models persist; if the baselines close the gap once tuned equally, the best-rank claim would not survive.","supporting_citations":[],"review_version":2}