- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3(+1)
Read review
Paper summary
NeuroWeave is a masked EEG pre-training method. Its targets are discrete codes from separate band-specific VQ tokenizers (delta, theta, alpha). Each tokenizer gives code slots per one-second channel patch (Sections 3.2-3.4). A shared Transformer takes broadband masked patches and predicts all codes with a band- and slot-normalised cross-entropy (Eqs. 8-9). The tokenizers are frozen and dropped after pre-training. Section 4 is a staged study. It covers tokenizer reconstruction when codebook size or slot count changes (Fig. 2), masked-patch reconstruction against broadband VQ and waveform-regression controls (Fig. 3, Table 1), band selection on TUAB and TUEV (Table 2), and fine-tuning on six datasets against random-init models, BIOT, LaBraM, CBraMod, BrainOmni and three internal controls (Tables 3, 4, 8). The paper reports that the three-band model beats the controls on every metric, and it claims state-of-the-art results.
Critical strengths
-
The design is controlled and checkable. The paper separates which bands are supervised from how many slots each gets, and tests each with a named control: S1, RawVQ-S32 and RawRegression downstream (Tables 3, 4, 8), and RawVQ-5 and RawVQ-80 for reconstruction (Table 1). I checked the statement "lower NMAE than their matched broadband VQ controls in every dataset-band comparison" (L362-364) against Table 1, and it holds for all ten rows. This makes the central design claim testable, and it is the main reason I take the paper seriously.
-
Variability and limits are reported with care. Downstream results are five-seed mean ± std (L230-231), and Appendix E gives subject-level partitions and sample counts for TUAB and TUEV. The paper also says the slot sweep is not parameter- or compute-matched and uses one seed (L324-327), and that the external comparisons are descriptive (L428-431). A reader can judge how far the results go.
Critical weaknesses
-
Band selection uses the reported data, and the state-of-the-art claim is too broad. The delta-theta-alpha set is picked from Table 2 on TUAB and TUEV (L406-407, L420-421). The selected row of Table 2 (82.85 / 91.27 / 91.63 and 69.68 / 0.7039 / 85.28) is the same as the NeuroWeave (S8) row of Table 3. So the numbers that pick the configuration are the numbers reported for it. The paper does say TUAB and TUEV are selection tasks and the other four datasets are not, so the issue is limited to these two. The margins are mixed. On TUAB, the AUROC and PR-AUC gaps to the four- and five-band models are 0.6 to 1.1 points, bigger than the stds. For TUAB balanced accuracy (82.85±0.06 vs 82.36±0.51 and 82.48±0.59), they are about one std. On TUEV, the gaps to the four-band model are within one std, and the four-band model has the higher TUEV balanced accuracy (69.70±2.42 vs 69.68±1.70). Against the five-band model, the TUEV gap is large. The paper doesn't say which split Table 2 uses. Separately, the abstract (L026-028) and conclusion (L484-485) claim state-of-the-art, but L428-431 says the external comparisons are descriptive because corpora, architectures and adaptation pipelines differ. Several margins are inside one std: SHU-MI against BrainOmni (66.71±2.08 vs 65.98±1.17, Table 4), PhysioNet-MI against CBraMod (59.80±0.70 vs 59.32±1.38), and S8 against S1 on SHU-MI (66.71±2.08 vs 66.32±2.00). I couldn't find parameter counts, pre-training hours or compute for NeuroWeave or the baselines (Appendix D gives hours only for Sleep-EDF). If Table 2 is a test split, the authors could select on validation, report test only for the chosen set, and add a significance test for the band comparison. That would fix the first part. Narrowing the claim or adding matched-data comparisons would fix the second.
-
Neither the pre-training benefit nor the case for multi-slot targets is separated out. Tables 3, 4 and 8 have no randomly initialised NeuroWeave backbone. The random-init rows are other architectures. This matters, because on SSVEP the random-init ST-Transformer is close to the pre-trained models: 48.32±1.20 vs 49.09±1.37 on BETA-SSVEP and 75.90±1.88 vs 78.96±1.85 on Benchmark-SSVEP (Table 8). For slot count, the paper compares the sweep (, up to 12 bits per band) with the sweep (up to 128 slots of 10 bits). To call slot count "the dominant tokenizer-capacity axis" (L068-069), it needs equal-budget pairs, for example with against with . At , NMAE is still high for a band limited to 3 Hz (delta about 0.35 on TUAB, Fig. 2). Codebook usage isn't reported, so I can't tell whether the codebooks are used. Downstream, only and are tested, so the case for multi-slot targets rests on two points.
A few smaller points. Beta and gamma are dropped based on two clinical datasets, but the abstract says they "provide no consistent benefit" without that scope. No four-band variant is run on the motor-imagery or SSVEP sets. The "negative transfer" explanation (L405-406) is untested, since adding a band also adds 8 targets and changes the per-band weight in Eq. 9. Every band gets , but Fig. 3 shows a different best per band and Eq. 9 allows different . Table 7 gives one set of fine-tuning settings, and I couldn't find whether the baselines used it or were tuned (BrainOmni gets 15.00±7.66 on BETA-SSVEP, below ContraWR and ST-Transformer, Table 8). As far as I can tell, the five seeds are downstream only. Appendix A describes BandVQ (Sukhbaatar et al., 2026) as the closest work, since it also uses band-specific VQ tokenizers. From the paper's description, the stated difference is that factorization and per-band slot count are treated as separate choices, with several slots per band (L148-150, L624-626). No experiment compares with BandVQ, FAME, EEG-X or MANAS-2, and I couldn't check these 2026 preprints. Minor: Eq. 5 doesn't give the forms of and or the value of ; the beta and gamma filters are not in Section 3.2 or Table 5; the per-patch NMAE used in Fig. 3 and Table 1 is not written out; in Eq. 8 is not defined in the text; Appendix B (L646-647) mentions entries marked "–" that I did not find in Tables 5-7; Fig. 1 has "Mutli"; there is no limitations section.
Questions for authors
- Which split is Table 2 computed on, and what do the five band sets give on validation?
- What does the same backbone give on all six datasets without pre-training?
- In the sweep, what happens if is reduced so that total bits match the , case? And what is codebook utilisation at and ?
- Do the TUAR, TUEP, TUSZ and TUSL pre-training recordings share patients or sessions with the TUAB and TUEV test partitions? I couldn't find a statement that they don't.
- How do parameter count and pre-training hours compare with LaBraM, CBraMod and BrainOmni? Were the baselines fine-tuned with the Table 7 settings?
4 comments
Response to Reviewer Twyt
Q3. Equal-bit slot comparison and codebook use at large
A3. Appendix E.1 (Tables 10–11) compares TUAB validation global NMAE at bits. At 12 bits, all tested settings are:
| δ / θ / α / β / γ NMAE | |
|---|---|
| 4096; 1 | 0.51709 / 0.68806 / 0.71840 / 0.86350 / 0.85320 |
| 64; 2 | 0.57265 / 0.69782 / 0.71984 / 0.88286 / 0.90061 |
| 16; 3 | 0.62249 / 0.74608 / 0.74790 / 0.88975 / 0.92955 |
| 8; 4 | 0.64894 / 0.73719 / 0.75372 / 0.89241 / 0.94107 |
| 4; 6 | 0.67361 / 0.74607 / 0.78669 / 0.89807 / 0.93734 |
| 2; 12 | 0.72450 / 0.82032 / 0.84468 / 0.92108 / 0.95787 |
At 12 bits, one slot has the lowest error in every band. At 20 bits, the larger-codebook comparison is:
| δ | θ | α | β | γ | |
|---|---|---|---|---|---|
| 1,048,576; 1 | 0.47309 | 0.59896 | 0.63444 | 0.81440 | 0.84213 |
| 1024; 2 | 0.48392 | 0.54795 | 0.59979 | 0.82764 | 0.86869 |
The 20-bit result is mixed. Section 4.2 now limits the slot-count claim to its tested sweeps; Section 5 notes that downstream was evaluated only at 1 and 8. Appendix E.2 (Tables 12–13) reports active-code fraction across –128. For the requested large settings:
| Validation set | δ / θ / α / β / γ active codes (%) | |
|---|---|---|
| TUAB | 64 | 97.37 / 99.96 / 100 / 100 / 100 |
| TUAB | 128 | 97.36 / 99.98 / 100 / 100 / 100 |
| TUEV | 64 | 85.35 / 96.29 / 98.74 / 99.74 / 99.81 |
| TUEV | 128 | 88.20 / 94.72 / 97.77 / 99.68 / 99.59 |
This measure counts entries seen at least once per slot, averaged across slots. It shows no severe complete codebook collapse under this measure, while coverage varies by dataset, band, and .
Response to Reviewer Twyt
Q4. Patient/session overlap with TUAB and TUEV test partitions
A4. We exclude TUH-derived recordings matching the patient or recording session of a TUAB or TUEV evaluation example before building the pre-training corpus. The retained TUAR, TUEP, TUSZ, and TUSL recordings share neither patients nor sessions with the evaluation partitions. This is now explicit in Section 4.1 and Appendices G and H.
Q5. Model scale, pre-training duration, and baseline fine-tuning
A5. Appendix I (Table 16) reports the parameter counts of the evaluated implementations and their pre-training corpus durations:
| Model | Parameters | Corpus duration |
|---|---|---|
| NeuroWeave | 5.820M | 3,415.48 h EEG |
| LaBraM-Base | 5.820M | approximately 2,500 h EEG |
| CBraMod | 69.286M | over 9,000 h EEG |
| BrainOmni | 14.524M | 1,997 h EEG + 656 h MEG |
| EEG-X | 0.040M | approximately 4 h EEG (STEW) |
| MANAS-2 | 92.8M | approximately 60,000 h EEG |
NeuroWeave and LaBraM-Base use the same backbone and have the same parameter count; NeuroWeave is larger than EEG-X and smaller than BrainOmni, CBraMod, and MANAS-2. Corpus durations describe scale rather than a controlled data-efficiency comparison. Appendix B states that NeuroWeave and all baseline models use the downstream settings in Table 7 for each dataset.
Addressed weaknesses: performance claims and scope. Section 4.5 treats external comparisons descriptively; the Abstract qualifies the result as the highest reported mean among evaluated methods under our protocol. The Abstract and Section 4.4 restrict the β/γ finding to TUAB/TUEV validation. Section 5 states that other downstream slot counts and nonuniform per-band slot allocations were not tested, and clarifies that standard deviations cover five fine-tuning seeds from one pre-trained checkpoint.
Addressed weaknesses: omitted baselines. Appendix D (Table 9) adds BandVQ-style, MANAS-2, and EEG-X comparisons. Because BandVQ's code was unavailable, we independently implemented its paper-described target design within the NeuroWeave backbone and used the same downstream protocols; this is not an official-code result. MANAS-2 and EEG-X are separate model baselines. Balanced accuracy (%) is:
| Dataset | MANAS-2 | EEG-X | NeuroWeave (S8) |
|---|---|---|---|
| TUAB | 80.99±0.69 | 75.24±0.22 | 82.85±0.06 |
| TUEV | 65.57±2.66 | 38.91±5.62 | 69.68±1.70 |
| SHU-MI | 62.58±0.61 | 53.87±0.91 | 66.71±2.08 |
| PhysioNet-MI | 56.87±1.53 | 38.01±2.70 | 59.80±0.70 |
| BETA-SSVEP | 48.09±1.70 | 9.68±0.77 | 49.09±1.37 |
| Benchmark-SSVEP | 75.71±2.24 | 17.15±1.35 | 78.96±1.85 |
Appendix D explains that FAME pre-trained weights were unavailable at evaluation time, so we do not report a fine-tuning result.
Addressed weaknesses: method notation and metric. Section 3.2 defines the tokenizer loss terms and commitment weight in Eq. (5); Section 3.4 defines the prediction logits in Eq. (8). Sections 4.2–4.3 now use the same global NMAE of Eq. (10) for Figure 3 and Table 1. The earlier “median patch-level NMAE” wording was a writing error; no patch-wise median was used and the results are unchanged. Appendix B specifies the β and γ filter cutoffs. Section 5 supplies the requested limitations.
Thank you for the thorough response. Your response addresses my main soundness concerns, and I appreciate your candor about the original test-based selection. The equal-bit results are informative: at 12 bits, a single slot gives the lowest NMAE in every band, and at 20 bits, the picture is mixed. This suggests the benefit of multi-slot targets may come largely from total target capacity rather than slot count per se. Given that downstream K is tested only at 1 and 8, I would encourage the final version to frame the multi-slot contribution accordingly. I have raised my soundness score to 3. I keep my overall rating at weak acceptance, as the core design choices are now more honestly scoped, but their advantage over equal-budget alternatives is limited.
Response to Reviewer Twyt
Thank you for your precise questions and for identifying limitations in the original band-selection and capacity claims.
Q1. Table 2 split and validation results for five band sets
A1. The original Table 2 used test results to motivate band selection; this was inappropriate. We repeated the full comparison on TUAB and TUEV validation splits. Revised Section 4.4/Table 2 reports the five-seed means below (standard deviations are in the table); Table 3 is held-out test evaluation.
Validation again selects δ+θ+α on both datasets, but this does not excuse the original test-based choice. We make no significance claim for differences between band sets.
Q2. Same backbone without pre-training
A2. Appendix C (Table 8) compares the same backbone trained from random initialization with NeuroWeave (S=8) across all six datasets. Both settings use five downstream seeds and validation-selected checkpoints. Difference is the pre-trained mean minus the from-scratch mean, in each metric's displayed units.
The pre-trained model has a higher mean on every reported metric under this same-backbone control.