{"id":"8e26df08-abf6-4630-894c-893c572eed01","arxiv_id":"2607.27308","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ZUNA1.1, an open-source 380M EEG diffusion autoencoder, reconstructs variable-length, flexibly masked EEG at least as well as its predecessor and far better than spherical spline interpolation.","lead":"A 380M-parameter diffusion model called ZUNA1.1 reconstructs missing or noisy EEG data across variable-length windows up to 30 seconds, many channel layouts, and arbitrary time segments. It matches its predecessor ZUNA1 on standard 5-second tests while beating spherical-spline interpolation, and its weights and code are released open source.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Patch extrapolation: 'arbitrary scalp locations' is the load-bearing claim; no held-out montage is tested, and the paper's own Stage-3 layout-scheme plateau (III.D) suggests spatial extrapolation is the least-learned capability.","rationale":"The reader's weakest-assumption identification matches my own: the flexibility claims rest on extrapolation of the discretized 4D-RoPE and variable-length training to unseen montages and temporal patterns, and the paper does not test a held-out montage. My stress pass adds the internal evidence from Fig 4 and Section III.D that the layout schemes—exactly the ones exercising spatial extrapolation—plateaued and were dropped from Stage 3. This makes the missing test load-bearing rather than merely desirable. The measured reconstruction comparisons (four public datasets, error bars, spline baselines) are credible, and the released code/weights are a genuine strength, so the verdict should remain CONDITIONAL rather than REJECT. The proposed test is cheap, uses the released checkpoint, and would decisively determine whether the abstract's capability claim is supported.","tokens_in":15438,"tokens_out":14914,"duration_ms":122709,"concrete_test":"Run the released ZUNA1.1 checkpoint on a held-out montage: take BCI2000 (or ANPHY-Sleep) and mask all but a synthetic 12-channel layout whose (x,y,z) coordinates are deliberately chosen to avoid all Table I and standard 10-20 positions (e.g., a ring of 12 electrodes at 30° rotation, radius 0.85). Reconstruct to the full 256-channel montage and compute NMSE. Compare against (a) spherical-spline interpolation on the same mask and (b) ZUNA1.1's own NMSE on an in-distribution Muse-4 mask from the same dataset. Repeat for three random rotations of the held-out layout. Also test a 1s contiguous interior temporal gap in all channels of a 20s window. If the held-out-layout NMSE is within ~1.5x of the in-distribution NMSE and still beats spline, the extrapolation concern is resolved; if it degrades sharply, the abstract's 'arbitrary' claim should be scoped down.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central novelty is flexibility, and the abstract's strongest wording is 'arbitrary number of EEG channels at arbitrary scalp locations' and 'arbitrary temporal intervals.' The temporal flexibility is supported by variable-length training and Fig 7b, though that figure lacks error bars and a spline baseline. The spatial flexibility, however, is only evaluated on montages that are in-distribution: the four evaluation datasets' own montages and the consumer layouts in Table I, which were explicitly used during training (Stage 1-2). A genuinely novel sparse montage with electrode coordinates not matching any trained device is never tested. Because the positional encoding is a discretized 4D-RoPE with 100 bins/axis, an unseen layout may place channels in sparsely populated or empty bins; the model's extrapolation to such coordinates is unknown. The paper itself provides a warning sign: Fig 4 shows the layout schemes plateauing at a high loss, and Section III.D states they were 'substantially more challenging' and were dropped from Stage 3. Thus the most advertised capability—handling arbitrary scalp locations—is also the least measured and, by the authors' own curves, the least successfully learned. If ZUNA1.1 degrades to near-spline performance on a held-out montage, the abstract's central claim would be empirically false as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ZUNA1.1 is a 380M-parameter transformer-based diffusion autoencoder for EEG reconstruction. It extends the earlier ZUNA1 model with variable-length training (0.5–30 s windows), a mixture of eight channel-dropout schemes, quality-aware preprocessing with thresholds applied at load time, and a larger training corpus (≈3.5M channel-hours). The paper claims that ZUNA1.1 can reconstruct variable-length sequences with an arbitrary number of channels at arbitrary scalp locations and arbitrary temporal intervals, that it is at least on par with ZUNA1, and that it substantially outperforms spherical spline interpolation. Reconstruction is evaluated on four held-out datasets (ANPHY-Sleep, BerlinBCI, BCI2000, AAD) under channel dropout and region occlusion, plus a duration sweep and structural dropout schemes. The model and code are released open source.","tokens_in":15786,"tokens_out":4010,"duration_ms":34030,"significance":"If the flexibility claims are substantiated, ZUNA1.1 would be a practically useful contribution: a single open-source model for EEG denoising, super-resolution, and temporal infilling, trained at scale and evaluated against a standard baseline. The paper's strengths include a detailed training recipe, honest discussion of observed limitations, and comparisons against spherical spline across multiple datasets with error bars in several figures. However, the central novelty—spatial and temporal flexibility—is not fully validated by the experiments as presented. The evaluation of 'arbitrary scalp locations' is limited to montages that are in-distribution, the only variable-length result lacks error bars and a baseline, and the 'at least on par' claim holds only on average, not uniformly. These gaps are load-bearing because they concern the abstract's primary claims.","major_comments":[{"comment":"The claim of 'an arbitrary number of EEG channels at arbitrary scalp locations' (Abstract and §II) is not supported by the evaluation. All reconstruction experiments in §IV.A use the four datasets' own montages (Figs 5, 6) at fixed 5 s samples, which are explicitly at the heart of the training distribution. No held-out test reconstructs from a previously unseen sparse consumer montage to a high-density layout, and no test evaluates coordinates outside the 100-bin-per-axis 4D-RoPE discretization. The paper's own Fig 4 and §III.D show that the four layout dropout schemes plateaued at high loss and were dropped from Stage 3, indicating spatial extrapolation is the least-learned capability. Please provide an explicit held-out montage experiment (e.g., a consumer layout not used in training) or qualify the abstract and §II claims to match the evaluated regime.","section":"§II, §III.D, Fig 4"},{"comment":"The only evidence for variable-length capability is Fig 7b (NMSE vs sample duration at 15% token dropout). This panel has no error bars, no spline baseline, and the caption merely says 'Evaluation dataset' without identifying the dataset or sample size. Since variable-length reconstruction is a central contribution, this figure needs error bars, a named dataset, and ideally a comparison to spline or to ZUNA1 to establish that the model generalizes across durations. All other figures use 5 s samples, so the temporal-range claim rests entirely on this single panel. In addition, 'arbitrary temporal intervals within channels' is not directly tested: the structural schemes (Fig 8) drop channels or time blocks, but no experiment isolates infilling a specific temporal interval in a single channel with other channels intact.","section":"Fig 7b, §IV.A"},{"comment":"The abstract and §I state that ZUNA1.1 is 'at least on par with our earlier ZUNA1 model', but Table IV shows this is only true on average. ZUNA1.1 is worse than ZUNA1 on BerlinBCI at all four dropout levels (e.g., 50%: 0.737±0.022 vs 0.674±0.017; 90%: 1.486±0.069 vs 1.135±0.025) and on AAD at 20%, 50%, and 75% dropout. The paper acknowledges this in §IV.A ('on average but not uniformly'), but the abstract does not carry that qualification. Either reword the abstract to explicitly state 'on average' or report an appropriate aggregate statistical test (e.g., paired comparison across datasets/conditions) to justify unqualified parity.","section":"Table IV, §IV.A"}],"minor_comments":[{"comment":"The notation 'e2(ρc,s−3.5)' is ambiguous. Please use a clearer exponential form, e.g., exp(2(ρc,s − 3.5)), or define the function.","section":"§III.B, Eq. (1)"},{"comment":"The caption says 'Evaluation dataset' without naming it. Specify which dataset and how many samples/trials were used, and include error bars.","section":"Fig 7 caption"},{"comment":"The scheme named 'full-time-pt-random' in Table I appears as 'full-time-pt-random' in Fig 4; ensure consistent naming. Also, Fig 4's y-axis label 'NMSE' could benefit from a note on the evaluation set and the fixed 5 s window.","section":"Table I and Fig 4"},{"comment":"The sentence beginning 'Note that 5 second samples are also at the heart of ZUNA1.1’s training distribution' is relevant to the interpretation of all results. It should appear earlier, in the experimental setup, so that the in-distribution nature of the main evaluations is transparent.","section":"§IV.A"},{"comment":"The checkpoint labels 'No MMD-50K', 'MMD–50K', 'Hybrid–400K', and 'Hybrid–575K' are not defined in the main text or table captions. A sentence explaining these training stages would improve reproducibility.","section":"Appendix C, Tables V–VI"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern lands: the paper's most advertised capability (arbitrary scalp locations) is the least measured, and the authors' own Fig 4/§III.D suggest it is the least successfully learned. The manuscript is honest about limitations, and the core reconstruction results against spline are solid, but the abstract overclaims. A major revision that adds a held-out montage evaluation or tempers the claims would be appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the thing about ZUNA1.1: it's a genuine engineering effort that largely delivers on variable-length reconstruction and beats spherical spline, but the headline flexibility claim—arbitrary channels at arbitrary scalp locations—is not actually tested. The stress-test note is right: no held-out montage is evaluated, and the authors' own Figure 4 shows the layout-based dropout schemes plateauing before being dropped in Stage 3. That is internal evidence that spatial extrapolation is the least-learned capability.\n\nWhat's new and worth crediting: variable-length training from 0.5–30s, quality-aware preprocessing with load-time thresholds, two filter variants, the eight-scheme dropout curriculum, and a larger corpus (3.5M channel-hours) trained for 580k steps. These are pragmatic improvements that address real deployment pain points. The results on the four datasets with error bars are credible: ZUNA1.1 clearly outperforms spherical spline and roughly matches the predecessor ZUNA1, which is expected given the shared architecture. The region-occlusion plots are a useful stress test. Appendix C is honest about the downstream classification degradation—that tempers the 'foundation model' framing but not the repair/super-resolution claim.\n\nNow the soft spots, in proportion. The main burden falls on the 'arbitrary scalp locations' claim. The model uses a discretized 4D-RoPE with 100 bins per axis; training only ever saw fixed montages and the template layouts used in the Stage 1–2 dropout schemes. No held-out, genuinely novel sparse montage is reconstructed to a high-density layout. The authors' own curves show layout schemes saturating at high loss and being dropped from Stage 3. So the most advertised capability is the least measured, and the internal evidence suggests it is the weakest. Second, 'at least on par with ZUNA1' holds on average but not uniformly—Table IV has several cells where ZUNA1 is clearly better (e.g., AAD at 20% dropout: 0.764 vs. 0.890). That's minor wording, but it should be tightened. Minor: Fig 7b lacks error bars and a spherical-spline baseline, so the duration sweep is weaker than it could be. Also, the code release lacks commit-level reproducibility details; the weights are on Hugging Face and the package is pip-installable, so this is addressable, but worth noting.\n\nI'd send this to peer review. The engineering is solid, the comparison to spline is a fair baseline, and the honest reporting of the downstream decline is a plus. A serious referee would require either a held-out montage generalization test or a softened claim about arbitrary spatial layouts. With that revision, this becomes a useful contribution for EEG practitioners who care about artifact repair and super-resolution.","headline":"Solid engineering follow-up with a real generalization gap between the abstract's 'arbitrary scalp locations' and the evidence.","tokens_in":16263,"tokens_out":2028,"would_cite":true,"duration_ms":17942,"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":"ZUNA1.1 shows that a single diffusion autoencoder can repair and up-sample EEG of variable length (0.5–30 s), with any number of channels at arbitrary scalp locations and arbitrary temporal gaps, matching a fixed-window specialist and beati","keywords":["EEG foundation model","diffusion autoencoder","channel reconstruction","spherical spline interpolation","4D-RoPE","variable-length EEG","denoising","super-resolution"],"falsifier":"Take a consumer headset montage that was never in the training set (e.g., a 6-electrode layout with electrode positions lying outside the 100-bin discretization grid used in training), record real EEG, delete all channels, and ask the model to reconstruct a 256-channel high-density montage; if NMSE degrades sharply compared to reconstructions from a standard 10–20 montage, the arbitrary-location claim is falsified. A simpler check: feed the model a 30-second window longer than any it has seen in evaluation and see whether the error grows discontinuously at the training-length boundary.","tokens_in":15319,"feed_emoji":"🧠","tokens_out":4114,"duration_ms":31472,"temperature":0.7,"pith_summary":"The paper introduces ZUNA1.1, a 380M-parameter diffusion autoencoder that performs EEG reconstruction — denoising, channel infilling, and up-sampling — on variable-length recordings from 0.5 to 30 seconds, with arbitrary channel counts at arbitrary scalp locations. Its central claim is that a single model, trained on roughly 3.5 million channel-hours with a diverse mixture of dropout patterns, can replace task-specific pipelines: it matches the previous fixed-5-second ZUNA1 model on that model's home turf and clearly outperforms classical spherical-spline interpolation. The authors argue this makes the model directly useful for real-world EEG workflows, where recordings have variable duration, consumer headsets have sparse layouts, and electrodes fail in spatially or temporally correlated bursts.","feed_headline":"One EEG model repairs any channel layout, up to 30 s","feed_subtitle":"A 380M-parameter diffusion autoencoder matches a fixed-window specialist and beats spherical spline interpolation on corrupted EEG.","key_machinery":"The load-bearing mechanism is the discretized 4D-RoPE positional encoding: each channel token carries a 3D scalp coordinate (x, y, z) and a coarse time index t (0.125 s per token), so attention geometry is defined by physical electrode location and time rather than by a fixed montage or window size. On top of this, variable-length training packs randomly cropped windows (0.5–30 s) into batches with sample-aware attention masks via flex attention, and eight channel-dropout schemes (four structural, four layout-based, including consumer-headset montages) teach the model to infill diverse corruption patterns. Quality-aware loading and two precomputed filter variants provide implicit augmentatio","core_discovery":"ZUNA1.1 is a transformer-based encoder–decoder diffusion autoencoder for masked EEG reconstruction, trained with a rectified-flow loss. The model's defining capability is flexibility: because it encodes electrode positions with a discretized 4D rotary positional embedding over (x, y, z, t), it can in principle process arbitrary sequences up to 30 seconds, any number of channels at any scalp locations, and reconstruct arbitrary temporal intervals within channels as well as whole channels. The paper reports that ZUNA1.1 performs at least on par with its predecessor ZUNA1 on fixed-5-second reconstruction benchmarks, while substantially outperforming spherical spline interpolation, with the gap","pith_inferences":["If 4D-RoPE generalizes as claimed, the same architecture could be applied to other spatially embedded biosignal arrays (e.g., high-density EMG or ECoG grids) without modification, since the positional encoding is agnostic to signal type.","The paper's own ablations show reconstruction accuracy and downstream classification accuracy diverge as training proceeds, suggesting that a reconstruction objective alone may not yield the best latent representations; a hybrid objective may be needed.","The robustness of the flexibility claim is untested at the extremes: all evaluations use fixed 5-second windows on standard montages, so the true 'arbitrary channel layout' claim awaits a held-out test with a never-seen consumer device or an out-of-distribution scalp coordinate.","The open release of weights at 380M parameters makes it feasible to fine-tune or distill the model for real-time streaming applications, a direction not explored in the paper."],"forward_implications":["A single model can replace montage-specific pipelines: a consumer four-electrode headset could be up-sampled to a dense research montage without retraining.","Temporal gaps in a channel, not just whole-channel loss, are covered: the model can reconstruct short intervals inside a channel using surrounding time points and other channels.","Variable-length inference removes the need to chop EEG into fixed 5-second segments, avoiding boundary artifacts in real recordings.","Because the model beats spherical spline interpolation by a widening margin as dropout increases, it offers practical robustness for noisy or partially failing recordings.","The authors observe that the hardest layout schemes (e.g., reconstructing 256 channels from four electrodes) show little improvement over training, implying some information is irrecoverably lost by such aggressive down-sampling."],"fun_headline_variants":["One EEG model, any channels, any intervals up to 30s","Flexible EEG model beats spline interpolation on any layout","EEG foundation model: variable channels, variable time windows","Match specialist on EEG denoise, beat spline on any setup","Arbitrary EEG channels and intervals, one model repairs all"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central flexibility claims rest on the discretized 4D-RoPE positional encoding extrapolating to scalp positions and temporal window lengths that were not seen during training; if that extrapolation fails, the model is only as flexible as its training distribution.","fun_headline_variants_meta":{"raw":{"variants":["One EEG model, any channels, any intervals up to 30s","Flexible EEG model beats spline interpolation on any layout","EEG foundation model: variable channels, variable time windows","Match specialist on EEG denoise, beat spline on any setup","Arbitrary EEG channels and intervals, one model repairs all"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1122,"prompt_tokens":678,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":422,"completion_tokens_details":{"reasoning_tokens":357}},"tokens_in":422,"tokens_out":444,"duration_ms":4330,"temperature":1.0,"reasoning_tokens":357,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:46:57.046736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a consumer headset montage that was never in the training set (e.g., a 6-electrode layout with electrode positions lying outside the 100-bin discretization grid used in training), record real EEG, delete all channels, and ask the model to reconstruct a 256-channel high-density montage; if NMSE degrades sharply compared to reconstructions from a standard 10–20 montage, the arbitrary-location claim is falsified. A simpler check: feed the model a 30-second window longer than any it has seen in evaluation and see whether the error grows discontinuously at the training-length boundary.","supporting_citations":[],"review_version":1}