acceptodds
Under review as a conference paper at ICLR 2027

ZUNA: Flexible EEG Superresolution with Position-Aware Diffusion Autoencoders

Abstract

We introduce ZUNA, a 380M-parameter diffusion autoencoder with four-dimensional rotary positional embeddings for flexible EEG signal reconstruction and superresolution. We train two variants at scale: ZUNA1, which uses a typical, fixed dropout scheme, and ZUNA1.1, which introduces a a new set of masking strategies designed to reflect realistic EEG corruption and missing-data patterns beyond fixed patch reconstruction. ZUNA1.1 is capable of reconstructing variable length sequences of up to 30 seconds, with an arbitrary number of EEG channels at arbitrary scalp locations, and can reconstruct arbitrary temporal intervals within channels in addition to reconstructing entire channels. Among open-source models, ZUNA shows state-of-the-art reconstruction results across a range of dropout tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
infilling, super-resolution, EEG foundation model, diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tuNGOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a 380M parameter position-aware diffusion autoencoder, which utilizes large-scale heterogeneous EEG pre-training and multiple channel/time masking strategies to achieve channel completion, time interval completion, and spatial super-resolution. The author claims that the model can reconstruct EEG at any scalp position and in any number of channels.

Critical strengths

The paper is clearly organized and successfully builds a large-scale EEG reconstruction pre-training system. It also considers missing-data patterns that more closely reflect real-world scenarios compared to fixed random masks.

The zero-shot evaluation across different datasets and the dataset-level overlap checking are valuable engineering practices.

Critical weaknesses

The core contribution aligns more closely with large-scale system integration rather than a novel modeling mechanism. Diffusion autoencoders, RoPE, masked reconstruction, channel dropout, and large-scale pre-training all have mature precedents. The paper does not sufficiently articulate why this specific combination of components constitutes a novel general principle for EEG representation learning.

The claim of generalization to "arbitrary scalp locations" is overstated. The paper tests channel occlusion on unseen datasets or test samples, not on spatial coordinates that were entirely excluded during the training process. If standard 10-20/10-10 coordinates repeatedly appear in the pre-training corpus, the model is merely generalizing to new signals, not to novel physical locations.

Pre-training scale does not automatically equate to a universal interpolation capability. Unlike traditional interpolation techniques, ZUNA's output still depends heavily on preprocessing, coordinate normalization, reference schemes, and the distribution of the training corpus. The paper fails to demonstrate that a frozen model can stably replace spherical splines under conditions involving new hardware, novel montages, different reference schemes, and distributional shifts.

The provided details are insufficient to reproduce the core conclusions. Specifically, the training corpus, coordinate cleaning methods, channel mappings, mask generation, data leakage audits, and full evaluation codes are lacking. Furthermore, the description of how many dropout strategies were ultimately used needs to be unified between the main text and the supplementary materials.

Questions for authors

Please perform a rigorous coordinate-level permanent holdout: the held-out electrode locations must not appear in pre-training, fine-tuning, normalization statistics, model selection, or coordinate templates.

Can a completely frozen ZUNA stably outperform spherical splines on new montages, new reference schemes, and new devices that were entirely excluded from training? Please clearly separate the results for zero-shot, lightweight adaptation, and training from scratch.

Please release, or at least provide an auditable inventory of, the pre-training data, channel coordinate mappings, deduplication rules, mask generators, dataset holdout protocols, and model weight hashes.

Do the core conclusions of the paper still hold when compared against strong EEGSR and coordinate-queryable methods using the exact same input masks and evaluation protocols?

2 comments
AuthorsOpenReview ↗

Questions for the Reviewer

We thank the reviewer for their feedback and comments.

We agree that we do not articulate our novelty well enough, particularly the large-scale diffusion model aspect. Of the 14 methods that NeuralBench tests, none are diffusion models. Of the diffusion EEG models that exist, which are relatively few (none with open-source code), none exist at the training corpus size or parameter count of ZUNA.

On that note, concerning "Please release, or at least provide an auditable inventory of..." that we have already done parts this where possible, and where we haven't that the request is either unreasonable or simply ignores the materials already provided. In Appendix H, we release a list of IDs in our dataset audit, where we also explained our choice of holdout datasets being conditioned on our training corpus to prevent leakage. We do not have rights to release this data (even open-source datasets like TUH have distributional limits), and even if we did, our training dataset is 10s of TBs. Our code includes mechanisms to evaluate the model on an MNE dataset. Also, the model weight hash is already included in our ZUNA supplementary code. Section 3.4, Table 1 specifies all eight strategies, their probabilities, and the three-stage training curriculum. The mask generators are implemented in the supplementary eeg_data.py, beginning with perform_token_dropout().

On the note of "arbitrary scalp locations," there may be a significant semantic disagreement. Our use refers to the model's ability to accept user-specified electrode coordinates within its supported spatial discretization, rather than requiring a fixed output montage like the neural methods examined in the paper. Unlike architectures with learned embeddings tied to specific electrode identities, ZUNA conditions directly on spatial coordinates. Consequently, it can reconstruct channels at positions not measured in the input recording without architectural modification. Of course that will depend on the training data: the hope is that our cumulative coverage is enough to fill in the gaps, which we have shown in our main experiments. All of our experiments are done with frozen checkpoints, and Figure 4 already demonstrates that our model beats spherical-spline on novel montages. Specifically, in the case of high dropout regimes (particularly <75-90% missing channels) we have demonstrated generalization to an incredibly sparse montage that spherical-spline cannot recover.

We do not claim that our evaluations establish uniformly strong performance at coordinates permanently excluded from the entire pretraining corpus, and would be happy to soften the language here. But the criticism addresses a stronger question than the demonstrated ability to reconstruct missing channels across held-out datasets and heterogeneous montages: that is, combinations of frequently used montages. If this concern is strongly held enough by the reviewer, we will oblige with their request. However, at the moment, "Please perform a rigorous coordinate-level permanent holdout: the held-out electrode locations must not appear in pre-training, fine-tuning, normalization statistics, model selection, or coordinate templates." is not well-specified enough to run an experiment that would satisfy what the reviewer appears to want. In particular, it is unclear what constitutes a held-out coordinate: an exact 3D position, a discretized spatial bin, or an entire neighborhood within some specified distance of an electrode. This distinction is consequential because nearby but nonidentical electrode positions occur across montages, and ZUNA models spatial coordinates rather than discrete electrode identities. And RoPE works on relative distances between coordinates, not the coordinates themselves. The prohibition on using held-out coordinates in "coordinate templates" is ambiguous: a common coordinate and reference system is necessary to define electrode positions, and using that reference system does not itself constitute exposure to held-out EEG signals during training.

"ZUNA's output still depends heavily on preprocessing, coordinate normalization..." both is and is not a fair criticism. It is trivially correct that the model is dependent on the distribution of the training corpus. Indeed, the same thing could be said for almost any model. The cost of zero-shot spline interpolation is that, as we show in Figures 1, 3, and 4 that increasing the distance between missing electrodes results in heavy error for these sorts of methods. As for preprocessing, as examined in the paper, we trained on multiple variants to loosen preprocessing requirements which is rare in FM literature. Indeed, this is done for users who want to interpolate heavily bandpass-filtered data, and our least restrictive requirement (ie normalization) seems unfair to qualify as "heavy."

Thank you for the clarifications. I acknowledge that the evaluations use frozen checkpoints, and appreciate the specific pointers to the supplementary materials. While, my main remaining concern is still existing: since you proposed a super-resolution method, but why didn't you give a quantitative report with EEG super-resolution method?

Reviewer_AhtxOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ZUNA, a 380M-parameter diffusion autoencoder for EEG signal reconstruction and superresolution. The model incorporates 4D rotary positional embeddings to support flexible electrode configurations. Building upon ZUNA1, ZUNA1.1 introduces diverse masking strategies, quality-aware preprocessing, and variable-length training. Experiments on multiple EEG datasets demonstrate competitive reconstruction performance across different missing-data scenarios. The authors also investigate the model's downstream classification performance and observe a potential trade-off between reconstruction quality and learned representation quality.

Critical strengths
  1. The quality-aware preprocessing pipeline and variable-length sampling strategy allow more flexible use of heterogeneous EEG recordings. Combined with sequence packing and FlexAttention, these designs offer practical benefits for large-scale EEG training.

  2. The 4D RoPE design enables flexible modeling of different electrode configurations and provides a promising approach to hardware-agnostic EEG reconstruction.

  3. The diverse masking strategies cover various spatial, temporal, and spatiotemporal corruption patterns, making the training framework more relevant to realistic EEG signal reconstruction scenarios.

Critical weaknesses
  1. Insufficient ablation studies. The paper lacks controlled ablations for its key designs. Specifically, the effectiveness of 4D RoPE should be compared with conventional channel embeddings or alternative spatial encodings; the diverse masking strategies should be evaluated against a single masking strategy; and variable-length training should be compared with fixed-length training. Since ZUNA1 and ZUNA1.1 also differ substantially in training data, training steps, and other implementation details, the individual contributions of these designs remain unclear. Controlled experiments under comparable training budgets are needed.

  2. Limited evaluation of reconstruction quality. The evaluation primarily relies on NMSE and spectral bandpower errors. Additional metrics, such as Pearson correlation, would help assess waveform fidelity. More importantly, evaluating whether reconstructed EEG signals improve downstream task performance would provide stronger evidence of their practical utility.

  3. Insufficient validation of hardware-agnostic superresolution. A central contribution of the paper is its ability to reconstruct EEG signals across arbitrary electrode configurations. However, the results do not sufficiently demonstrate reliable reconstruction under sparse or unseen electrode layouts. In particular, Figure 7 (Appendix B) shows that the Consumer EEG masking scheme plateaus at an NMSE of approximately 1.1, indicating limited reconstruction effectiveness in this challenging setting. More direct evaluations using sparse configurations (e.g., 4, 8, 16, or 32 channels) to reconstruct higher-density EEG, along with experiments on unseen electrode layouts, would better substantiate this central claim.

  4. Unclear baseline evaluation protocols. In Table 6, models such as CBraMod and LUNA use different patch sizes and reconstruction mechanisms. Further clarification is needed regarding masking alignment and preprocessing consistency to establish a fair comparison.

Questions for authors
  1. Could the authors provide controlled ablation results to clarify the individual contributions of 4D RoPE, diverse masking strategies, and variable-length training? In particular, how do these designs compare with conventional channel embeddings, single masking strategies, and fixed-length training under comparable data and computational budgets?

  2. Figure 7 shows that the Consumer EEG masking scheme plateaus at an NMSE of approximately 1.1. Given the emphasis on hardware-agnostic superresolution, how well does ZUNA1.1 perform when reconstructing high-density EEG from sparse electrode configurations (e.g., 4, 8, 16, or 32 channels)? Are there quantitative results for previously unseen electrode layouts?

  3. Could the authors provide additional evidence of reconstruction fidelity beyond NMSE and spectral bandpower errors, such as Pearson correlation? More importantly, does reconstructing missing EEG signals improve downstream task performance compared with conventional interpolation methods?

  4. Could the authors clarify how masking patterns and preprocessing are aligned across different reconstruction baselines in Table 6, particularly CBraMod and LUNA, which use different patch sizes and reconstruction mechanisms? How is comparability ensured across these models?

2 comments
AuthorsOpenReview ↗

Thanks + followup

We thank the reviewer for their feedback and comments. Indeed, they rightfully hit upon a fundamental weakness of the paper, which is our insufficient reported ablation studies. We ourselves believe this to be our biggest limitation at a scientific level and would not object to rejection upon this basis if it is the principal concern. I would, however, like to appeal to the reviewer's interpretation of our contributions.

As one reviewer notes, we are "more closely with large-scale system integration rather than a novel modeling mechanism." Indeed, ZUNA relies on well-established sequence modeling techniques, but the scale, scope, and focus of our model is unprecedented. Unlike a foundation model like REVE, which is architecturally similar, we have taken reconstruction to its limit. The central finding is that doing this is antithetical to downstream representation learning, which is significant because the overwhelming majority of EEG FMs are masked pretrained autoencoders. Given that we have shown superior reconstruction to models also trained/evaluated for reconstruction (ie our large scale system works very well), we ask our scientific contributions to be based on the representational claim rather than "4D RoPE/this particular masking strategy/etc" is better. The shift from ZUNA1 to ZUNA1.1 is primarily about making the model more flexible for researchers, not claiming that specific changes are universally better. As our results show, ZUNA1 can indeed outperform ZUNA1.1 in specific cases despite having a smaller training corpus and less diverse masking scheme, but ZUNA1.1 on average can handle far more realistic scenarios where someone would want to use our model for superresolution. That is an engineering set of constraints, not a scientific claim. In particular, "diverse masking strategies, and variable-length training" are particularly hard to disentangle due to the scale of our training data and the time required to train a model like ZUNA on 10's of TBs of data.

On the other weaknesses: (2) Yes of course, correlation is already part of the evaluation harness and is easily added. However, "evaluating whether reconstructed EEG signals improve downstream task performance ...", we have already done this. Indeed, one of the central findings of the paper is that improving reconstruction seems to hurt downstream performance.

(3) "More direct evaluations using sparse configurations (e.g., 4, 8, 16, or 32 channels) to reconstruct higher-density EEG, along with experiments on unseen electrode layouts, would better substantiate this central claim." I do not believe the author has noted the significance of Figure 4 in this respect. Indeed the case of high dropout (ie 75% or 90%) is a sparse configuration. For example, a 32-channel like BCI2000 each model sees just 8 or 3 channels. That is also a novel, if highly nonstandard layout. We have indeed tried, as discussed in the paper and shown in Figure 7 of Appendix B, that our consumer EEG montage training scheme is the worst performing dropout scheme. Muse 2, for example, only has 4 electrodes on the forehead and behind the ears. Indeed, in Figure 4 as well, the methods sharply increase at these sparse layouts, which would correspond to merely a single channel for Muse. Our usage of heterogeneity here refers to combinations of the most frequent layouts that appear in our training data. ZUNA is not restricted to one particular backbone of channels, but that does not mean it can interpolate channels at arbitrarily large distances away from a small subset. Indeed, those maximum allowed distances seem to be larger than spherical spline and ST-EEGFormer, but there is a hard signal processing bound on this that we certainly do not claim to cross

(4) Specifically, for "In Table 6, models such as CBraMod and LUNA use different patch sizes and reconstruction mechanisms"...the patch sizes must be different because the native temporal sampling rates are different for each model. However, those patch sizes still correspond to the same physical time intervals as noted in Figure 5. That is, 50% of the same shared time is dropped across all models. Given that models have different token lengths, this can correspond to more or less tokens per model, but 50% of time is dropped in any case. We have given a preprocessing protocol in Appendix E.2 if the author has more specific questions here.

In sum, I believe the author has noted some valuable changes. Appropriate camera-ready improvements will be the inclusion of the Pearson correlation, which is an easy enough addition. If insisted upon, we can include a RoPE ablation. Our proposed baseline is an axial sinusoidal embedding: like our RoPE, we split the embedding 4 ways across x,y,z, and t, each part of the embedding gets its own set of dimensions. We would appreciate the reviewer's feedback on whether these clarifications address their primary concerns, and would be happy to discuss any remaining questions.

Thank you for the detailed response and clarification of the paper's engineering contributions. While I acknowledge the practical value of the proposed system, the lack of controlled ablation studies and additional experimental evidence leaves my main concerns unresolved. Therefore, I will maintain my original rating.

Reviewer_jXwSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ZUNA, a 380M-parameter diffusion autoencoder for EEG reconstruction and superresolution. ZUNA represents channel–time patches with four-dimensional rotary positional encodings over electrode coordinates and time, allowing the model to process variable montages, arbitrary channel subsets, and sequences of up to 30 seconds. Evaluation on four held-out datasets shows strong random-patch and region-occlusion reconstruction results, while a separate analysis finds that continued optimization for reconstruction degrades the usefulness of the learned representations for downstream classification.

Critical strengths
  1. The paper addresses an important and practically relevant problem. A single model that can accommodate heterogeneous montages, arbitrary electrode locations, full-channel loss, and temporally localized corruption would be valuable for both research-grade and consumer EEG. The use of spatial coordinates rather than a fixed channel vocabulary is well aligned with this objective.

  2. The data and training pipeline are substantial. It is trained on approximately millions of channel-hours. The paper also reports the principal architectural and optimization settings in useful detail.

  3. The evaluation is broad and includes informative negative evidence. The authors report both waveform NMSE and spectral-bandpower error, evaluate several masking regimes, and compare against neural and classical interpolation baselines. The analysis of the tension between reconstruction quality and downstream representation quality is candid and potentially useful to the EEG foundation-model community.

Critical weaknesses
  1. The contribution of the proposed components is not established by controlled ablations. ZUNA1 and ZUNA1.1 differ simultaneously in training-corpus size, number of optimization steps, preprocessing, sequence-length distribution, masking curriculum, spatial discretization, normalization, and loss design. Consequently, their comparison cannot identify the effect of the diverse masking mixture, 4D RoPE, or any other individual design choice.

  2. The experiments evaluate synthetic missingness rather than realistic EEG corruption. Electrode detachment, motion artifacts, cross-channel contamination, and hardware noise are not generally equivalent to deleting a known set of clean samples. A generative model may also produce spectrally plausible signals while altering event timing or clinically relevant waveforms. The paper does not evaluate naturally corrupted recordings, an end-to-end corruption-detection-and-repair pipeline, or whether reconstruction improves downstream sleep, BCI, or clinical-event decisions; the claims about in-the-wild denoising and repair are therefore stronger than the evidence supports.

  3. The empirical advantage is not uniform, and the baseline comparison mixes substantial data and adaptation differences. ZUNA1.1 uses far more training data and computation than most baselines, whereas the external models are evaluated as frozen public checkpoints under model-specific preprocessing. Moreover, Table 7 shows that ZUNA1.1 is clearly worse than both ZUNA1 and ST-EEGFormer on BerlinBCI at 75% and 90% channel dropout, and it is also not best in several AAD and ANPHY-Sleep settings.

  4. Reliability and deployment cost are insufficiently characterized. Inference uses a 380M-parameter model and 50 Euler steps, but the paper reports neither latency, memory, throughput, nor an accuracy–cost comparison with lightweight interpolation methods.

Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Following up + thanks for the read

Many thanks to the reviewer for taking the time to read and thoughtfully engage with our paper. As with another reviewer, they hit upon a fundamental weakness of the paper, which is our insufficient reported ablation studies. This is likely our biggest limitation at a scientific level and would not object to rejection upon this basis if it is the principal concern. I would, however, like to appeal to the reviewer's interpretation of our contributions: first as an engineering system, next our result that pretrained masked autoencoders (MAEs) can be great at their objective but bad at crafting generalizable features for downstream tasks (an important finding given that most EEG FMs are MAEs).

Specifically, on weaknesses (1) and (3), these are interrelated. The shift from ZUNA1 to ZUNA1.1 is primarily about making the model more flexible for researchers, not claiming that specific changes are universally better. As one reviewer notes, we are "more closely with large-scale system integration rather than a novel modeling mechanism." As our results show, ZUNA1 can indeed outperform ZUNA1.1 in specific cases despite having a smaller training corpus and less diverse masking scheme, but ZUNA1.1 on average can handle a far wider set of data and scenarios than ZUNA1, or any other model here (except perhaps linear interpolation, which as shown has its own set of difficulties). Indeed, ZUNA1 is specifically trained for whole-channel dropout on 5-second samples, ie the scenario exactly represented by Figure 4. However, ZUNA1 degrades in performance on samples outside of a small neighborhood around 5 seconds (plus or minus 1/2 second, roughly), whereas for ZUNA1.1 in Figure 8 we see that performance across samples from 0 to 30 seconds is fairly stable across the majority of dropout schemes. Most of the changes from ZUNA1 to ZUNA1.1 are added engineering constraints, not scientific claims. In addition, the set of scenarios covered by ZUNA1.1 is far wider than any previous neural FM. The superresolution literature is at a more basic state than implied by this criticism.

On the note of "ZUNA1.1 uses far more training data and computation..." we note that more training data and computation are not merely "cheating" but substantive improvements in scaling if increasing those also increases performance. If ZUNA achieves better performance averaged across many different tasks as a result of greater scale, that should be to its credit, not its weakness. Some models have to be excluded from certain tasks because they can't even handle them. ZUNA is also frozen with its own set of preprocessing requirements (all models receive their native preprocessing), and all models see the same physical time intervals in masked reconstruction. On average, it outperforms its competitors for each of the tasks presented here even if there is some variance in individual datapoints.

On weakness (2), I agree that the current language in the introduction is too strong and implies too much about our results. For example, mentioning "motion artifacts can disrupt entire groups of channels at a time" can imply that we can automatically denoise blinking/head movements/etc. without user specification of those intervals. We have removed such examples in the paper. However, our dropout schemes already cover a much wider range of denoising scenarios than any previous superresolution model, and those do correspond to meaningful tasks. As seen in Figure 4, ZUNA can work in highly sparse-channel regimes. It can remove entire channels or individual segments down to a resolution of 0.125s (in Figure 8, from 0-30s performance is incredibly stable across the majority of the dropout regimes). For dead or detached electrodes, even if our quality-aware preprocessing did not filter this, ZUNA can interpolate whole channels unlike CBraMod or LUNA. While linear interpolation can also do that, it is clearly dependent on the distance between electrodes and degrades much faster in performance. Lastly, for "A generative model may also produce spectrally plausible signals while altering event timing or clinically relevant waveforms." this is why we evaluated both spectral and NMSE, as timing should shown up in the waveform amplitude.

On weakness (4), this not a central focus/claim of the paper. ZUNA inference runs comfortably on consumer-grade GPUs like a 4090 and even has CPU/Mac Metal support. Most of the models here, anecdotally, do seem to run faster than ZUNA, but none of our data collection is processed online. Thus latency is not a major issue for us: even if linear interpolation can run at 1ms versus 1s for ZUNA, yes that is 2 OOM slower but not of practical concern to a researcher analyzing data offline. A few seconds versus a minute for a few hundred samples is not a big enough of a difference for us to have explored this tradeoff between accuracy and cost. 380M is large for an EEG FM, but not for modern hardware

Thanks for the response and clarifications. However, my main concern about the lack of controlled ablation studies remains unresolved, so I will keep my original rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.