acceptodds
Under review as a conference paper at ICLR 2027

SCOPE: Structured Prototype-Guided Adaptation for EEG Foundation Models with Limited Labels

Abstract

Electroencephalography foundation models (EFMs) have shown strong potential for transferable representation learning, but their real-world adaptation remains challenging when only a few labeled subjects are available. We trace this challenge to a structural mismatch between noisy, limited supervision and the highly plastic parameter space of EFMs, reflected in three key failure modes: overconfident miscalibration, transient optimization collapse, and disruption of pretrained representations under unconstrained updates. To address this mismatch, we propose SCOPE, a Structured COnfidence-aware Prototype-guided framework for label-limited EFM adaptation. SCOPE first constructs cohort-level external supervision to provide persistent guidance and further derives confidence-aware pseudo-labels to select reliable unlabeled samples. Building on this supervision, SCOPE introduces ProAdapter, a lightweight prototype-conditioned adapter that modulates frozen EFMs to preserve pretrained representations. We comprehensively evaluate SCOPE across 60 adaptation settings spanning 6 tasks, 5 EFM backbones, and 5%-100% labeled subject ratios. In label-limited settings, SCOPE outperforms full fine-tuning by 4.6 multiclass Kappa and 4.4 binary PRAUC points on average.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG foundation model, Label-efficient adaptation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3KrqOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SCOPE: prototype-guided adaptation for EEG foundation models (EFMs) under limited labeled data. This work highlights issues in adapting EFMs: overconfident miscalibration, training collapse, and strong representation drift. SCOPE conjectures that these issues stem from limited labeled data and the full fine-tuning approach, which motivates them to 1) build a pipeline to create pseudo-labels for unlabeled data via limited labeled data, and 2) use these combined cohorts to selectively fine-tune the EEG foundation model. SCOPE shows superior performance to other adaptation approaches across several EFMs and datasets.

Critical strengths
  1. Well-motivated idea of generating “pseudo” label data to adapt EEG foundation models, despite quality concerns

  2. Interesting idea of limited adaptation with confidence metrics

Critical weaknesses

Concerns about the quality of pseudo-labels: SCOPE does in-loop label creation: trains a TPN classifier with a labeled dataset; uses TPN to create a prototype set of unlabeled data; updates the prototype set using the labeled dataset; then uses the prototype most similar to the TPN embedding. This creates a loop and over-relies on TFN prediction. If EFM is not performant for cross-subject prediction, what makes TFN predictions more reliable? In the paper, TFN predictions are not shown to be reliable (Ablation on the training strategy shows that training should anchor to true labeled data; Tab. 27: performance drops as we use more pseudo data). The authors need to work further on this, as it is the main contribution of the work.

Minor issue: The main paper lacks explanations of the scale-shift modulation design. The authors explain why limited adaptation is necessary but do not provide substantial reasons for their modulation design choice.

Questions for authors

Weaknesses above

  1. Accuracy of predicting pseudo-labels
  2. How does the number of prototypes affect performance? In Fig. 10, performance varies with the number of prototypes without a clear pattern; could the authors further investigate this? Does the current design need additional propulsive force to push in-class prototypes apart, which is missing from Eq. 5 and Eq. 6?
  3. I do not see a clear functional separation in the complex in-loop label creation. Further explanations could be useful to make it clear
6 comments
AuthorsOpenReview ↗

Response to Reviewer 3Krq [1/n]

We sincerely thank you for recognizing our study’s well-motivated approach and interesting ideas.

TL;DR

① Pseudo-labels (Weakness1 & Question1 & Question3; see our Responses 1-4)

We would like to distinguish two challenges in pseudo-labeling:

  1. Pseudo-label quality: How to construct reliable pseudo-labels, directly measured by pseudo-label prediction performance.
  2. Pseudo-label utilization: How to effectively leverage inherently imperfect pseudo-labels, evaluated through downstream performance.

With this distinction, we would like to emphasize two important clarifications:

  1. Pseudo-label construction: SCOPE does not directly adopt TPN predictions as pseudo-labels, but further refines and selects them through prototype-based evidence and confidence-aware filtering.
  2. Interpretation of ablations: Our ablations on pseudo-label ratios and semi-supervised training strategies investigate how to better utilize imperfect pseudo-labels.

To provide a more structured explanation, we organize our responses as follows:

ResponseWhat We AddressYour Concern
Response 1Directly evaluate pseudo-label accuracyQuestion 1
Response 2Explain the roles of individual components in pseudo-label constructionQuestion 3
Response 3Quantify the stepwise improvements in pseudo-label qualityWeakness 1
Response 4Explain design choices for utilizing imperfect pseudo-labelsWeakness 1

② Intra-class prototype design (Question2 ; see our Response 5)

We provide further analyses of prototype numbers and additional ablations, with the following key findings:

  • Too few prototypes cannot adequately capture intra-class variability, while too many introduce instability under limited labeled data.
  • Both class-specific prototype allocation and explicit intra-class repulsion degrade performance, so we remain our current design.

③ Scale-and-shift modulation (Weakness2; see our Response 6)

We added the rationale for using feature-wise modulation to incorporate prototype-derived structural supervision.


We provide further details and address each of your concerns point by point below.

AuthorsOpenReview ↗

Response to Reviewer 3Krq [2/n]

1) Q1 Pseudo-label accuracy

In the original manuscript, we reported pseudo-label accuracy in Table 39 of Appendix I, where ρ (confidence) = 0.5 corresponds to the accuracy of the filtered pseudo-labels under our main experimental setting (30% labeled training data).

We additionally report pseudo-label accuracy on ISRUC with 20%, 30%, and 50% labeled training data to further evaluate its reliability across different label ratios.

Labeled Training Data RatioKappaWeighted-F1
20%66.0074.02
30%72.1579.25
50%87.5190.77

These results show the quality of our pseudo-labels improves as more labeled training data becomes available.


2) Q3 Functional separation in pseudo-label generation and selection

Pseudo-labeling pipeline\color{blue}{\textbf{Pseudo-labeling pipeline}}

We provide a more detailed functional separation about the pseudo-label generation pipeline.

  1. Progressive construction of cohort-level structured prototypes and two complementary predictions.
    • Task-prior learning: The TPN, regularized by ETF geometry, provides a global class prior and cluster-friendly embeddings, producing task-prior pseudo-labels.
    • Prototype construction: Building upon the TPN embeddings, we further construct prototypes through Sinkhorn-constrained clustering and refine them with labeled supervision to capture local intra-class structure, producing complementary prototype-based pseudo-labels.
  2. Confidence-aware fusion and selection.
    • Rather than directly using either prediction, we fuse their class probabilities to emphasize agreement. Pseudo-labels are retained only when the two predictions agree and the fused confidence exceeds a predefined threshold.

Key clarifications\color{blue}{\textbf{Key clarifications}}

We would like to note that your description in the Weakness section accurately captures the first stage of our pipeline. Importantly, the second stage, confidence-aware fusion and selection, is also important for improving pseudo-label reliability.

Revisions made\color{blue}{\textbf{Revisions made}}

Following your suggestion, we have added the following functional separation at the beginning of the External Structured Supervision Construction subsection in the Method section (lines 146-158):

To provide EFMs with reliable supervision, we adopt a two-stage pipeline consisting of three steps:

Stage 1: Cohort-level structured prototypes construction.

  • Task-prior learning. Train a task-prior network (TPN) with inter-class geometric regularization to obtain cluster-friendly embeddings and task-prior pseudo-labels.
  • Prototype construction. Cluster these embeddings and refine the resulting prototypes to obtain complementary prototype-based pseudo-labels.

Stage 2: Confidence-aware pseudo-label selection.

  • Fusion and selection. Combine the two class-probability distributions, and retain pseudo-labels only when both predictions agree and the fused entropy-based confidence exceeds a threshold.

Stage 1 is detailed in Section 3.2.1, and Stage 2 in Section 3.2.2.

AuthorsOpenReview ↗

Response to Reviewer 3Krq [3/n]

3) W1.1 Reliability of external supervision

Key clarification\color{blue}{\textbf{Key clarification}}

SCOPE does not directly rely on TPN predictions as reliable pseudo-labels. TPN provides an initial task-specific prior, which is further complemented by two levels of geometric regularization (ETF and Sinkhorn) and confidence-aware selection to construct reliable pseudo-labels. We provide quantitative analyses below to demonstrate the effectiveness of these strategies.

Empirical evidence for improved pseudo-label quality through geometric regularization and confidence-aware selection\color{blue}{\textbf{Empirical evidence for improved pseudo-label quality through geometric regularization and confidence-aware selection}}

We analyze pseudo-label quality on ISRUC with 30% labeled training subjects to demonstrate how our three strategies improve supervision reliability beyond the initial TPN predictions.

StrategyPrimary Function
1. ETF regularizationImproves inter-class separability
2. Sinkhorn-balanced clusteringImproves pseudo-label quality for minority classes
3. Confidence-aware fusion and selectionSelects more reliable pseudo-labels
  1. ETF regularization improves inter-class separability. This improves task prior pseudo-label quality from 46.36 to 51.60 in Kappa and from 55.74 to 60.89 in weighted-F1.

  2. Sinkhorn-balanced clustering preserves minority-class representation. Appendix I (Tables 39–40) validates this effect on the imbalanced ISRUC dataset. Without Sinkhorn, the minority N1 class nearly collapses (F1 = 2.09 at ρ = 0), whereas Sinkhorn substantially improves its representation (F1 = 41.57 at ρ = 0).

  3. Confidence-aware fusion and selection improve pseudo-label reliability. We retain pseudo-labels only when TPN and prototype predictions agree, achieving 62.49 Kappa and 70.67 weighted-F1 before confidence thresholding (ρ = 0). Further filtering at ρ = 0.5 improves these metrics to 72.15 Kappa and 79.25 weighted-F1 (Appendix I, Table 39).

    To further evaluate the effectiveness of our fusion strategy, we compare our multiplicative fusion with two simpler alternatives: average fusion and element-wise maximum fusion. Our multiplicative fusion emphasizes joint support from both prediction sources and penalizes conflicting evidence. We conduct this comparison on ISRUC with 30% labeled training subjects, using the same confidence threshold (ρ = 0.5), and evaluate downstream performance across five EFM backbones:

    AverageMaxOurs
    BackboneKappaWeighted-F1KappaWeighted-F1KappaWeighted-F1
    pseudo-label accuracy68.6674.2369.8975.9172.1579.25
    LaBraM53.83±0.8563.71±1.2353.51±1.8563.73±1.3356.19±0.6765.15±0.47
    CBraMod62.33±1.7868.52±1.3461.86±1.1468.12±1.2263.29±1.1270.74±0.64
    CSBrain60.81±1.0967.93±0.9359.90±1.5668.78±1.8561.51±0.5569.86±0.69
    EEGMamba61.17±1.6570.21±1.7061.92±1.7969.80±1.5263.07±0.8971.41±0.79
    CodeBrain62.30±1.2469.34±0.8261.29±1.5569.81±1.4763.46±0.9871.11±0.74

Additional clarification on EFM confidence reliability\color{blue}{\textbf{Additional clarification on EFM confidence reliability}}

External supervision addresses unreliable EFM confidence, rather than inaccurate EFM predictions. The primary motivation for introducing the external supervision construction module is that EFM confidence may not reliably reflect prediction correctness, making direct self-training prone to error amplification.

  • Evidence of unreliable EFM self-training:
    • Figure 2(A) reveals overconfidence in EFM predictions
    • Appendix C.2 shows direct EFM self-training yields inconsistent improvements over full fine-tuning
    • Table 1 shows unstable performance of FixMatch and FineSSL, two general-purpose self-training baselines
  • Reliability of our constructed supervision: As shown in Table 39 (Appendix I), increasing the confidence threshold (ρ) improves the quality of retained pseudo-labels. For example:
    • ρ = 0: Kappa = 62.49, weighted-F1 = 70.67
    • ρ = 0.5: Kappa = 72.15, weighted-F1 = 79.25
    • ρ = 0.8: Kappa = 82.93, weighted-F1 = 88.51
Reviewer_MjxqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose SCOPE, a framework for EEG foundation models adaptation when only a fraction of the training subjects is labeled. To this end, a convolutional network trained on the labeled subjects is used to build class prototypes and to assign confidence-aware pseudo-labels to the unlabeled subjects, and the frozen foundation model is adapted through lightweight adapters in its last layers that are conditioned on prototype similarity. The method is evaluated with five backbones on six tasks, and with labeled-subject ratios ranging from 5% to 100%.

Critical strengths

First, the evaluation is comprehensive and consistent, and hence, the main claim, namely that SCOPE is more stable than generic parameter-efficient and semi-supervised methods when labels are limited is considered as credible.

Second, the method is efficient since it trains only a small fraction of the parameters and requires substantially less memory than full fine-tuning. In addition, the comparison with direct self-training suggests that externally constructed pseudo-labels are more reliable than those derived from the model's own confidence.

Critical weaknesses

First, it is unclear where the improvement actually comes from. The prototypes and pseudo-labels come from a separate network, and its prototype similarities are fed into the adapters, so the final predictor effectively combines two models. It is also not explained how inference is done for the test subjects. The EEGNet baseline in Table 1, which is architecturally close to this separate network, performs poorly on its own, but a weak and a strong model can still complement each other. Without a late-fusion baseline, or results with the same pseudo-labels used with LoRA or full fine-tuning, I find it hard to tell whether the gain comes from the adaptation method itself or simply from ensembling and distillation.

Second, the statistical evidence is rather thin, and some results are questionable. Several test splits contain only two to five subjects (three for SEED, two for CHB-MIT), the reported variability covers training seeds only, and some margins over the strongest baseline are very small, e.g. 63.29 versus 63.16 Kappa against LoRA with CBraMod on the ISRUC dataset. In addition, several entries in the appendix tables appear to be duplicated across backbones (e.g., the TASTEFUL rows for EEGMamba and CodeBrain in Table 18), which should be checked.

Questions for authors

Is the prototype conditioning also computed for the test subjects at inference? If so, the performance of the task-prior network on its own, alongside a simple late-fusion baseline, should be reported. It would also be useful to see results for LoRA and full fine-tuning trained on the same filtered pseudo-labels, in order to get a better sense of how much of the gain is due to the supervision and how much is due to the adapter.

As for the experimental protocol, it is not made clear whether the sensitivity analyses were done on the validation or the test subjects, and how the default hyperparameters were chosen. It is also not clear whether all the baselines were tuned in the same way (e.g., FixMatch with CBraMod on ISRUC is considerably below the frozen backbone, which suggests a configuration issue rather than a fundamental limitation of the method). For the smaller datasets, it would be useful to report per-subject results or subject-level confidence intervals. The entries that appear to be duplicates in Tables 18–20 need to be checked.

9 comments
AuthorsOpenReview ↗

Response to Reviewer Mjxq [1/n]

We sincerely thank you for your positive feedback and for recognizing our study's comprehensive evaluation, robust performance and computational efficiency.

TL;DR

① Sources of Performance Improvements (Weakness 1 & Question 1; see our Responses 1-2)

We conducted all the additional baseline comparisons you suggested, leading to three main findings:

  • Beyond prediction-level ensembling: SCOPE outperforms both the task-prior network (TPN) and late-fusion baselines, indicating that its improvements cannot be attributed solely to prediction-level ensembling. A key distinction is SCOPE's selective utilization of TPN information through structured prototypes and high-confidence pseudo-labels supported by agreement across prediction sources.
  • Benefits of external supervision: Other adaptation methods also benefit from incorporating the same filtered pseudo-labels, but SCOPE maintains its performance advantage over these methods.
  • Contribution of prototype conditioning: Removing prototype conditioning reduces SCOPE's performance, demonstrating its additional contribution beyond pseudo-label supervision.

Together, these findings support our central claim that SCOPE's improvements arise from 1) reliable and selected external supervision, and 2) prototype-conditioned adaptation.

② Experimental Setups and Clarifications (Weakness 2 & Question 2; see our Responses 3-4)

We also appreciate your careful reading and detailed examination of our experimental results. We address the remaining concerns as follows:

ResponseAdditional analyses and clarificationsYour concern
Response 3Report per-subject performance on datasets with small test sets.Question 2.3
Response 4 (Part I)Clarify that hyperparameters were selected using validation data, while sensitivity analyses report test-set results.Question 2.1
Response 4 (Part II)Provide detailed diagnostics and additional experiments for the FixMatch baseline.Question 2.2
Response 4 (Part III)Correct the identified typo and recheck the related Tables.Question 2.4

We provide further details and address each of your concerns point by point below.

AuthorsOpenReview ↗

Response to Reviewer Mjxq [2/n]

1) W1&Q1 (1/2) Ablation study on ensemble effects

Inference procedure\color{blue}{\textbf{Inference procedure}}

Yes, your understand is correct. The prototype conditioning is also computed for the test subjects at inference.

TPN and late-fusion baselines\color{blue}{\textbf{TPN and late-fusion baselines}}

We evaluate TPN, full finetuning, and late fusion (TPN + full fine-tuning) as baselines on ISRUC with 20%, 30%, and 50% labeled training data across five EFM backbones.

Labeled training subject ratio20%30%50%
KappaWeighted-F1KappaWeighted-F1KappaWeighted-F1
TPN46.26±1.9455.83±1.6652.61±1.5261.36±1.6760.88±0.7968.47±0.76
LaBraM
Full finetuning48.55±1.2553.87±1.0351.26±1.2661.73±1.3063.83±1.6171.47±1.52
late fusion50.48±1.2760.62±0.8756.81±0.7864.81±0.6364.21±0.8671.40±0.78
SCOPE (Ours)51.67±1.0560.95±0.5756.19±0.6765.15±0.4765.52±1.3973.27±1.12
CBraMod
Full finetuning59.52±1.2163.81±1.3061.95±1.0269.78±0.9570.29±1.1076.67±0.89
late fusion60.01±0.8167.68±0.8062.43±0.6669.33±0.0469.75±0.4075.92±0.23
SCOPE (Ours)60.81±0.3868.96±0.3663.29±1.1270.74±0.6472.48±0.8078.51±0.75
CSBrain
Full finetuning52.59±0.9762.33±0.9060.97±1.0469.19±1.0963.83±1.6171.70±1.40
late fusion59.07±2.4967.98±2.2060.83±0.8268.82±1.1262.91±1.2670.62±1.06
SCOPE (Ours)60.59±0.9168.90±0.0961.51±0.5569.86±0.6965.48±0.9573.13±0.79
EEGMamba
Full finetuning53.59±0.8664.39±0.8062.02±0.8768.71±0.9168.79±1.3575.77±1.27
late fusion59.88±1.0268.47±0.7162.70±0.9570.51±0.8967.71±0.3874.76±0.16
SCOPE (Ours)60.24±0.2569.63±0.2363.07±0.8971.41±0.7969.33±0.5476.21±0.49
CodeBrain
Full finetuning49.79±1.7759.90±1.6258.46±1.4267.05±1.5761.89±1.5769.71±1.28
late fusion54.54±2.1562.27±2.0956.31±1.6263.88±1.6060.34±0.5768.80±0.59
SCOPE (Ours)57.48±1.1065.76±1.2463.46±0.9871.11±0.7464.81±0.6571.92±0.73

The results show that SCOPE generally outperforms late fusion, suggesting that SCOPE's improvements cannot be attributed solely to prediction-level ensemble effects. Unlike late fusion, which directly combines TPN and EFM predictions, SCOPE selectively leverages information from TPN through prototype construction and confidence-aware pseudo-label selection.

To further investigate ensemble effects, we quantify the average gains of late fusion over full fine-tuning across backbones:

Labeled training subject ratioΔ KappaΔ Weighted-F1
20%+3.99+4.54
30%+0.88+0.18
50%−0.74−0.76

Late fusion provides positive average gains when labeled data are scarce, but these gains diminish as the labeled-subject ratio increases and become negative when 50% of the training subjects are labeled. In contrast, SCOPE maintains positive improvements over full fine-tuning. These observations further support the importance of selectively leveraging external supervision rather than simply combining model predictions.

Revisions made\color{blue}{\textbf{Revisions made}}

We have added Appendix G.7, “Ablation Study on Ensemble Effects” (Lines 2318-2370), to the revised manuscript and included a reference to this additional analysis in Section 4.4, "Ablation Study" (Lines 483–485).

AuthorsOpenReview ↗

Response to Reviewer Mjxq [3/n]

2) W1&Q1 (2/2) Ablation study on adaptation methods

Results\color{blue}{\textbf{Results}}

Thank you for your thoughtful question. We evaluate the three adaptation baselines used in our paper, TASTEFUL, LoRA, and full fine-tuning, with the same pseudo-label supervision loss on ISRUC with 20%, 30%, and 50% labeled training data across five EFM backbones.

To further isolate the sources of improvement, we also extend the evaluation of SCOPE w/o prototype conditioning to these additional settings, representing the combination of pseudo-label supervision and vanilla FiLM adaptation, with partial results already reported in the ablation study in Table 2 of the original manuscript.

Labeled training subject ratio20%30%50%
KappaWeighted-F1KappaWeighted-F1KappaWeighted-F1
LaBraM
Pseudo label + TASTEFUL49.94±1.7758.53±2.7853.33±1.7662.58±1.5356.52±1.4765.16±1.36
Pseudo label + LoRA51.12±0.7760.11±1.9656.24±1.1864.74±2.2258.32±0.6367.18±0.71
Pseudo label + Full finetuning55.75±0.2664.40±0.3258.10±1.0865.64±1.5764.75±1.7072.31±1.55
SCOPE w/o prototype conditioning49.77±1.4258.15±1.3355.31±0.4364.50±0.4459.41±0.9968.79±1.20
SCOPE (Ours)51.67±1.0560.95±0.5756.19±0.6765.15±0.4765.52±1.3973.27±1.12
CBraMod
Pseudo label + TASTEFUL56.53±0.7264.68±0.7062.57±1.8969.94±1.7269.13±1.1875.96±1.24
seudo label + LoRA58.32±0.6366.39±1.2561.76±2.2369.29±2.0570.16±1.1276.88±0.99
Pseudo label + Full finetuning59.25±0.5467.55±0.4162.09±1.6769.49±1.8870.34±1.1376.98±0.95
SCOPE w/o prototype conditioning59.88±1.6068.25±1.7162.01±1.5769.23±1.4268.36±1.4475.13±1.45
SCOPE (Ours)60.81±0.3868.96±0.3663.29±1.1270.74±0.6472.48±0.8078.51±0.75
CSBrain
Pseudo label + TASTEFUL53.08±1.4660.94±3.1760.37±0.8868.75±0.8566.21±0.1472.92±0.61
Pseudo label + LoRA55.56±1.7465.74±1.5661.08±0.7068.57±1.0064.99±1.6172.32±1.61
Pseudo label + Full finetuning57.23±2.2566.62±2.1961.02±1.0668.99±0.9965.52±1.2173.74±1.20
SCOPE w/o prototype conditioning56.73±2.2865.92±2.0859.66±1.3567.59±1.1065.50±1.2172.66±1.34
SCOPE (Ours)60.59±0.9168.90±0.0961.51±0.5569.86±0.6965.48±0.9573.13±0.79
EEGMamba
Pseudo label + TASTEFUL53.17±1.3862.58±1.1055.37±1.6564.05±1.5959.60±2.5666.53±2.30
Pseudo label + LoRA59.22±2.0067.44±3.5763.18±0.6270.90±0.7166.90±0.6973.03±1.40
Pseudo label + Full finetuning58.83±2.9966.73±1.2262.94±1.1869.54±0.8769.77±1.3175.89±1.43
SCOPE w/o prototype conditioning58.94±1.9668.58±3.4663.41±1.1270.53±1.0469.20±1.5775.53±1.47
SCOPE (Ours)60.24±0.2569.63±0.2363.07±0.8971.41±0.7969.33±0.5476.21±0.49
CodeBrain
Pseudo label + TASTEFUL50.41±1.2656.69±1.4356.42±1.8864.11±1.8562.32±1.6268.86±1.61
Pseudo label + LoRA54.68±3.8663.01±3.7648.31±4.0854.79±4.4154.79±4.4163.49±3.50
Pseudo label + Full finetuning53.87±2.6563.96±2.4859.20±1.2768.82±1.2062.08±2.2070.67±2.02
SCOPE w/o prototype conditioning53.71±0.9462.16±2.2759.60±1.3366.97±1.4662.03±0.9269.81±0.65
SCOPE (Ours)57.48±1.1065.76±1.2463.46±0.9871.11±0.7464.81±0.6571.92±0.73

We summarize the results across the 15 backbone–label-ratio settings below:

MethodAverage KappaAverage Weighted-F1Best Kappa countBest Weighted-F1 count
Pseudo-label + TASTEFUL57.6665.491/150/15
Pseudo-label + LoRA58.9866.930/150/15
Pseudo-label + Full fine-tuning61.3869.423/153/15
SCOPE w/o prototype conditioning60.2368.251/150/15
SCOPE (Ours)62.4070.3710/1512/15

Overall, SCOPE achieves the highest average performance on both metric across the evaluated settings, with the alternative adaptation methods receiving the same confidence-weighted pseudo-label supervision.

Reviewer_P2gcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces the Structured COnfidence-aware Prototype-guided adaptation framework for EEG foundation models (SCOPE), which addresses the following key challenges of EEG foundation models: (c1) overconfidence undermines reliable self-guidance, (c2) models are fragile under imperfect supervision, and (c3) unconstrained fine-tuning risks overwriting the pretrained representations of EEG foundation models. These challenges are tackled by Task-Prior Induction via Inter-Class Geometry, Prototype Clustering via Intra-Class Variability, and a Prototype-Conditioned Adapter. The paper provides a comprehensive evaluation on 6 tasks and 5 EEG foundation models in a limited-label setting.

Critical strengths
  • Comprehensive, reproducible benchmarking across 6 decoding tasks, 3 Transformer-based models (LaBraM, CBraMod, CSBrain), and 2 state-space-style models (EEGMamba, CodeBrain).
  • Detailed ablation studies.
Critical weaknesses
  • Limited novelty: SCOPE is a combination of known ML techniques, namely a task-prior network with an Equiangular Tight Frame (Papyan et al.; Yang et al.), prototype learning (Snell et al.), pseudo-labeling (Rizve et al.), and FiLM adaptation (Perez et al.).
Questions for authors
  • In Table 26, the ablation of the ETF regularizer shows similar results (CSBrain: 60.04 ± 2.39 without ETF vs. 61.51 ± 0.55 with ETF) with high variance. How does the ETF regularization help the model?
  • Did the authors compare the pseudo-label loss and prototype conditioning with other adaptation methods?
5 comments
AuthorsOpenReview ↗

Response to Reviewer P2gc [1/n]

We sincerely thank you for your positive feedback and for recognizing our study's comprehensive evaluations and detailed ablation studies.

TL;DR

① Novelty (Weakness; see our Response 1)

SCOPE is the first to systematically diagnose the challenges of adapting EEG foundation models in realistic label-limited settings and introduce a novel adaptation paradigm that leverages prototype-based external supervision.

② ETF regularization (Question 1; see our Response 2)

ETF improves inter-class separability, providing two key benefits for the external supervision construction process

  1. improving task-prior pseudo-label quality
  2. providing cluster-friendly embeddings for subsequent prototype learning

③ Controlled comparisons (Question 2; see our Response 3)

We added the controlled comparisons with other adaptation methods as you suggested, and made several key observations:

  • SCOPE's performance advantage remains even when pseudo-labels are incorporated into other adaptation methods.
  • Prototype conditioning is an integral, tightly coupled component of our ProAdapter, rather than a readily separable add-on to other adaptation methods, and contributes importantly to its performance gains.
  • Our proposed pseudo-labels and confidence-weighting strategy provide transferable benefits across different adaptation methods.

We provide further details and address each of your concerns point by point below.

1) W1 Novelty

We acknowledge that SCOPE's primary contribution is a novel adaptation paradigm for EEG foundation models (EFMs). To the best of our knowledge, SCOPE is the first framework to specifically address the critical real-world challenge of adapting EFMs under limited-label settings. Our contribution lies in systematically diagnosing the key failure modes of existing fine-tuning strategies and creatively integrating complementary technical solutions into a coherent adaptation framework.

Specifically, recognizing that EFMs cannot reliably guide their own adaptation due to overconfident predictions, sensitivity to imperfect supervision, and the risk of overwriting pretrained representations, we introduce prototype-based external supervision to provide reliable and structured guidance throughout the adaptation process.


2) Q1 ETF Regularization

Thank you for this important question. There are two roles of ETF regularization:

  1. Improving task-prior pseudo-label quality. ETF regularization directly improves the quality of task-prior pseudo-labels generated by the TPN. For example, on ISRUC with 30% labeled training subjects, ETF regularization improves pseudo-label quality from 46.36 to 51.60 in Kappa and from 55.74 to 60.89 in weighted-F1.
  2. Providing a better geometric foundation for prototype clustering. ETF regularization encourages maximally separated class representations, thereby providing a geometric inductive bias that helps the TPN learn more cluster-friendly embeddings (see Appendix A.1 for the mathematical formulation). This establishes the first step of the hierarchical geometric structure, providing a more discriminative embedding space for subsequent intra-class prototype clustering.

We acknowledge that the improvements in final performance can be relatively modest on the CSBrain backbone. Nevertheless, the intermediate analysis demonstrates that ETF regularization makes meaningful contributions to both task-prior pseudo-label quality and the geometric foundation for prototype construction. These contributions are important to the external structured supervision construction process.

Revisions made\color{blue}{\textbf{Revisions made}}

We have added the following explanation to Appendix G.5 (lines 2249–2253) to further clarify the role of ETF regularization in improving intermediate representations.

Beyond downstream performance, we examine the effect of ETF regularization on task-prior pseudo-label quality. On ISRUC with 30% labeled training subjects, incorporating ETF regularization improves TPN pseudo-label Kappa from 46.36 to 51.60 and weighted F1 from 55.74 to 60.89. In addition to improving task-prior pseudo-label quality, ETF regularization provides a geometric inductive bias that supports subsequent prototype construction.

AuthorsOpenReview ↗

Response to Reviewer P2gc [2/n]

3) Q2 Ablation study on adaptation methods

Thank you for this thoughtful question. To assess the performance of pseudo-label supervision and prototype conditioning with other adaptation methods, we conduct two complementary comparisons. The pseudo-label loss can be directly applied to different adaptation methods, while prototype conditioning is part of the ProAdapter’s dedicated modulation interface.

Therefore, we examine the contributions of pseudo-label supervision and prototype conditioning through two complementary comparisons on ISRUC with 20%, 30%, and 50% labeled training subjects across five EFM backbones:

  1. To assess the benefit of pseudo-label supervision across different adaptation methods, we evaluate TASTEFUL, LoRA, and full fine-tuning with the same pseudo-label supervision loss used in the original manuscript.
  2. To isolate the additional contribution of prototype conditioning, we compare full SCOPE with SCOPE w/o prototype conditioning, which retains vanilla FiLM adaptation and identical pseudo-label supervision. We extend this ablation, reported in Table 2 of the original manuscript, to all the settings above.

Results\color{blue}{\textbf{Results}}

Labeled training subject ratio20%30%50%
KappaWeighted-F1KappaWeighted-F1KappaWeighted-F1
LaBraM
Pseudo label + TASTEFUL49.94±1.7758.53±2.7853.33±1.7662.58±1.5356.52±1.4765.16±1.36
Pseudo label + LoRA51.12±0.7760.11±1.9656.24±1.1864.74±2.2258.32±0.6367.18±0.71
Pseudo label + Full finetuning55.75±0.2664.40±0.3258.10±1.0865.64±1.5764.75±1.7072.31±1.55
SCOPE w/o prototype conditioning49.77±1.4258.15±1.3355.31±0.4364.50±0.4459.41±0.9968.79±1.20
SCOPE (Ours)51.67±1.0560.95±0.5756.19±0.6765.15±0.4765.52±1.3973.27±1.12
CBraMod
Pseudo label + TASTEFUL56.53±0.7264.68±0.7062.57±1.8969.94±1.7269.13±1.1875.96±1.24
seudo label + LoRA58.32±0.6366.39±1.2561.76±2.2369.29±2.0570.16±1.1276.88±0.99
Pseudo label + Full finetuning59.25±0.5467.55±0.4162.09±1.6769.49±1.8870.34±1.1376.98±0.95
SCOPE w/o prototype conditioning59.88±1.6068.25±1.7162.01±1.5769.23±1.4268.36±1.4475.13±1.45
SCOPE (Ours)60.81±0.3868.96±0.3663.29±1.1270.74±0.6472.48±0.8078.51±0.75
CSBrain
Pseudo label + TASTEFUL53.08±1.4660.94±3.1760.37±0.8868.75±0.8566.21±0.1472.92±0.61
Pseudo label + LoRA55.56±1.7465.74±1.5661.08±0.7068.57±1.0064.99±1.6172.32±1.61
Pseudo label + Full finetuning57.23±2.2566.62±2.1961.02±1.0668.99±0.9965.52±1.2173.74±1.20
SCOPE w/o prototype conditioning56.73±2.2865.92±2.0859.66±1.3567.59±1.1065.50±1.2172.66±1.34
SCOPE (Ours)60.59±0.9168.90±0.0961.51±0.5569.86±0.6965.48±0.9573.13±0.79
EEGMamba
Pseudo label + TASTEFUL53.17±1.3862.58±1.1055.37±1.6564.05±1.5959.60±2.5666.53±2.30
Pseudo label + LoRA59.22±2.0067.44±3.5763.18±0.6270.90±0.7166.90±0.6973.03±1.40
Pseudo label + Full finetuning58.83±2.9966.73±1.2262.94±1.1869.54±0.8769.77±1.3175.89±1.43
SCOPE w/o prototype conditioning58.94±1.9668.58±3.4663.41±1.1270.53±1.0469.20±1.5775.53±1.47
SCOPE (Ours)60.24±0.2569.63±0.2363.07±0.8971.41±0.7969.33±0.5476.21±0.49
CodeBrain
Pseudo label + TASTEFUL50.41±1.2656.69±1.4356.42±1.8864.11±1.8562.32±1.6268.86±1.61
Pseudo label + LoRA54.68±3.8663.01±3.7648.31±4.0854.79±4.4154.79±4.4163.49±3.50
Pseudo label + Full finetuning53.87±2.6563.96±2.4859.20±1.2768.82±1.2062.08±2.2070.67±2.02
SCOPE w/o prototype conditioning53.71±0.9462.16±2.2759.60±1.3366.97±1.4662.03±0.9269.81±0.65
SCOPE (Ours)57.48±1.1065.76±1.2463.46±0.9871.11±0.7464.81±0.6571.92±0.73
AuthorsOpenReview ↗

Response to Reviewer P2gc [3/n]

3) Q2 Ablation study on adaptation methods (Continue)

Observations\color{blue}{\textbf{Observations}}

There are three main observations from the results.

  1. SCOPE retains advantage over alternative adaptation methods under the same pseudo-label supervision. We summarize the results across the 15 backbone–label-ratio settings below. All average gains below are computed with equal weight across the evaluated backbone-label-ratio settings and reported as absolute metric-point differences.

    MethodAverage KappaAverage Weighted-F1Best Kappa countBest Weighted-F1 count
    Pseudo-label + TASTEFUL57.6665.491/150/15
    Pseudo-label + LoRA58.9866.930/150/15
    Pseudo-label + Full fine-tuning61.3869.423/153/15
    SCOPE w/o prototype conditioning60.2368.251/150/15
    SCOPE (Ours)62.4070.3710/1512/15

    Overall, SCOPE achieves the highest average performance on both metric across the evaluated settings, with the alternative adaptation methods receiving the same confidence-weighted pseudo-label supervision.

  2. Prototype conditioning provides additional gains beyond vanilla FiLM adaptation.

    Compared with SCOPE w/o prototype conditioning, full SCOPE improves average Kappa by 2.16 points and weighted-F1 by 2.11 points. This controlled comparison supports the added value of using prototype-derived information to condition feature modulation.

  3. Pseudo-label supervision and confidence-aware loss weighting provide transferable benefits across adaptation methods. We report the average performance gains of TASTEFUL, LoRA, and full fine-tuning after incorporating our pseudo-label supervision, compared with their results without pseudo-labels reported in the manuscript, across 15 settings:

    Adaptation methodΔ Kappa (with − without pseudo-label supervision)Δ Weighted-F1 (with − without pseudo-label supervision)
    TASTEFUL+2.84+2.06
    LoRA+4.86+4.55
    Full finetuning+2.23+2.35

Revisions made\color{blue}{\textbf{Revisions made}}

We have added Appendix G.8, "Ablation Study on Adaptation Methods" (lines 2372-2451), to the revised manuscript, presenting controlled comparisons and analyses of other adaptation methods under the same pseudo-label supervision.

We also included a reference to this additional analysis in Section 4.4, "Ablation Study" (Lines 483–485).


Thank you once again for your thorough and thoughtful review, and encouraging feedback. We hope our clarifications and additional analyses address your concerns. We look forward to your feedback, and are happy to provide further information if needed.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.