acceptodds
Under review as a conference paper at ICLR 2027

RPT: Resultant-Power Transport for Data-Free Continual EEG Model Merging

Abstract

Electroencephalography (EEG) supports diverse brain–computer interface and monitoring applications, requiring models that adapt to new subjects while retaining previously acquired knowledge. However, inter-subject variability can undermine knowledge retention during adaptation, and replay-based remedies require storing sensitive EEG recordings, raising privacy concerns and increasing storage demands. Continual model merging (CMM) offers a scalable alternative by incrementally merging subject-specific fine-tuned models without storing or replaying historical EEG. Nevertheless, conflicting parameter updates and differing normalization states across subjects can disrupt accumulated knowledge during merging. We introduce Resultant-Power Transport (RPT), which recursively merges subject models without accessing EEG or calibration samples. RPT uses task-vector agreement to scale layer-specific targets and selects incoming contributions through a closed-form, history-weighted projection that preserves the relative weights of historical contributions. Affine-map BN merging (AM-BN) complements this update by merging the inference transformations jointly defined by batch normalization statistics and affine parameters. Experiments across three EEG tasks demonstrate that RPT outperforms competing CMM methods, enabling data-free continual adaptation to new subjects while sustaining performance on previously encountered subjects.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
continual model merging, EEG decoding, continual learning, data-free model merging, batch normalization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jnuSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Resultant-Power Transport, a data-free continual model merging framework for subject-incremental EEG learning. RPT introduces a resultant-guided task-vector transport mechanism that uses cross-subject agreement to construct layer-specific targets and a closed-form projection to determine incoming contributions while preserving historical weights. Additionally, the proposed Affine-Map Batch Normalization mechanism merges normalization transformations using compatibility-weighted affine maps derived from stored BN statistics. Experiments on 3 EEG datasets demonstrate strong performance.

Critical strengths
  • The paper addresses data-free continual EEG model merging, which is particularly relevant for scenarios where storing or replaying historical EEG recordings is undesirable.
  • The proposed history-weighted projection provides a computationally efficient mechanism.
Critical weaknesses
  • Table 3 reports extremely low performance on FACED when AM-BN is removed. These results are also substantially lower than the weight-averaging baselines reported in Table 2.
  • Although the experiments cover three EEG tasks, they all use variants of the same CNN-Transformer backbone. The effectiveness of task-vector transport may depend on the architecture and the relative contributions of convolutional and linear layers. Additional evaluations on different EEG architectures would strengthen the generality of the proposed framework.
Questions for authors
  • Table 3 shows that removing AM-BN causes FACED performance to drop to approximately 8%, substantially below the WA baseline. Could the authors explain the source of this degradation?
  • Could the authors evaluate RPT on different backbone architectures, particularly EEG models with substantially different convolutional or Transformer configurations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer jnuS[1/N]

Thank you for your constructive review and for recognizing the merits of our approach. We appreciate the opportunity to address your questions and provide further context for our findings.

W1 / Q1. Why FACED accuracy drops without AM-BN

The 7.8% accuracy in Table 3 arises from independently averaging BN affine parameters and running statistics. In the first BN layer of each CNN branch, subject 23 accounted for about 99.9% of the summed expert running variances, and subject 4 accounted for about three quarters of the remainder. The large averaged variances reduced the effective BN scales, and the ablated model assigned about 98% of test trials to a single class in each of the ten orders.

Native weight averaging (WA) in Table 2 keeps the first expert’s BN running statistics. Subject 23 arrived after the first expert in all ten orders. In one order, replacing only the running statistics in the first BN layer of each branch with those of subject 23 reduced WA accuracy from 50.98% to 10.20%. In the same order, with the ablation’s non-BN weights fixed, using the first expert’s running statistics raised accuracy from 7.84% to 55.69%. Replacing the full BN state with its AM-BN counterpart raised accuracy to 58.43%.

We repeated merging after excluding subject 23 alone and after excluding both subjects 23 and 4. The table reports mean final ACC (%) over ten matched orders on the same 83 test subjects (Appendix C.7, Table 15, p. 33).

Subject checkpoints excluded from mergingRPT w/o AM-BNFull RPT
None7.6358.59
Subject 2327.8758.76
Subjects 23 and 443.0158.76

With AM-BN, RPT achieved 58.59% accuracy with all 85 experts included, within 0.17 percentage points of either exclusion run. With both checkpoints excluded and AM-BN held fixed, adding task-vector transport still improved accuracy from 52.61% to 58.76%.


W2 / Q2. Evaluation with the CBraMod backbone

Thank you for suggesting this evaluation. We compared RPT with continual model merging baselines on all three datasets using CBraMod [1], a distinct EEG backbone that combines convolutional patch embeddings with criss-cross attention. We followed the same subject-incremental experimental setup described in the manuscript. For each dataset, all merging methods received the same CBraMod checkpoints (50 epochs of pre-training for the shared model and 25 epochs of fine-tuning per subject) in the same arrival order.

The table reports CBraMod results, with accuracy (ACC) and macro-F1 (MF1) expressed as percentages and backward transfer (BWT) in percentage points (pp).

MethodPhysionet-MI ACC/MF1Physionet-MI BWT-ACC/BWT-MF1FACED ACC/MF1FACED BWT-ACC/BWT-MF1ISRUC ACC/MF1ISRUC BWT-ACC/BWT-MF1
Pre-trained47.89 / 41.18— / —39.56 / 30.67— / —61.58 / 48.82— / —
Independent FT53.36 / 48.66— / —44.18 / 34.89— / —73.08 / 57.04— / —
WA51.49 / 45.88-0.32 / -0.1841.65 / 32.56+0.74 / +0.7262.28 / 48.72-1.21 / -1.10
Task Arithmetic40.69 / 29.71-4.29 / -6.999.85 / 3.79-12.35 / -12.1521.64 / 8.42-16.16 / -16.30
TIES51.33 / 45.70-2.06 / -2.9242.06 / 32.87+0.60 / +0.6462.58 / 49.47-4.47 / -3.38
MagMax50.39 / 44.63-0.63 / -0.7641.91 / 32.64+0.30 / +0.3362.82 / 48.77-0.53 / -0.97
OPCM52.11 / 47.10+0.16 / +0.5640.29 / 31.34-1.49 / -1.3362.99 / 48.69-1.76 / -1.63
NUFILT52.27 / 46.19-1.75 / -2.3142.38 / 33.010.00 / -0.3061.32 / 46.30-7.75 / -7.42
WUDI-Merging43.51 / 36.38-7.30 / -9.1413.09 / 8.44-18.45 / -15.0062.23 / 47.78-9.47 / -7.50
DOP52.27 / 46.45-1.27 / -2.1941.76 / 32.57+0.74 / +0.0461.49 / 46.76-8.58 / -7.09
RPT53.68 / 48.26+1.59 / +1.4542.94 / 33.44+1.19 / +0.7763.87 / 49.96-0.70 / -0.72

With CBraMod, RPT achieved the highest final ACC and macro-F1 among the merging methods on all three datasets and led five of the six BWT comparisons. MagMax led BWT-ACC on ISRUC.

[1] Wang J, Zhao S, Luo Z, et al. Cbramod: A criss-cross brain foundation model for eeg decoding. ICLR 2025


We hope our responses and additional results adequately address the points you raised. We appreciate your consideration and would be pleased to discuss any remaining questions. For further details, please refer to the revised manuscript, where changes to both the main text and the appendix are marked in blue.

Thank you for the additional experiments, and I have no further concerns. I will keep my score.

Thank you for maintaining your positive rating! We appreciate the time and care you put into reviewing our paper. Your constructive feedback helped us strengthen our work!

Reviewer_3ccEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses catastrophic forgetting and inter-subject variability in continual EEG learning by proposing Resultant-Power Transport (RPT), a data-free continual model merging framework. Unlike replay-based methods, RPT incrementally integrates subject-specific fine-tuned models without requiring access to historical EEG recordings during merging. Experiments on three EEG datasets covering motor imagery, emotion recognition, and sleep staging demonstrate improvements over existing continual model merging baselines in classification accuracy, macro-F1, and prior-subject retention.

Critical strengths
  1. A principled and efficient framework for data-free continual EEG model merging. The paper introduces a coherent framework that combines cross-subject task-vector agreement with history-weighted projection to balance knowledge retention and new-subject adaptation. The closed-form update avoids storing historical EEG recordings during merging and maintains constant auxiliary storage with respect to the number of subjects. This provides a practically relevant solution to continual EEG adaptation under data-access and computational constraints.
  2. Comprehensive experimental validation and complementary treatment of normalization. The evaluation covers three EEG tasks and compares RPT with multiple continual model merging and conventional continual learning baselines. Importantly, the authors conduct comparisons under matched normalization settings, factorial ablations, and mechanism-specific controls. These experiments provide evidence that both task-vector transport and affine-map BN merging contribute to the reported performance gains, rather than attributing improvements solely to the combination of multiple components.
Critical weaknesses
  1. Insufficient analysis of individual-level forgetting during continual learning. Although the paper evaluates historical knowledge retention using backward transfer (BWT) and arrival-position-based analyses, these results primarily reflect aggregate performance across previously encountered subjects. Such aggregate metrics may conceal substantial forgetting on individual subjects, particularly those introduced early in the learning sequence. The paper does not sufficiently characterize how recognition performance on specific historical subjects evolves as more new subjects are incorporated, or whether improvements on some subjects compensate for severe forgetting on others. Consequently, the evidence supports improved average retention but does not fully establish consistent preservation of previously acquired knowledge.
  2. Limited investigation of subject-order sensitivity and its relationship to the proposed merging mechanism. Although the authors evaluate ten random subject orders and report relatively stable overall performance, the proposed recursive merging procedure remains potentially order-dependent because the incoming contribution is determined using the previously merged model. The current experiments do not systematically investigate how different subject-order structures, particularly sequences involving substantial inter-subject distribution shifts, affect historical knowledge retention. Moreover, while the layer-specific agreement controls demonstrate performance benefits, they do not directly establish whether the proposed transport mechanism effectively mitigates order-dependent interference between subject-specific parameter updates. This limits our understanding of the method's robustness under more challenging continual learning scenarios.
Questions for authors
  1. Individual-level forgetting and long-term knowledge retention. The paper reports favorable BWT and prior-subject retention results. However, could the authors provide retention trajectories for individual early-arriving subjects as additional subjects are incorporated? In particular, are there subjects that experience substantial performance degradation despite the favorable average BWT? Reporting worst-case forgetting or the distribution of subject-level forgetting would help clarify whether RPT consistently preserves historical knowledge.
  2. Sensitivity to subject arrival order. Although ten random subject orders are evaluated, how does RPT behave under structured arrival sequences, such as introducing similar subjects consecutively or placing subjects with substantial distribution shifts earlier versus later? Would the same subject exhibit different final recognition performance depending on its position in the sequence? Such experiments would help assess whether the reported retention advantage remains robust under challenging subject orders.
  3. Relationship between task-vector agreement and historical knowledge preservation. The proposed method uses task-vector agreement and history-weighted projection to determine incoming contributions. However, does higher task-vector agreement consistently correspond to lower interference or better retention of previously learned subjects? Could the authors provide an analysis relating task-vector similarity, subject distribution differences, and subsequent forgetting? This would help explain why the proposed mechanism improves continual learning performance beyond the observed aggregate results.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 3ccE[1/N]

We sincerely appreciate the care and attention you devoted to reviewing our manuscript. Your detailed assessment and constructive suggestions have helped us strengthen the evaluation and improve the presentation of our work.

W1 / Q1. Individual-level forgetting and retention

Thank you for pointing out that average retention can conceal individual losses. In Appendix C.8 (p. 34–36), Figures 13–14 show ACC and macro-F1 trajectories averaged across ten arrival orders, and Figures 15–16 show individual trajectories. RPT’s trajectories were generally stable as more subjects were incorporated.

How often do large drops occur? The table summarizes the ACC results in Table 16 (p. 34). Counts are subject–order pairs whose ACC fell more than 10 percentage points (pp) below their score at incorporation. Early arrivals are the first 10% of each stream. The later group includes the other prior subjects.

DatasetEarly, any later mergeEarly, final mergeLater, final mergeLargest early final drop (pp)
Physionet-MI42/8030/80 (37.5%)61/620 (9.8%)33.33
FACED15/908/90 (8.9%)17/750 (2.3%)33.33
ISRUC8/705/70 (7.1%)5/590 (0.8%)38.75

Large final ACC drops were more frequent among early arrivals on all three datasets, with the largest reaching 38.75 pp on ISRUC. On FACED, early arrivals gained 4.44 pp in mean ACC, while 8 of 90 subject–order pairs ended more than 10 pp below their incorporation score. Table 17 (p. 36) shows the largest final drop among early arrivals in each dataset.

How does RPT compare with NUFILT? We compared the two methods under matched AM-BN. Each pair reads NUFILT → RPT.

DatasetEarly mean final ACC (%)Early final drops >10 ppLargest early final drop (pp)Final drops >10 pp among all prior subjects (%)
Physionet-MI49.44 → 51.2233/80 → 30/8044.44 → 33.3334.29 → 13.00
FACED54.81 → 56.676/90 → 8/9033.33 → 33.337.02 → 2.98
ISRUC65.80 → 67.297/70 → 5/7056.25 → 38.754.55 → 1.52

Compared with NUFILT, RPT achieved higher mean final ACC for early arrivals and fewer large final drops across all prior subjects on all three datasets.


W2a / Q2. Sensitivity to structured subject orders

Thank you for suggesting this test. We constructed five structured arrival orders. We used the frozen pre-trained model to extract features from training EEG before the classifier. We measured distances between feature distributions using squared maximum mean discrepancy (MMD²) with a Gaussian kernel [1].

  • Similar subjects together. We used pairwise MMD² and average-linkage clustering to form three groups. Within each group, we started with its most central subject and repeatedly chose the remaining subject closest to the last one added. We joined the groups in order of increasing distance from their central subjects to the pre-training reference.
  • High-shift group first or last. We selected the 20% of subjects (rounded up) with the largest MMD² to the pooled pre-training feature distribution, giving each pre-training subject equal weight. We placed this same group first or last, preserving the relative order within both groups from the clustered sequence.
  • Lower or higher pre-trained ACC first. We ranked subjects by the frozen model’s accuracy on each subject’s own training data, from lowest to highest or from highest to lowest.

The table below reports RPT’s final ACC and macro-F1 (%) for each of the five arrival orders.

DatasetSimilar subjects togetherHigh-shift group firstHigh-shift group lastLower pre-trained ACC firstHigher pre-trained ACC first
Physionet-MI53.56 / 46.6853.56 / 46.6853.56 / 46.6853.56 / 46.6853.40 / 46.52
FACED58.82 / 49.0658.43 / 48.5458.82 / 49.0658.43 / 48.5458.43 / 48.54
ISRUC72.36 / 59.5172.36 / 59.4772.32 / 59.4672.41 / 59.5772.32 / 59.46

Across the five orders, RPT’s mean final ACC varied by less than 0.40 percentage points (pp) on each dataset. Moving the high-shift group from first to last changed final ACC for 0/71, 1/85, and 6/67 subjects on Physionet-MI, FACED, and ISRUC, respectively.

Using the prior-subject score defined in Figure 3(c) (p. 7), which averages final ACC and macro-F1 over all subjects except the last arrival, RPT outperformed the strongest baseline in 5/5 structured orders on FACED, 4/5 on ISRUC, and 2/5 on Physionet-MI. On FACED, all high-shift subjects reached the same final ACC in both placements. Among the 16 group members who underwent further merges in both orders, mean ACC from incorporation to the final merge fell by 2.08 pp when the group arrived first and rose by 2.08 pp when it arrived last.

[1] Gretton A et al. A Kernel Two-Sample Test. JMLR 13:723–773, 2012.

AuthorsOpenReview ↗

Response to Reviewer 3ccE[2/N]

W2b / Q3. Task-vector agreement and forgetting

Layer-averaged agreement showed little association with immediate ACC drops on earlier subjects. Across ten orders per dataset, the correlation was close to zero after accounting for merge position and pre-merge ACC. Stronger alignment between the incoming task vector and the merged model’s displacement from the pre-trained model was associated with smaller drops on Physionet-MI (0.84 to 0.48 pp) and ISRUC (0.55 to 0.21 pp). On ISRUC, larger shifts between incoming and earlier subjects’ EEG feature distributions were associated with larger drops (0.27 to 0.52 pp).

In RPT, agreement sets each layer’s target scale for the history-weighted projection. At seven points in one order per dataset (merge 2 and 10%, 25%, 50%, 75%, 90%, and 100% of the stream), we recomputed the projection with agreement scaling enabled or with the scale fixed at 1, using the same pre-merge model and incoming expert. We averaged ACC drops over earlier subjects, counting gains as zero, and measured the incoming subject’s ACC change from before to after the merge. The table averages both measures over the seven points and reads scaling off → on.

DatasetMean ACC drop on earlier subjects (pp)Mean ACC gain on the incoming subject (pp)
Physionet-MI1.40 → 0.284.76 → 3.17
FACED0.12 → 0.060.00 → 0.00
ISRUC1.03 → 1.100.68 → 0.86

Scaling reduced immediate drops on Physionet-MI and FACED. On ISRUC, both the incoming gain and the earlier-subject drop increased. With AM-BN fixed, adding task-vector transport improved mean ACC-based backward transfer on all three datasets. Across the ten original orders, the gains were 0.87, 1.35, and 0.25 pp, respectively (Table 3, p. 8). A possible refinement is to cap the contribution of poorly aligned updates and test whether this improves retention while preserving incoming-subject gains under high-shift arrival orders.


Thank you again for your careful reading and valuable comments. Your feedback helped us strengthen the manuscript, and we would be glad to discuss any remaining questions. We hope that our responses and additional results adequately address your concerns. Changes to both the main text and the appendix are marked in blue in the revised manuscript.

Dear Reviewer 3ccE,

Thank you again for your careful reading and insightful review. We sincerely appreciate your recognition of the methodological and experimental strengths of our work, as well as the thoughtful suggestions you offered. Your feedback has been valuable in helping us deepen the analysis and strengthen the manuscript.

As the discussion period draws to a close, we would like to follow up on our responses and revised manuscript. Your suggestions led us to add experiments and analyses addressing these points.

  • Individual retention trajectories and the distribution of forgetting.
  • Sensitivity to structured subject arrival orders.
  • Task-vector agreement, subject distribution shifts, and the role of task-vector transport.

We would be very grateful to hear your thoughts on whether these additions address your concerns. If any points would benefit from further explanation, we would welcome the opportunity to clarify them and continue the discussion.

Thank you once again for the time, care, and thoughtful consideration you have devoted to our work.

Best regards,

The Authors

Reviewer_dd2AOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper tackles the problem of adapting EEG models and proposes RPT in line with continual model merging approaches. RPT, in short, provides a way to merge subject models without accessing EEG or calibration samples. It consists of two components including, resultant guided task vector transport and AM-BN.

Critical strengths
  • The proposed method is well designed to address the challenges of conflicting parameter updates and differences in normalization statistics across subjects, while preserving decoding performance on previously encountered subjects. This is a key strength of the approach, as it addresses important challenges in continual subject adaptation.
Critical weaknesses
  • A major concern is practical applicability. Although the merged model requires no labels at inference, incorporating each new subject requires a fine-tuned checkpoint, which in turn requires a substantial amount of that subject’s labeled EEG (90% based on the training-to-testing ratio is applied to every newly introduced subject, according to the manuscript). In realistic deployment, a new user typically arrives with unlabeled data, so RPT may fail and cannot adapt.

  • Another concern is the motivation for merging when labeled data is available for each subject. Since every subject checkpoint is fine-tuned on that subject’s labeled data and subject identity is known at inference, the subject’s own fine-tuned model could simply be used directly. According to Table 2, independent fine-tuning substantially outperforms the final merged model on Physionet-MI and ISRUC. Additionally, the applicability of RPT to EEG foundation models is unclear. Although the merging step is O(P), the pipeline requires fully fine-tuning and storing a complete checkpoint for every subject, which becomes computationally and storage-intensive for large models.

Questions for authors
  • Could the authors clarify the motivation for merging subject-specific models when labeled data is available for each subject? Since each subject-specific checkpoint is fine-tuned using that subject’s labeled data, why not directly use the corresponding fine-tuned model?

  • Could the authors report the performance of the final merged model on unseen subjects outside those used for adaptation, compared to the original pretrained model?

  • How does performance vary with the number or fraction of labeled samples used to fine-tune each subject checkpoint (e.g., 10%, 25%, 50% instead of 90%)? Does RPT’s advantage over the baselines persist when subject checkpoints are trained on few labels, and how does the merged model compare with independent fine-tuning in that low-label regime?

  • Why are the results of RPT with and without AM-BN (CMM and CMM with AM-BN) identical in Table 2 but not in Table 3?

  • What do independent and sequential fine-tuning mean?

  • Given the growing adoption of EEG foundation models, could the authors discuss whether the proposed approach can be extended to these larger architectures without incurring substantial computational overhead?

  • Are the gains from the proposed approach statistically significant?

6 comments
AuthorsOpenReview ↗

Response to Reviewer dd2A[1/N]

Thank you for your thoughtful assessment of our manuscript. We appreciate the constructive questions you raised and the opportunity to clarify and expand on our work.

W1. Label requirements for new subjects

Thank you for raising this practical concern. We address it in two parts.

  1. Label requirements of the merging rule. RPT uses only model parameters for merging and can accept compatible subject checkpoints from either supervised or unsupervised fine-tuning. In either case, gradients of the fine-tuning objective drive parameter updates from a shared pre-trained model, and the task vector captures their cumulative effect (Appendix A.1, p. 15, and Refs. [1, 2]). RPT then uses agreement among these vectors and the current merged model to determine how much of the incoming update to incorporate, with the aim of limiting interference with earlier adaptations.

  2. Choice of benchmark protocol. Following prior continual model merging (CMM) evaluations [2, 3], we used the available training labels to fine-tune subject-specific models and supplied the same checkpoints to all merging methods. This allowed us to compare merging rules under the same fine-tuning conditions. Using labeled EEG from incoming subjects for model updates is also a common experimental setting in EEG continual learning [4, 5].

How well RPT performs when new subjects arrive with only unlabeled EEG remains an open question. Your suggestion motivates a follow-up study that draws on unsupervised EEG adaptation strategies such as those in BrainUICL [6] to generate compatible subject checkpoints and test whether merging them improves decoding for incoming subjects while preserving performance on earlier subjects. We now note this direction in Appendix D (Limitations, p. 36).

[1] Cheng R, Xiong F, Wei Y, et al. Whoever started the interference should end it: Guiding data-free model merging via task vectors. ICML 2025

[2] Yang E, Tang A, Shen L, et al. Continual model merging without data: Dual projections for balancing stability and plasticity. NeurIPS 2025

[3] Qiu Z, Wang L, Cao Y, et al. Null-space filtering for data-free continual model merging: Preserving stability, promoting plasticity. ICLR 2026

[4] Duan T, Wang Z, Shen L, et al. Retain and adapt: Online sequential eeg classification with subject shift[J]. IEEE Transactions on Artificial Intelligence, 2024, 5(9): 4479-4492.

[5] Li D, Shin H B, Lee S W. Toward Memory-Efficient Continual Adaptation for MI-EEG Decoding in BCIs[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2025, 56(1): 766-778.

[6] Zhou Y, Zhao S, Wang J, et al. BrainUICL: An unsupervised individual continual learning framework for EEG applications. ICLR 2025


W2a / Q1. Why merge subject-specific models?

Thank you for raising this deployment trade-off. Independent FT requires storing and maintaining a separate model for every subject. In an EEG monitoring service that continually enrolls users, each new user adds another checkpoint to maintain, and every inference request must be routed to the correct model. This becomes impractical as the number of users grows. RPT incorporates each new subject’s adaptation into one shared model per EEG task. The service deploys a single model for all enrolled users and can decode their EEG without using subject identity to select a checkpoint.

Table 2 (p. 7) compares the two deployment settings. For m subjects, Independent FT evaluates m separately fine-tuned models, each on its corresponding subject’s held-out EEG, and provides a reference for personalized decoding. RPT evaluates one final merged model on the same m test sets. We clarified the Independent FT protocol in Section 4.2 (p. 6) of the revised manuscript.


Q2. Performance on unseen subjects

Thank you for suggesting this evaluation. For each dataset, we evaluated ten splits of the incremental subject pool, using approximately 80% of subjects for adaptation and holding out the remaining 20%. RPT merged checkpoints from the adaptation subjects, and both models were evaluated on the same held-out subjects. The table reports mean scores (%) across ten splits, with sample variances (pp²) in parentheses.

DatasetPre-trained ACCRPT ACCPre-trained macro-F1RPT macro-F1Gain ACC / macro-F1
Physionet-MI44.70(0.22)51.15(0.10)42.24(0.26)49.36(0.11)+6.44 / +7.13
FACED39.83(0.22)55.04(0.18)36.04(0.19)53.28(0.10)+15.21 / +17.24
ISRUC70.85(0.40)74.74(0.15)64.65(0.36)69.28(0.18)+3.88 / +4.63

On held-out subjects, the RPT merged model achieved higher mean ACC and macro-F1 than the pre-trained model on all three datasets.

AuthorsOpenReview ↗

Response to Reviewer dd2A[2/N]

Q3. Performance with fewer labels per subject

We fine-tuned subject checkpoints using 10%, 25%, or 50% of each subject’s total samples, drawn from the original training pool. All CMM methods used the same AM-BN. Independent FT evaluated each checkpoint on its corresponding subject’s test set. All entries report final ACC / macro-F1 (%).

10% labeled samples

MethodPhysionet-MIFACEDISRUC
Independent FT51.67 / 45.4250.98 / 40.9678.59 / 64.32
WA51.18 / 44.9351.37 / 41.5670.45 / 57.07
Task Arithmetic26.20 / 17.4048.63 / 37.8852.84 / 28.83
TIES51.30 / 44.5950.20 / 40.2468.41 / 55.49
MagMax50.43 / 43.0049.80 / 40.1171.27 / 58.16
OPCM22.54 / 12.7223.92 / 15.8726.34 / 9.97
NUFILT51.87 / 44.2751.37 / 41.1470.87 / 57.34
WUDI-Merging47.80 / 40.3136.86 / 28.0166.39 / 51.67
DOP51.14 / 43.4550.98 / 41.3467.53 / 53.67
RPT(Ours)53.12 / 46.0652.55 / 42.7971.97 / 58.86

25% labeled samples

MethodPhysionet-MIFACEDISRUC
Independent FT56.08 / 50.3752.55 / 42.5181.17 / 67.59
WA51.87 / 45.3350.59 / 40.8071.38 / 58.50
Task Arithmetic26.62 / 12.1533.33 / 24.4153.97 / 31.44
TIES52.65 / 46.2249.80 / 40.0569.74 / 56.74
MagMax50.74 / 43.7748.24 / 37.9971.46 / 59.60
OPCM26.34 / 13.2217.65 / 11.6531.54 / 11.15
NUFILT50.73 / 45.2052.55 / 42.9270.98 / 59.61
WUDI-Merging47.42 / 40.2241.18 / 31.4270.93 / 56.98
DOP51.64 / 45.9552.16 / 42.2569.94 / 57.35
RPT(Ours)53.28 / 46.6256.47 / 47.1271.98 / 59.32

50% labeled samples

MethodPhysionet-MIFACEDISRUC
Independent FT56.96 / 51.5049.80 / 39.3182.75 / 69.71
WA52.83 / 46.1651.37 / 41.1071.91 / 58.93
Task Arithmetic26.15 / 11.1234.12 / 24.5961.93 / 37.69
TIES52.34 / 45.9649.80 / 40.1170.56 / 58.10
MagMax50.12 / 43.5752.16 / 42.0571.85 / 58.93
OPCM25.06 / 11.7215.29 / 10.3522.99 / 9.47
NUFILT51.96 / 45.4050.59 / 41.0271.15 / 58.70
WUDI-Merging49.47 / 43.0547.06 / 36.2169.46 / 57.35
DOP52.36 / 46.4650.98 / 40.4969.91 / 56.24
RPT(Ours)53.43 / 46.7058.43 / 48.7272.24 / 59.42

Both RPT scores increased with the label budget on all three datasets. Among CMM methods, RPT achieved the highest ACC in all nine settings and the highest macro-F1 in eight. Compared with Independent FT, RPT scored higher on both metrics at all three budgets on FACED and at 10% on Physionet-MI.


Q4. AM-BN in Tables 2 and 3

Thank you for pointing out the ambiguity. Both RPT rows in Table 2 report the full method, including task-vector transport and AM-BN. We clarified this in the revised caption (p. 7, lines 327–328). The upper block uses each baseline’s original normalization rule. The lower block applies AM-BN to the eight compatible baselines so that their parameter merging rules can be compared with RPT under matched normalization.

Table 3 (p. 8) isolates the effect of AM-BN by holding task-vector transport fixed. Its “w/o AM-BN” variant averages the BN affine parameters and running statistics separately.


Q5. Independent and Sequential Fine-Tuning

Independent FT starts each subject’s fine-tuning from a fresh copy of the pre-trained model. Each resulting model is evaluated on its own subject’s held-out EEG.

Sequential FT carries the fine-tuned model forward from one subject to the next. The first step starts from the pre-trained model. The final model is evaluated on the held-out EEG of all subjects in the stream.

We clarified both protocols in Section 4.2 of the revised manuscript (p. 6, lines 319–321). We use these protocols as references, following prior work on continual model merging [1–3].

[1] Yang E, Tang A, Shen L, et al. Continual model merging without data: Dual projections for balancing stability and plasticity. NeurIPS 2025

[2] Qiu Z, Wang L, Cao Y, et al. Null-space filtering for data-free continual model merging: Preserving stability, promoting plasticity. ICLR 2026

[3] Qiu Z, Xu Y, He C, et al. MINGLE: Mixture of Null-Space Gated Low-Rank Experts for Test-Time Continual Model Merging. NeurIPS 2025

AuthorsOpenReview ↗

Response to Reviewer dd2A[3/N]

W2b / Q6. Storage and computation for EEG foundation models

RPT processes one incoming checkpoint at a time, using O(P)O(P) operations and O(P)O(P) working memory per merge for a model with PP parameters, independent of the number of subjects. After each update, RPT releases the incoming checkpoint and retains the pre-trained model and the current merged model described in Algorithm 1 (p. 18). The merger therefore does not need to keep a growing collection of subject checkpoints [1].

We measured the mean merging time per incoming checkpoint and peak host RAM for both the original backbone and CBraMod [2] using four CPU threads on the same server.

DatasetOriginal Backbone time / peak RAM (s / MiB)CBraMod time / peak RAM (s / MiB)
Physionet-MI0.2972 / 1063.700.6232 / 1586.07
FACED0.3854 / 1111.540.7179 / 1825.51
ISRUC0.3978 / 1034.940.9087 / 2420.01

We also evaluated CBraMod (50 epochs of pre-training and 25 epochs of fine-tuning per subject) under the same subject-incremental setup. RPT achieved the highest final ACC and macro-F1 among the merging methods on all three datasets and led five of the six BWT comparisons. MagMax led BWT-ACC on ISRUC.

MethodPhysionet-MI ACC/MF1Physionet-MI BWT-ACC/BWT-MF1FACED ACC/MF1FACED BWT-ACC/BWT-MF1ISRUC ACC/MF1ISRUC BWT-ACC/BWT-MF1
Pre-trained47.89 / 41.18— / —39.56 / 30.67— / —61.58 / 48.82— / —
Independent FT53.36 / 48.66— / —44.18 / 34.89— / —73.08 / 57.04— / —
WA51.49 / 45.88-0.32 / -0.1841.65 / 32.56+0.74 / +0.7262.28 / 48.72-1.21 / -1.10
Task Arithmetic40.69 / 29.71-4.29 / -6.999.85 / 3.79-12.35 / -12.1521.64 / 8.42-16.16 / -16.30
TIES51.33 / 45.70-2.06 / -2.9242.06 / 32.87+0.60 / +0.6462.58 / 49.47-4.47 / -3.38
MagMax50.39 / 44.63-0.63 / -0.7641.91 / 32.64+0.30 / +0.3362.82 / 48.77-0.53 / -0.97
OPCM52.11 / 47.10+0.16 / +0.5640.29 / 31.34-1.49 / -1.3362.99 / 48.69-1.76 / -1.63
NUFILT52.27 / 46.19-1.75 / -2.3142.38 / 33.010.00 / -0.3061.32 / 46.30-7.75 / -7.42
WUDI-Merging43.51 / 36.38-7.30 / -9.1413.09 / 8.44-18.45 / -15.0062.23 / 47.78-9.47 / -7.50
DOP52.27 / 46.45-1.27 / -2.1941.76 / 32.57+0.74 / +0.0461.49 / 46.76-8.58 / -7.09
RPT53.68 / 48.26+1.59 / +1.4542.94 / 33.44+1.19 / +0.7763.87 / 49.96-0.70 / -0.72

For larger EEG foundation models, one extension would freeze a shared backbone, train a lightweight LoRA adapter [3] for each subject, and merge the resulting weight updates with RPT. Given the computational resources and time available, we leave a full evaluation on large EEG foundation models to future work.

[1] Tang A, Yang E, Shen L, et al. Merging on the fly without retraining: A sequential approach to scalable continual model merging. NeurIPS 2025

[2] Wang J, Zhao S, Luo Z, et al. Cbramod: A criss-cross brain foundation model for eeg decoding. ICLR 2025

[3] Hu E J, Shen Y, Wallis P, et al. LoRA: Low-rank adaptation of large language models. ICLR 2022


Q7. Statistical significance

We conducted an additional statistical analysis to assess RPT’s gains over the eight baselines (Table 2, p. 7). The table below reports statistical significance tests on ACC and macro-F1.

Comparison under matched AM-BNPhysionet-MI (ACC / MF1)FACED (ACC / MF1)ISRUC (ACC / MF1)
RPT vs. WA0.0020** / 0.0020**0.0020** / 0.0020**0.0020** / 0.0020**
RPT vs. Task Arithmetic0.0020** / 0.0020**0.0020** / 0.0020**0.0020** / 0.0020**
RPT vs. TIES0.0020** / 0.0020**0.0020** / 0.0020**0.0020** / 0.0020**
RPT vs. MagMax0.0020** / 0.0020**0.0020** / 0.0020**0.0020** / 0.0020**
RPT vs. OPCM0.0020** / 0.0020**0.0020** / 0.0020**0.0020** / 0.0020**
RPT vs. NUFILT0.0020** / 0.0039**0.0020** / 0.0020**0.0020** / 0.1230
RPT vs. WUDI-Merging0.0020** / 0.0020**0.0020** / 0.0020**0.0020** / 0.0020**
RPT vs. DOP0.0020** / 0.0039**0.0020** / 0.0020**0.0117* / 0.0215*

RPT differed significantly from the baselines in 47 of the 48 pairwise comparisons. The table reports the corresponding p-values (* p < 0.05, ** p < 0.01).


Thank you again for the care you brought to the review. We are grateful for the opportunity to strengthen the manuscript through your feedback and would welcome any further comments. We hope that our responses and additional results adequately address your concerns. For further details, please refer to the revised manuscript, where changes to both the main text and the appendix are marked in blue.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary and Key Evidence

Dear AC, SAC, PC and Reviewers,

We sincerely thank the reviewers for their careful assessment and constructive feedback, and the Area Chair for overseeing the review process. We appreciate the reviewers’ recognition of our approach to continual EEG learning without historical replay (dd2A, 3ccE, jnuS), method design and knowledge retention (dd2A, 3ccE), computational efficiency (3ccE, jnuS), and comprehensive controlled evaluation (3ccE).

We summarize the key clarifications and additional evidence for Resultant-Power Transport (RPT):

  • Single-model deployment and computational cost (Reviewer dd2A). Compared with Independent FT, RPT consolidates subject adaptations into one shared decoder per EEG task, avoiding the need to store and select a separate model for each subject. Compared with Sequential FT, it achieves better final decoding performance and substantially less forgetting, with subsecond mean merging time, no further training of the merged model, and auxiliary storage that does not grow with the number of subjects.

  • Label requirements (Reviewer dd2A). We clarified that supervision is used to generate the benchmark checkpoints under protocols consistent with prior continual model merging and EEG continual learning. RPT’s merging rule can also accept compatible checkpoints from unsupervised adaptation, opening a route to unlabeled subject streams.

  • Label efficiency and transfer to unseen subjects (Reviewer dd2A). Across the evaluated label budgets, RPT retained its final accuracy advantage over the merging baselines. Its fixed decoder also achieved higher accuracy and macro-F1 than the pre-trained model on held-out subjects across all evaluated datasets, without subject-specific fine-tuning.

  • Additional backbone evaluation (Reviewers dd2A and jnuS). We evaluated RPT with CBraMod, a distinct EEG backbone. RPT achieved the highest final accuracy and macro-F1 among merging methods across all evaluated datasets and led most of the backward transfer comparisons.

  • Normalization and transport contributions (Reviewers jnuS and 3ccE). BN-state replacement and checkpoint-exclusion analyses explain the FACED ablation. Controls with affine-map BN merging (AM-BN) held fixed further support a separate contribution from task-vector transport.

  • Individual retention and arrival-order sensitivity (Reviewer 3ccE). Our additional analyses showed higher mean final accuracy for early-arriving subjects and fewer large final drops across all prior subjects than NUFILT under matched AM-BN. Structured tests with clustered subjects and early or late high-shift arrivals also showed stable overall accuracy.

The revised manuscript incorporates the protocol clarifications and additional analyses of normalization and individual retention; further experiments are documented in the rebuttal discussion. We sincerely thank you for your time and consideration of the revised manuscript and full rebuttal evidence in your final assessment.

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.