Pith. sign in

REVIEW 3 major objections 6 minor 26 references

Stacked LoRA on EEG foundation models separates shared motor-imagery patterns from each subject’s neural quirks and wins most accuracy head-to-heads.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 04:56 UTC pith:TOYZLFP7

load-bearing objection Clean PEFT study that shows dual-path (global + subject) LoRA helps most EEG foundation-model MI setups, with a useful population-size finding, but the pooled split softens the inter-subject claims. the 3 major comments →

arxiv 2607.03094 v1 pith:TOYZLFP7 submitted 2026-07-03 cs.LG

Stacked LoRA for Subject-Adaptive EEG Foundation Models in Motor Imagery Decoding

Aymen Sarhane , Fouad Lbakali , Mouad Souissi , Jonathan Lys , Giulia Lioi This is my paper
classification cs.LG
keywords EEGBrain-computer interfaceMotor imageryFoundation modelLow-rank adaptationParameter-efficient fine-tuningSubject variability
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Motor-imagery brain–computer interfaces still need costly per-person recalibration because people’s EEG patterns differ so much. Even large pretrained EEG foundation models cannot be frozen and used as-is; they need adaptation. This paper shows that a dual low-rank update—one Global adapter trained on everyone plus a small Subject-Specific adapter per person—structurally decouples the shared task signal from individual variability. Across three foundation models and three motor-imagery datasets (including a clinical patient set), the stacked version reaches the best or tied-best accuracy in seven of nine settings. The right mix of global versus personal adapters depends on the population: a shared adapter suffices for large diverse cohorts, while subject-specific adapters become decisive when clinical inter-session noise is high.

Core claim

The authors claim that Stacked LoRA—adding a Global low-rank adapter shared across all subjects together with one Subject-Specific low-rank adapter per subject on a frozen EEG foundation-model backbone—mitigates inter-subject variability in motor-imagery classification more effectively than either path alone, attaining the best or tied-best balanced accuracy in the large majority of backbone–dataset combinations, with the optimal balance shifting according to cohort size and clinical variability.

What carries the argument

Stacked LoRA: at each adapted layer the low-rank update is the sum of a Global adapter (Ag, Bg) trained on all subjects and a hard-routed Subject-Specific adapter (Ak, Bk) that fires only for trials of subject k; the backbone stays frozen.

Load-bearing premise

The evaluation uses one global random 70/15/15 split of pooled trials from all subjects and sessions, then filters those same partitions by subject; if that split already leaks subject-specific structure into the shared train set, the reported gains of the subject-specific adapters are inflated.

What would settle it

Re-run the identical Stacked-versus-Global comparison under a strict leave-one-subject-out or leave-one-session-out protocol on the same three datasets; if Stacked LoRA no longer beats Global LoRA on BCI IV-2a and Zuo2025, the claimed benefit of the subject-specific path disappears.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes Stacked LoRA, a dual-path parameter-efficient fine-tuning scheme for EEG foundation models in motor imagery decoding. At each adapted layer the low-rank update is the sum of a Global adapter trained on all subjects and a hard-routed Subject-Specific adapter (Eq. 2). Three strategies—Global LoRA, Subject-Specific LoRA, and Stacked LoRA—are compared against linear probing on frozen REVE, LaBraM, and LUNA backbones across BCI Competition IV-2a, PhysioNet MI, and the clinical Zuo2025 dataset. Table 1 reports that Stacked LoRA is best or tied-best in seven of nine backbone–dataset combinations; the authors further argue that a shared adapter suffices for large diverse cohorts while subject-specific adapters dominate under high clinical inter-session variability. Parameter footprints at train and deploy time are quantified in Table 2, and limitations (no zero-shot subjects, fixed ranks, within-dataset splits only) are stated explicitly.

Significance. Subject-adaptive PEFT for large EEG foundation models is a timely and practically relevant problem: even strong pretrained backbones fail as frozen extractors, and full per-subject fine-tuning is costly. The dual-path design cleanly separates shared task structure from residual individual signatures, is instantiated as a drop-in module on three distinct architectures, and is evaluated on both healthy and clinical MI data with transparent trainable-parameter accounting. The population-dependent guidance (when to prefer Global vs Subject-Specific vs both) is a useful practical takeaway if the empirical patterns hold under stronger evaluation protocols. The work is incremental relative to DELoREAN and subject-conditioned layers, but the systematic multi-backbone, multi-dataset comparison and the clinical benchmark add concrete value for the BCI community.

major comments (3)
  1. Section 4 (Preprocessing and Data Splitting) and the central claim of mitigating inter-subject variability: all results rest on a single global random stratified 70/15/15 split of pooled multi-subject trials, then filtered into subject-level loaders. Trials from the same subject (and, for multi-session sets, potentially the same session) can appear in both train and test. This is a within-subject held-out-trial protocol, not leave-one-subject-out or cross-session evaluation. Because Subject-Specific adapters are trained only on that subject’s training trials, residual within-subject / within-session correlations can inflate the reported Stacked gains over Global LoRA (e.g., REVE on BCI IV-2a: 80.3% vs 76.4% in Table 1). The paper itself flags the lack of cross-session validation as a limitation (§7), which confirms that the evaluation is the softest load-bearing support for the populatio
  2. Section 2.5 and Eq. (2): ranks (rg=32, rs=8) and scales (αg, αs) are fixed for all backbones and datasets with no sensitivity analysis. The claimed complementarity of the two paths—and therefore the recommendation of when Stacked is preferable—depends on these free parameters. A modest rank/scale ablation (or at least a statement that results are stable under a small grid) is needed before the population-dependent design rule can be treated as robust rather than tied to one hyperparameter choice.
  3. Table 1, PhysioNet rows for LaBraM and LUNA: Stacked LoRA underperforms pure Global LoRA (59.9 vs 60.2; 51.9 vs 55.1). The discussion correctly notes diminishing returns of per-subject adapters on large cohorts, but does not analyse whether the subject-specific path actively interferes (e.g., via capacity or optimisation) or is merely neutral. A short diagnostic—accuracy when the subject-specific branch is zeroed at inference, or cosine similarity of the two low-rank updates—would strengthen the claim that the paths learn complementary subspaces rather than that stacking is simply “never much worse.”
minor comments (6)
  1. Throughout the manuscript (abstract, §1, §2): numerous missing spaces and concatenated words appear (“proposeaframework”, “usinga singlefine-tuningstage”, “structurallydecouples”, etc.). These look like PDF extraction artefacts but must be cleaned for camera-ready text.
  2. §3: LUNA is cited as [24] in one place and [25] in the reference list; LaBraM and LUNA share the same bracket in the introduction. Fix citation numbering.
  3. Eq. (2): the indicator routing is clear, but it would help the reader to state explicitly that at inference only the matched subject adapter is loaded (as later quantified in Table 2), so that the N-fold training cost does not imply N-fold deployment cost.
  4. Table 2 caption is dense; a short sentence in the main text restating that the deployed Stacked model uses only head + Global + one rank-8 adapter would improve readability.
  5. §5.1: “As shown These improvements” — incomplete sentence; fix grammar.
  6. Related work: Klein et al. (subject-conditioned layers) and DELoREAN are appropriately cited; a one-sentence contrast with multi-task LoRA variants already mentioned (MTLoRA) on whether subject identity is treated as a “task” would tighten the positioning.

Circularity Check

0 steps flagged

No circularity: empirical PEFT comparison on held-out splits; dual-path design is an application of external dual-adapter ideas, not a self-referential derivation.

full rationale

The paper’s load-bearing claims are empirical accuracies (Table 1) for three LoRA strategies (Global, Subject-Specific, Stacked) on three public MI datasets and three frozen backbones. Equation (2) defines Stacked LoRA as the sum of a shared low-rank update and a hard-routed per-subject update; that is an architectural definition, not a derivation that equates a predicted quantity to a fitted input. Training is standard supervised fine-tuning with early stopping on validation accuracy; test metrics are measured on held-out partitions and are not forced by construction from any fitted constant. Self-citations (REVE as backbone; prior clinical MI work) supply models or background and do not underwrite uniqueness or force the ranking of strategies. The dual-path idea is attributed to external DELoREAN/MTLoRA-style work, not smuggled as a uniqueness theorem from the same authors. There is no self-definitional loop, no fitted parameter renamed as prediction, and no renaming of a known empirical law. Evaluation-design concerns (pooled split vs LOSO) affect correctness risk, not circularity. Score 0; steps empty.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The work is empirical PEFT; load-bearing choices are the dual-path architecture, fixed ranks/scales, frozen backbones, and the global stratified split. No new physical entities or free constants fitted to produce a claimed law; free parameters are the usual PEFT hyperparameters.

free parameters (4)
  • Global LoRA rank rg = 32
    Fixed to 32 for all experiments without per-dataset or per-backbone search; directly controls capacity of the shared path.
  • Subject-Specific LoRA rank rs = 8
    Fixed to 8; controls capacity of each personal adapter and total trainable parameters.
  • LoRA scaling coefficients αg, αs
    Present in Eq. 2; values not reported as tuned, treated as fixed hyperparameters shared across strategies.
  • Learning rate and training schedule = 1e-4, 25 epochs, patience 5
    AdamW 1e-4, cosine with 10% warmup, max 25 epochs, early stopping patience 5; chosen by hand and shared.
axioms (4)
  • domain assumption Low-rank adapters of modest rank can capture residual subject-specific EEG variability once a frozen foundation model and a global adapter are present.
    Core modeling premise of Stacked LoRA (Section 2.3); inherited from LoRA/DELoREAN literature and not independently verified for EEG.
  • domain assumption Hard subject-identity routing (indicator 1{x∈Sk}) is available at both training and inference.
    Eq. 2 and Limitations section; without subject ID the subject-specific path cannot be used.
  • ad hoc to paper A single global random stratified 70/15/15 split of pooled multi-subject trials, then filtered per subject, yields valid estimates of subject-adaptive performance without leakage.
    Section 4 Data Splitting; non-standard relative to leave-one-subject-out BCI practice and load-bearing for the reported gains.
  • standard math Standard LoRA forward pass and PEFT training dynamics (Eq. 1).
    Section 2.2; taken from Hu et al. 2022.
invented entities (1)
  • Stacked LoRA (Global + Subject-Specific dual-path adapter) no independent evidence
    purpose: Structurally decouple population-level task knowledge from per-subject neural signatures inside each adapted layer.
    Named framework of the paper; direct repurposing of DELoREAN’s dual-path idea from tasks to subjects. No independent evidence outside the reported accuracy tables.

pith-pipeline@v1.1.0-grok45 · 15707 in / 2981 out tokens · 23574 ms · 2026-07-12T04:56:47.746537+00:00 · methodology

0 comments
read the original abstract

Electroencephalography (EEG) decoding for brain-computer interfaces (BCIs) faces a major challenge: substantial inter-subject variability limits effective cross-subject generalization. Consequently, practical systems still rely largely on subject-specific models trained from scratch and requiring individual recalibration. EEG foundation models have recently emerged as a promising alternative; however, even large pretrained models cannot simply be used as fixed feature extractors and still require additional adaptation before they can be reliably applied to downstream tasks. In this work, we address this challenge through targeted adaptation strategies. Building on recent EEG foundation models such as REVE, LaBraM, and LUNA, we examine the impact of different low-rank adaptation strategies on motor imagery classification. We propose a framework that structurally decouples subject-invariant knowledge from subject-specific neural signatures: the low-rank update at each adapted layer is split into a Global adapter, trained jointly across all subjects, and Subject-Specific adapters, each absorbing individual variability. To assess the contribution of each path, we compare three adaptation strategies: (i) subject-specific LoRA (ii) global LoRA and (iii) stacked LoRA, combining both Global and Subject Specific adapters. Experiments on BCI Competition IV-2a, PhysioNet Motor Imagery, and the clinical Zuo2025 benchmark show that Stacked LoRA effectively mitigates inter-subject variability, achieving the best accuracy in the large majority of backbone and dataset combinations. Our analysis further reveals that the optimal balance between the global and subject-specific paths depends on the target population: a shared adapter is sufficient for large, diverse cohorts, whereas subject-specific adaptation is decisive under the high inter-session variability of clinical recordings.

Figures

Figures reproduced from arXiv: 2607.03094 by Aymen Sarhane, Fouad Lbakali, Giulia Lioi, Jonathan Lys, Mouad Souissi.

Figure 1
Figure 1. Figure 1: Stacked LoRA framework: the frozen backbone (pretrained EEG foundation model) is augmented with a shared Global Adapter (Ag, Bg) and N Subject-Specific Adapters (Ak, Bk). Global adapter. Since (Ag, Bg) receives gradients from every trial in the batch, its optimum depends on the mixture of all subject distributions rather than on any single subject distribution Sj . It therefore converges toward the portion… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 1 canonical work pages

  1. [1]

    and Valdes-Sosa, P

    Mushtaq,F.,Welke,D.,Gallagher,A.,Pavlov,Y.G.,Kouara,L.,Bosch-Bayard,J., ... and Valdes-Sosa, P. (2024). One hundred years of EEG for brain and behaviour research. Nature human behaviour, 8(8), 1437-1443

    2024

  2. [2]

    Frontiers in Neuroscience6, 55 (2012).https: //doi.org/10.3389/fnins.2012.00055

    Tangermann, M., Müller, K.-R., Aertsen, A., Birbaumer, N., Braun, C., Brunner, C., Leeb, R., Mehring, C., Miller, K.J., Müller-Putz, G.R., Nolte, G., Pfurtscheller, G., Preissl, H., Schalk, G., Schlögl, A., Videaurre, C., Waldert, S., Blankertz, B.: Review of the BCI Competition IV. Frontiers in Neuroscience6, 55 (2012).https: //doi.org/10.3389/fnins.2012.00055

    Pith/arXiv arXiv doi:10.3389/fnins.2012.00055 2012

  3. [3]

    Frontiers in Human Neuroscience14, 37 (2020).https://doi.org/10.3389/fnhum.2020.00037

    Lioi, G., Butet, S., Fleury, M., Bannier, E., Lécuyer, A., Bonan, I., Barillot, C.: A multi-target motor imagery training using bimodal EEG-fMRI neurofeedback: a pilot study in chronic stroke patients. Frontiers in Human Neuroscience14, 37 (2020).https://doi.org/10.3389/fnhum.2020.00037

    doi:10.3389/fnhum.2020.00037 2020

  4. [4]

    IEEE Trans- actions on Biomedical Engineering51(6), 1034–1043 (2004).https://doi.org/ 10.1109/TBME.2004.827072

    Schalk, G., McFarland, D.J., Hinterberger, T., Birbaumer, N., Wolpaw, J.R.: BCI2000: A general-purpose brain-computer interface (BCI) system. IEEE Trans- actions on Biomedical Engineering51(6), 1034–1043 (2004).https://doi.org/ 10.1109/TBME.2004.827072

    doi:10.1109/tbme.2004.827072 2004

  5. [5]

    Lotte, F., Congedo, M., Lécuyer, A., Lamarche, F., and Arnaldi, B. (2007). A re- view of classification algorithms for EEG-based brain–computer interfaces. Journal of neural engineering, 4(2), R1-R13

    2007

  6. [6]

    and Wu, D

    Liu, D., Chen, Y., Chen, Z., Cui, Z., Wen, Y., An, J., ... and Wu, D. (2026). EEG Foundation Models: Progresses, Benchmarking, and Open Problems. arXiv preprint arXiv:2601.17883

    arXiv 2026

  7. [7]

    16000–16009

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalablevisionlearners.In:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition (CVPR), pp. 16000–16009. IEEE, New Orleans (2022)

    2022

  8. [8]

    Curran Associates, Red Hook (2025).https://arxiv

    El Ouahidi, Y., Lys, J., Thölke, P., Farrugia, N., Pasdeloup, B., Gripon, V., Jerbi, K., Lioi, G.: REVE: A foundation model for EEG adapting to any setup with large- scalepretrainingon25,000subjects.In:AdvancesinNeuralInformationProcessing Systems (NeurIPS 2025). Curran Associates, Red Hook (2025).https://arxiv. org/abs/2510.21585

    arXiv 2025

  9. [9]

    arXiv preprint arXiv:2605.08495 (2026)

    Banville, H., d’Ascoli, S., Dahan, S., Rapin, J., Careil, M., Benchetrit, Y., King, J.R.: NeuralBench: A unifying framework to benchmark NeuroAI models. arXiv preprint arXiv:2605.08495 (2026)

    Pith/arXiv arXiv 2026

  10. [10]

    In: International Confer- ence on Learning Representations (ICLR)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: International Confer- ence on Learning Representations (ICLR). (2022).https://arxiv.org/abs/2106. 09685

    2022

  11. [11]

    In: NeurIPS 2025 Workshop: Foun- dation Models for the Brain and Body (2025)

    Klein, T., Minakowski, P., Sager, S.: Mitigating subject dependency in EEG de- coding with subject-specific low-rank adapters. In: NeurIPS 2025 Workshop: Foun- dation Models for the Brain and Body (2025)

    2025

  12. [12]

    Mtlora: A low-rank adaptation approach for efficient multi-task learning,

    A. Agiza, M. Neseem, and S. Reda, “Mtlora: A low-rank adaptation approach for efficient multi-task learning,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 196–16 205

    2024

  13. [13]

    Fine-tuning with reserved majority for noise reduction,

    Jiang et al., “Fine-tuning with reserved majority for noise reduction,” in The Thirteenth International Conference on Learning Representations, 2025. [Online]. https://openreview.net/forum?id=ZV7CLf0RHK Stacked LoRA for Subject-Adaptive EEG Decoding 13

    2025

  14. [14]

    In: 34th European Signal Processing Conference (EUSIPCO)

    Balykov, A., Preuss, K., Boukli-Hacene, G., Gripon, V., Cardinaux, F., Mauch, L.: DELoREAN: Disposable experts for low-rank efficient adaptation to novelty. In: 34th European Signal Processing Conference (EUSIPCO). IEEE (2026)

    2026

  15. [15]

    Institute for Knowledge Discovery (Laboratory of Brain-Computer Interfaces), Graz University of Technology (2008)

    Brunner, C., Leeb, R., Müller-Putz, G.R., Schlögl, A., Pfurtscheller, G.: BCI Com- petition 2008 – Graz data set A. Institute for Knowledge Discovery (Laboratory of Brain-Computer Interfaces), Graz University of Technology (2008)

    2008

  16. [16]

    In: Advances in Neural Information Processing Systems, vol

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems, vol. 30, pp. 5998–6008. Curran Associates, Long Beach (2017)

    2017

  17. [17]

    In: Advances in Neural Information Processing Systems, vol

    Zhang, B., Sennrich, R.: Root mean square layer normalization. In: Advances in Neural Information Processing Systems, vol. 32, pp. 1–12. Curran Associates, Van- couver (2019)

    2019

  18. [18]

    arXiv preprint arXiv:2002.05202 (2020).https://arxiv.org/abs/2002.05202

    Shazeer, N.: GLU variants improve transformer. arXiv preprint arXiv:2002.05202 (2020).https://arxiv.org/abs/2002.05202

    Pith/arXiv arXiv 2002

  19. [19]

    In: International Conference on Learning Representations (ICLR)

    Dao, T.: FlashAttention-2: Faster attention with better parallelism and work parti- tioning. In: International Conference on Learning Representations (ICLR). Vienna (2024).https://arxiv.org/abs/2307.08691

    Pith/arXiv arXiv 2024

  20. [20]

    Journal of Neural Engineering15(6), 066011 (2018).https://doi.org/10

    Jayaram, V., Barachant, A.: MOABB: trustworthy algorithm benchmarking for BCIs. Journal of Neural Engineering15(6), 066011 (2018).https://doi.org/10. 1088/1741-2552/aadea0

    2018

  21. [21]

    arXiv preprint arXiv:2404.15319 (2024)

    Chevallier, S., Carrara, I., Aristimunha, B., Guetschel, P., Sedlar, S., Lopes, B., Velut, S., Khazem, S., Moreau, T.: The largest EEG-based BCI reproducibility study for open science: the MOABB benchmark. arXiv preprint arXiv:2404.15319 (2024)

    Pith/arXiv arXiv 2024

  22. [22]

    Circulation101(23), e215–e220 (2000).https://doi.org/10

    Goldberger, A.L., Amaral, L.A.N., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mietus, J.E., Moody, G.B., Peng, C.K., Stanley, H.E.: PhysioBank, Phys- ioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation101(23), e215–e220 (2000).https://doi.org/10. 1161/01.CIR.101.23.e215

    2000

  23. [23]

    Scientific Data12, 1451 (2025).https: //doi.org/10.1038/s41597-025-05767-2

    Zuo, C., Yin, Y., Wang, H., Zheng, Z., Ma, X., Yang, Y., Wang, J., Wang, S., Huang, Z.-g., Ye, C.: Enhancing classification of a large lower-limb motor imagery EEG dataset for BCI in knee pain patients. Scientific Data12, 1451 (2025).https: //doi.org/10.1038/s41597-025-05767-2

    work page doi:10.1038/s41597-025-05767-2 2025

  24. [24]

    In: International Conference on Learning Rep- resentations (ICLR) (2024)

    Jiang, W., Zhao, L., Lu, B.: Large brain model for learning generic representations with tremendous EEG data in BCI. In: International Conference on Learning Rep- resentations (ICLR) (2024)

    2024

  25. [25]

    In: Advances in Neural Infor- mation Processing Systems (NeurIPS) (2025)

    Döner, B., Ingolfsson, T.M., Benini, L., Li, Y.: LUNA: Efficient and topology- agnostic foundation model for EEG signal analysis. In: Advances in Neural Infor- mation Processing Systems (NeurIPS) (2025)

    2025

  26. [26]

    In: International Conferenceon LearningRepresentations(2019).https://openreview.net/forum? id=Bkg6RiCqY7

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conferenceon LearningRepresentations(2019).https://openreview.net/forum? id=Bkg6RiCqY7

    2019