acceptodds
Under review as a conference paper at ICLR 2027

Beyond Layout Compatibility: Understanding and Reusing Spatial Priors in EEG Foundation Models

Abstract

EEG foundation models aim to transfer knowledge from large-scale pre-training across heterogeneous acquisition settings, where positional representations link signal content to electrode location. Yet accommodating an electrode layout does not ensure effective use of pre-trained positional knowledge. We theoretically analyze how positional parameterization shapes electrode-specific learning, motivating the reuse of independently parameterized electrode vectors. Building on this analysis, we propose RADAR, a framework centered on Spatial Positional Remapping (SPR). SPR uses pre-trained electrode vectors as spatial anchors to initialize target encodings through geometric remapping, with backbone-specific interfaces enabling cross-model reuse. RADAR further incorporates Spatio-Temporal Regularization (STR) and Distributionally Robust Optimization (DRO) to examine whether positional reuse complements data- and optimization-level improvements. Across multiple EEG decoding tasks and foundation models, RADAR improves downstream performance, while SPR outperforms alternative positional encodings in most evaluated settings. Further analyses show that the learned spatial information remains useful outside its original electrode configurations and model architecture, with SPR providing additional gains under matched auxiliary training conditions. Taken together, these results show that layout-compatible models can still benefit from better use of positional knowledge, highlighting spatial prior reuse as a way to translate large-scale pre-training into greater downstream gains.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG foundation models, positional encoding, spatial prior reuse

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_6Xc9OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper looks at the issue that when you take a pre-trained EEG foundation model and apply it to a new EEG headset, such as one with 128 or 256 sensors, the model doesn't know where the new sensors are located on the scalp (as it has never seen this headset during training). Instead of training new sensor-position embeddings from scratch (which LaBraM (which is an EEG Foundation Model) normally does on unseen sensors) or interpolating the raw brainwave signals, the authors took LaBraM’s pre-trained lookup table of 90 standard sensor positions and calculated the embeddings for new sensor locations by taking a distance-weighted average of the 4 closest known sensors. They also plugged this same 90-sensor lookup table into other foundation models (CBraMod, CSBrain, LUNA, REVE). Finally, they bundled this with two standard training tricks, channel dropout + Gaussian noise and reweighting losses of random mini-batch subgroups, and called the whole package RADAR. The paper further presents a first-order analysis contrasting independent embedding tables with shared coordinate encoders, along with stability and approximation bounds for the remapped initialization.

For validation the authors use a cross-subject protocol (5 folds × 3 seeds) on three high-density datasets (HGD, MAMEM1, and CCD for both classification and regression) and five low-density benchmarks. Within each backbone, the authors compare the native positional encoding, a REVE-style coordinate encoding, and SPR. SPR improves average classification on LaBraM (62.5 to 69.6 balanced accuracy), CBraMod, CSBrain, and LUNA, and is roughly neutral on REVE. Ablations cover frozen versus trainable embeddings, random and unaligned pre-trained tables, signal-level and ROI remapping, alternative remapping rules, and a linear probe that decodes electrode coordinates from the remapped embeddings.

Critical strengths
  1. The method is simple and practical, which makes it interesting and quite useful to consider for other models as well. SPR reuses LaBraM’s pre-trained electrode table to initialize embeddings for unseen electrodes through spherical distance-weighted interpolation. It requires no additional pre-training. Because the remapping acts only on positional parameters, all recorded high-density channels are retained. On LaBraM, SPR outperforms signal-level remapping to LaBraM’s 90 positions (69.6 vs. 66.3 average balanced accuracy for Signal-NN, Table 11 --> average column).

  2. Informative analysis of positional content is very insightful, and the authors give a clear input into what is happening. The Random PE, Unaligned Pretrained PE, and SPR controls help separate the contributions of pre-trained content and geometric alignment. By comparing SPR against Random PE and Unaligned Pretrained PE controls (Section 4.3), they cleanly separate the benefits of independent vector parameterization, pre-trained content, and geometric alignment. Furthermore, the linear coordinate probing experiment against Random and Shuffled controls (Section 4.5, Table 8) demonstrates that LaBraM’s pre-trained positional table encodes linearly decodable 3D scalp geometry (7.18∘7.18^\circ–8.36∘8.36^\circ error vs. >49∘>49^\circ for controls). This is consequential because it validates the paper's core premise: pre-trained electrode tables contain genuine, transferable spatial structure rather than merely serving as arbitrary channel identifiers.

Critical weaknesses
  1. While Table 1 reports a +7.1 percentage point average gain on LaBraM (62.5% to 69.6%) and +1.6 to +2.2 points on CBraMod, CSBrain, and LUNA, I have found three experimental confounds that weaken the claim that these gains stem from transferring pre-trained spatial knowledge, I present this as one weakness as it all connects to these numbers presented in table 1.
  • The LaBraM baseline is artificially low. Appendix E.8 reveals that "Native LaBraM" replaces the entire positional table with random vectors on high-density datasets, even on HGD, whose standard 10-20 electrode names partially overlap with LaBraM's vocabulary. Furthermore, Table 13 shows that simple nearest-neighbor copying (K=1) already recovers nearly the entire gain on HGD (0.664 to 0.723, vs. 0.727 for SPR) and most of the gain on MAMEM1 (0.447 to 0.541, vs. 0.579 for SPR), leaving only ~0.4 and ~3.8 points attributable to the distance-weighted interpolation itself.
  • Integrating SPR into non-LaBraM models modifies their architectures (Appendix E.8), removing positional convolutions in CBraMod and CSBrain and bypassing LUNA’s coordinate branch. Table 5 shows that simply applying this architectural change with random independent vectors (Random PE) already accounts for a large fraction of the gain over Native (CBraMod: 0.586 Native --> 0.590 Random PE --> 0.603 SPR; CSBrain: 0.570 --> 0.581 --> 0.586; LUNA: 0.501 --> 0.526 Random PE vs. 0.523 SPR, where pre-trained content provides zero benefit). Meanwhile, on the strongest coordinate-based backbone (REVE), SPR slightly degrades performance (0.659 to 0.656).
  • For HGD (14 subjects) and MAMEM1 (11 subjects), each of the 5 folds holds out only a single test subject (Appendix E.1). Thus, only 5 subjects per dataset are ever evaluated (repeating 3 seeds adds no independent test data), and the cross-backbone gains (~1-2 points) are an order of magnitude smaller than the fold standard deviations, with no statistical significance tests reported.
  1. The paper packages a straightforward engineering heuristic (spherical 4-NN average of lookup-table embeddings) in heavy abstraction and pairs it with unrelated training tricks that do not consistently help. One thing that stood out to me was that the paper writing is quite dense and the prose can be heavily simplified to improve readability. Throughout the paper, simple operations are obscured behind inflated terminology (e.g., "functionally leveraged," "better unleashing the decoding potential," "preserving high-fidelity spatial semantics") rather than stating the concrete mathematical operations upfront. An example of this is the fact that elementary derivations are framed as a "theoretical basis". Equations 1, 4, and 6 are standard textbook properties, a first-order chain-rule step under Euclidean SGD (whereas all experiments use AdamW, which preconditions coordinates), the ℓ∞\ell_\infty non-expansiveness of a row-stochastic matrix, and the triangle inequality under Lipschitz continuity with unobservable reference discrepancy δ(tj)\delta(t_j). Moreover, Appendix B.1 explicitly concedes that Eq. 1 imposes no performance ordering between tables and coordinate generators, contradicting Section 3.1's use of Eq. 1 to motivate choosing independent tables.
Questions for authors
  • Appendix E.8 states that Native LaBraM replaces the entire positional table with random vectors on high-density datasets, yet Appendix E.6 notes that HGD uses standard electrode names that partially overlap with standard montages. Why does the Native LaBraM baseline not retain name-matched pre-trained embeddings for the overlapping standard 10-20 channels on HGD? What is LaBraM's performance on HGD when overlapping channels retain their pre-trained vectors and only unseen channels are randomly initialized?

  • Please confirm whether Random PE in Table 5 uses the exact same modified integration interface as SPR (i.e., with positional convolutions removed in CBraMod/CSBrain and the coordinate branch bypassed in LUNA, Appendix E.8). Given that Random PE already accounts for most of the gain on CSBrain (0.570 to 0.581 vs. 0.586 for SPR) and outperforms SPR on LUNA (0.526 vs. 0.523), and given that only 5 test subjects are evaluated across the 5 folds on HGD and MAMEM1, can the authors provide paired statistical significance tests (e.g., across all subjects via leave-one-subject-out, or across the existing folds) comparing SPR directly against Random PE?

  • Table 6 shows that adding STR+DRO to SPR degrades average classification performance on LUNA (0.523 for SPR vs. 0.509 for RADAR) and that full RADAR underperforms Native + STR + DRO on REVE (0.654 vs. 0.691). Furthermore, Appendix E.3 states that DRO hyperparameters are tuned per task on validation data, giving RADAR a larger hyperparameter search budget than SPR or Native baselines. Finally, Appendix C.3 acknowledges that randomly partitioning a shuffled mini-batch does not target meaningful subpopulations and relates the objective to tilted ERM. Can the authors please clarify (a) whether SPR/Native baselines received an equivalent hyperparameter tuning budget, and (b) why random-batch tilted ERM is framed as "Distributionally Robust Optimization" rather than evaluated as true Group DRO over the known training subject identities?

  • This is more of a future work request here, but Signal-NN and Signal-KNN4 (Table 11) compress 128/256 channels down to 90 channels, confounding interpolation quality with channel reduction. Comparing against classic spherical spline interpolation (Perrin et al., 1989) and evaluating whether SPR's gains persist on Large/Huge foundation model checkpoints would be valuable extensions for future work.

11 comments
AuthorsOpenReview ↗

Response to Reviewer 6Xc9 (Part 1/7): Full-Cohort LOSO Evaluation [W1, Q1, Q2]

We are deeply grateful for your time and effort, as your exceptionally thorough and rigorous review has substantially strengthened our work. Below, we first present the full Leave-One-Subject-Out (LOSO) results on HGD and MAMEM1, and then address each specific concern and question in turn.

1. Full-Cohort LOSO Evaluation and Revised HGD Baseline [W1, Q1, Q2]

"...only 5 test subjects are evaluated across the 5 folds on HGD and MAMEM1, can the authors provide paired statistical significance tests (e.g., across all subjects via leave-one-subject-out, or across the existing folds) comparing SPR directly against Random PE?"

Following your suggestion, we introduce a new full-cohort Leave-One-Subject-Out (LOSO) protocol with 14-fold evaluation on HGD (N=14N=14) and 11-fold evaluation on MAMEM1 (N=11N=11) across all five backbones under matched settings, so that every subject serves once as the held-out test set. In each LOSO fold, one subject serves as the test set, one subject serves as the validation set, and the remaining N−2N-2 subjects form the training set, yielding an (N−2)(N-2)/1/1 subject split (12/1/1 on HGD and 9/1/1 on MAMEM1). Each run is repeated across three random seeds (42, 43, and 44) and averaged within each fold before computing cross-subject statistics. Below we report balanced accuracy as mean ± subject-level SD (%) and subject-level two-sided pp-values formatted as paired tt-test / continuity-corrected Wilcoxon signed-rank test.

DatasetModelNativeRandom PESPRp vs. Nativep vs. Random
HGD (N=14)LaBraM73.70 ± 7.18*70.43 ± 8.6276.34 ± 7.160.002/0.002<0.001/0.001
CBraMod69.41 ± 5.3368.84 ± 6.6371.02 ± 6.370.015/0.0280.006/0.008
CSBrain69.59 ± 6.3868.36 ± 7.6170.89 ± 6.470.012/0.0210.003/0.008
LUNA66.75 ± 6.9867.84 ± 8.2969.31 ± 8.170.013/0.0070.110/0.079
REVE67.78 ± 8.4262.91 ± 6.5468.53 ± 6.710.614/0.4900.005/0.017
MAMEM1 (N=11)LaBraM58.73 ± 26.46†58.73 ± 26.4664.69 ± 25.980.001/0.0050.001/0.005
CBraMod44.85 ± 18.0345.57 ± 21.1350.70 ± 19.640.017/0.0450.018/0.029
CSBrain45.62 ± 19.0945.93 ± 22.1149.20 ± 21.450.016/0.0290.211/0.423
LUNA32.17 ± 10.1032.57 ± 10.3034.82 ± 11.530.318/0.2660.298/0.284
REVE60.87 ± 23.2556.68 ± 22.5062.70 ± 24.650.597/0.6560.301/0.625

*HGD Native LaBraM retains pretrained vectors for all 78 name-matched channels and randomly initializes the 50 unseen channels. †MAMEM1 (GSN-256) has zero name overlap with LaBraM's 10-20 table, so Native equals Random PE.

Notably, once evaluated under the stricter full-cohort LOSO protocol, the minor 5-fold fluctuations disappear and SPR achieves the best mean performance across every single backbone and dataset without exception.

AuthorsOpenReview ↗

Response to Reviewer 6Xc9 (Part 2/7): HGD Baseline & PE Interface [W1, Q1, Q2]

2. Addressing the Three Confounds in Table 1 [W1, Q1, Q2]

2.1 Revised HGD Baseline and K=4 vs. K=1 Remapping [W1, Q1]

"Why does the Native LaBraM baseline not retain name-matched pre-trained embeddings for the overlapping standard 10-20 channels on HGD? What is LaBraM's performance on HGD when overlapping channels retain their pre-trained vectors and only unseen channels are randomly initialized? ... Furthermore, Table 13 shows that simple nearest-neighbor copying (K=1) already recovers nearly the entire gain on HGD ... and most of the gain on MAMEM1..."

Our initial pipeline reinitialized the full table whenever a layout exceeded LaBraM's standard set. We agree that retaining overlapping channels on HGD is a fairer Native baseline and will adopt it in the revision. Under LOSO, retaining the 78 matched channels lifts LaBraM from 70.43% to 73.70%, while interpolating the 50 unseen channels via SPR further raises accuracy to 76.34% (p = 0.002/0.002, improving 13 of 14 subjects). Even against a strong baseline that already preserves every available standard anchor, SPR still extracts consistent gains across nearly every subject from the remaining unseen electrodes, demonstrating that it reliably extends pretrained spatial priors to new sensor locations.

For Table 13 under our original 5-fold evaluation, both K=1 and K=4 retain exact source vectors at matched electrodes, differing only at unseen locations. While K=1 snaps each unseen electrode discontinuously to a single nearest anchor regardless of direction, K=4 triangulates across four surrounding anchors via a parameter-free distance softmax with zero extra trainable parameters. On HGD, 78 of 128 channels match source anchors identically under both rules and the remaining 10-5 sites sit tightly within the 10-20 grid, leaving limited room between K=1 and K=4 even though K=4 still leads. On MAMEM1, however, GSN-256 sensors deviate substantially from the 10-20 montage with zero exact matches, where single-anchor copying suffers spatial quantization error (54.1%) and uniform 4-neighbor averaging blurs local geometry (53.7%). Distance-weighted K=4 raises accuracy to 57.9%, outperforming K=1 by 3.8 percentage points and uniform K=4 by 4.2 points (accounting for nearly 30% of the total 13.2-point improvement over the 44.7% baseline). That K=1 already captures coarse anchor transfer as the simplest case of geometric reuse does not diminish the clear value of distance-weighted K=4, which consistently leads across layouts and delivers a substantial 3.8-point gain on non-standard 256-channel montages at zero extra parameter or training cost.

2.2 Interface Confirmation and Pretrained Spatial Content [W1, Q2]

"Please confirm whether Random PE in Table 5 uses the exact same modified integration interface as SPR ... Table 5 shows that simply applying this architectural change with random independent vectors (Random PE) already accounts for a large fraction of the gain over Native..."

We confirm that Random PE in Table 5, Table 9, and the LOSO table uses the exact same modified interface as SPR (Appendix E.8), removing positional convolutions in CBraMod and CSBrain and bypassing LUNA's coordinate branch under identical parameter counts and training settings.

Under the exact same modified interface and parameter count, full-cohort LOSO confirms that switching to random independent vectors (Random PE) brings no reliable gain over Native, whereas initializing that interface with SPR's remapped priors improves accuracy across all five backbones on both datasets. Specifically, replacing Native modules with Random PE slightly lowers accuracy on CBraMod and CSBrain on HGD, degrades REVE on both datasets, and barely moves MAMEM1 elsewhere, showing that an independent-vector interface alone cannot unlock high-density decoding. Initializing that same interface with SPR overturns the minor 5-subject fluctuations in Table 5 and delivers up to 6 percentage points of gain over Random PE. Consistent with our low-density benchmarks where SPR beats Random PE in 24 of 25 pairs (Table 9), SPR turns mere layout compatibility into effective spatial prior reuse without extra pretraining.

AuthorsOpenReview ↗

Response to Reviewer 6Xc9 (Part 3/7): Statistical Power & STR/DRO [W1, W2, Q2]

2.3 Cross-Subject Variance vs. Paired Statistical Power [W1, Q2]

"...and the cross-backbone gains (~1-2 points) are an order of magnitude smaller than the fold standard deviations, with no statistical significance tests reported."

Subject-level standard deviations reflect intrinsic inter-subject differences rather than method instability. While HGD subjects are relatively stable (subject-level SD around 5% to 9%), MAMEM1 exhibits extreme subject-to-subject spread (subject-level SD around 10% to 26%), where Subject 9 exceeds 90% accuracy whereas on Subject 8 every model and method barely exceeds 30% and most hover near 20%. Evaluating all variants on the exact same held-out subjects cancels these large subject-specific offsets. Once paired by subject, SPR improves mean performance across all 20 baseline comparisons in the LOSO table, reaching dual parametric and non-parametric significance in 11 of 12 comparisons across the three electrode-level backbones without target-layout pretraining exposure (LaBraM, CBraMod, and CSBrain) and in 8 of 10 comparisons across all five backbones on HGD.

Overall, 13 of the 20 comparisons are significant under both tests, and the remaining cases on REVE, LUNA, and MAMEM1 reflect clear pretraining, architectural, and sample-size boundaries rather than ineffective transfer. On REVE, SPR achieves a highly significant +5.62-point gain over Random PE on HGD (p=0.005/0.017p = 0.005/0.017) and a +6.02-point gain on MAMEM1, confirming that SPR effectively injects valid spatial priors into REVE, whereas the narrower margin over Native simply reflects REVE's stronger 25,000-subject pretraining with direct exposure to HGD and HBN GSN-128 layouts (Appendix E.6), which already endows Native REVE with target-layout priors (and SPR still edges it out on both datasets). On LUNA, compressing 128 or 256 sensors into fixed latent queries via early cross-attention (Section 2) alongside weaker baseline capacity (~32% on 5-class MAMEM1) attenuates fine-grained positional differences, although SPR still significantly beats Native on HGD (p=0.013/0.007p = 0.013/0.007). Finally, MAMEM1's small cohort (N=11N=11, 1,081 trials total in Table 2) inherently caps paired statistical power even when mean gains are substantial, such as +6.02 points on REVE and +3.27 points on CSBrain.

3. Practical Motivation and Modular Role of STR and DRO [W2]

"The paper packages a straightforward engineering heuristic (spherical 4-NN average of lookup-table embeddings) in heavy abstraction and pairs it with unrelated training tricks that do not consistently help."

Our inclusion of STR and DRO is driven directly by practical downstream adaptation needs rather than bundling unrelated tricks to inflate a headline score. As described in Section 3.3, we introduce them for two concrete practical reasons. First, when adapting EEG foundation models in practice, researchers rarely rely on a single modification in isolation and routinely apply other means of regularization to improve decoding. Evaluating SPR alongside STR and DRO tests whether SPR maintains its spatial advantage when established data- and optimization-level techniques are already active, rather than merely substituting for them, which Table 6 confirms through clear incremental gains over Native + STR + DRO on LaBraM, CBraMod, and CSBrain.

Second, EEG heterogeneity is a fundamental practical bottleneck that spans layout mismatch, signal perturbations, and distribution shifts simultaneously. Combining SPR, STR, and DRO addresses this single root problem from the three orthogonal dimensions of model representation, data regularization, and optimization. Because these three components are strictly orthogonal and decoupled, practitioners are never locked into an all-or-nothing package and can freely enable or disable each module according to their backbone and task. Rather than assuming every auxiliary module helps universally, we report exhaustive ablations across all five backbones to equip researchers with rich empirical priors, clear selection criteria, and direct validation of where each component adds value in practice.

Reviewer_mRZ9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RADAR, a framework centered on Spatial Positional Remapping, to address spatial knowledge transfer across different electrode layouts in EEG foundation models. SPR transfers pretrained positional embeddings from LaBraM to target electrode layouts through geometry-based interpolation, with backbone-specific interfaces enabling cross-model reuse. RADAR further incorporates Spatio-Temporal Regularization and Distributionally Robust Optimization to improve downstream performance. Experiments demonstrate the effectiveness of SPR across multiple EEG foundation models and downstream tasks.

Critical strengths

1、The paper highlights that compatibility with different electrode layouts does not necessarily imply effective reuse of pretrained spatial knowledge. By introducing SPR, the authors provide a valuable direction for transferring spatial representations across electrode layouts and model architectures.

2、The authors analyze the differences in gradient updates between independently learnable positional embeddings and shared coordinate-based encodings, illustrating how parameterization affects electrode-specific learning. They further establish initialization stability and local interpolation error bounds for SPR, providing theoretical support for the proposed approach.

Critical weaknesses

1、The core SPR method essentially performs distance-weighted interpolation of pretrained positional embeddings from LaBraM. While simple and effective, its technical novelty is relatively limited. Moreover, STR and DRO are existing training techniques with limited direct connection to the central problem of spatial prior transfer. The three components appear more like a combination of independent techniques than a unified framework motivated by a common research objective.

2、The proposed SPR method exclusively uses pretrained positional embeddings from LaBraM as its spatial prior. Although the authors compare against REVE-PE, the two approaches differ in both positional parameterization and transfer mechanism, making it difficult to assess the general applicability of SPR to different pretrained spatial representations. Evaluating spatial priors from other EEG foundation models would help determine whether the observed improvements depend on LaBraM-specific embeddings or reflect a more general transfer mechanism.

3、Different EEG tasks may rely on distinct functional brain regions, such as motor and visual areas. Further analysis of task-dependent positional representations and their relationships with functional brain regions would help explain the performance gains of SPR and strengthen the neuroscientific significance of the work.

Questions for authors

Can the spatial mapping matrix W in SPR be updated during downstream training? W is only used to initialize the target positional embeddings, which are subsequently fine-tuned independently, is the main advantage of SPR simply a better initialization compared with random initialization? Furthermore, electrode relationships may depend not only on physical distance but also on functional brain regions and task-specific neural activity. Could the authors further justify the use of fixed geometry-based weights and investigate whether learnable interpolation weights could provide additional improvements?

9 comments
AuthorsOpenReview ↗

Response to Reviewer mRZ9 (Part 1/5): W1

We sincerely thank you for your thoughtful and well-rounded review, whose constructive suggestions have guided us to broaden both the generality and neurophysiological depth of our study. Below, we address each concern and question in turn.

W1. What is SPR's technical contribution, and what unifies RADAR?

The core SPR method essentially performs distance-weighted interpolation of pretrained positional embeddings from LaBraM. While simple and effective, its technical novelty is relatively limited. Moreover, STR and DRO are existing training techniques with limited direct connection to the central problem of spatial prior transfer. The three components appear more like a combination of independent techniques than a unified framework motivated by a common research objective.

Our core contribution is formulating pretrained electrode embeddings as portable, cross-backbone spatial anchors, advancing EEG foundation model adaptation from layout compatibility to transferable spatial prior reuse. Guided by Section 3.1, SPR selects independently learned electrode vectors to preserve decoupled signal-location associations, using distance-weighted interpolation only as a parameter-free operator to extend anchors to unseen sites while retaining exact matches. Crucially, SPR goes well beyond interpolating LaBraM's table. It improves cross-backbone decoding on five low-density benchmarks without any interpolation (Section 4.2, Appendix Table 9) and transfers effectively when replacing LaBraM with EEGPT anchors (see W2 below), making pretrained spatial priors reusable across lookup-table, convolutional, and coordinate-based architectures without extra pretraining.

Mechanistically, SPR shifts layout alignment from signal space to positional-parameter space and separates parameter-free geometric initialization from decoupled electrode-wise fine-tuning. Unlike signal-level remapping, which projects and mixes recorded waveforms onto a lower-channel reference montage, SPR adapts positional representations to the observed layout while keeping all native EEG channels intact, outperforming signal-interpolation and ROI baselines (Section 4.2, Appendix Tables 11 and 12). Keeping geometric remapping parameter-free and restricted to initialization is likewise intentional. Learning W confines target vectors to frozen-anchor mixtures and underperforms SPR (see Q3 below), whereas SPR pairs topography-preserving initialization (Table 8) with unconstrained electrode-specific fine-tuning.

At the framework level, RADAR centers on SPR to address practical adaptation to heterogeneous EEG across orthogonal representation, data, and optimization dimensions, functioning as a modular framework rather than a rigid bundle. As detailed in Section 3.3, we incorporate STR and DRO from two practical perspectives. First, real-world EEG adaptation faces layout mismatch, signal perturbations, and uneven sample difficulty simultaneously. Pairing representation-level prior reuse (SPR) with data-level perturbation robustness (STR) and optimization-level group reweighting (DRO) targets this shared bottleneck across three complementary axes. Second, practitioners rarely modify positional encodings in isolation and routinely apply regularization during fine-tuning. Evaluating SPR alongside established STR and DRO techniques directly tests whether spatial prior transfer still delivers non-redundant gains when data and optimization regularizers are already active, rather than merely substituting for them (confirmed in Section 4.4, Table 6).

Importantly, RADAR is not meant to be applied as a mandatory all-or-nothing package. Because SPR, STR, and DRO are strictly decoupled, researchers can freely enable or omit each module depending on their backbone and task. Rather than assuming every auxiliary module helps universally, we report systematic ablations across all five backbones to provide a comprehensive empirical reference, enabling researchers to freely select the configuration best suited to their setting in practice.

AuthorsOpenReview ↗

Response to Reviewer mRZ9 (Part 2/5): W2

W2. Does SPR generalize to spatial priors beyond LaBraM?

The proposed SPR method exclusively uses pretrained positional embeddings from LaBraM as its spatial prior. Although the authors compare against REVE-PE, the two approaches differ in both positional parameterization and transfer mechanism, making it difficult to assess the general applicability of SPR to different pretrained spatial representations. Evaluating spatial priors from other EEG foundation models would help determine whether the observed improvements depend on LaBraM-specific embeddings or reflect a more general transfer mechanism.

We conducted additional experiments using EEGPT's pretrained 512-dimensional electrode embeddings as SPR's source anchors in place of LaBraM's. We retain SPR's geometry-guided remapping mechanism and transfer the resulting positional representations to CBraMod and CSBrain. This directly evaluates another pretrained spatial prior within the same transfer framework.

To keep embedding dimensionality strictly controlled without mixing widths across donors, we compare EEGPT-based SPR with a randomly initialized 512-dimensional target positional table under the same positional interface and training settings. Beyond our original 5-fold splits, our new experiments adopt a newly introduced full-cohort Leave-One-Subject-Out (LOSO) protocol with 14-fold evaluation on HGD (N=14N=14) and 11-fold evaluation on MAMEM1 (N=11N=11), where every subject serves once as the held-out test set under an (N−2)(N-2)/1/1 train/validation/test subject split per fold (12/1/1 on HGD and 9/1/1 on MAMEM1), alongside our 5-fold release splits on CCD (9/1/1 across 11 releases). Each run is repeated across three random seeds (42, 43, and 44) and averaged within each fold. We also report each backbone's native positional configuration to assess the benefit over its original spatial representation, reporting balanced accuracy (%) for classification and NRMSE for regression as mean ± subject-level SD (release-level on CCD).

TaskReceiving backboneNativeRandom target tableEEGPT-based SPR
HGD ↑CBraMod69.41 ± 5.3369.96 ± 5.7270.85 ± 6.18
HGD ↑CSBrain69.59 ± 6.3868.63 ± 7.3870.89 ± 6.61
MAMEM1 ↑CBraMod44.85 ± 18.0347.09 ± 21.4848.45 ± 19.99
MAMEM1 ↑CSBrain45.62 ± 19.0943.58 ± 19.3446.29 ± 20.52
CCD-class ↑CBraMod72.60 ± 3.4073.12 ± 2.8273.99 ± 2.92
CCD-class ↑CSBrain70.14 ± 2.8069.68 ± 3.2970.44 ± 3.04
CCD-reg ↓CBraMod0.961 ± 0.0140.962 ± 0.0110.955 ± 0.011
CCD-reg ↓CSBrain0.963 ± 0.0140.963 ± 0.0130.961 ± 0.013

EEGPT-based SPR outperforms both the native configuration and the matched random target table across all eight settings, demonstrating that SPR reflects a general spatial transfer mechanism rather than relying on LaBraM-specific embeddings. Relative to the native models, it improves classification accuracy by up to 3.60 percentage points and reduces regression error on both backbones. These gains show that transferred spatial anchors improve upon the receiving models' own positional representations beyond providing a better alternative to random initialization. Crucially, replacing LaBraM with EEGPT preserves these benefits through the same geometry-guided transfer mechanism, confirming that SPR makes effective use of spatial knowledge learned by a different source model. Compared with 200-dimensional LaBraM-based SPR, EEGPT-based SPR reaches comparable performance on HGD and CCD, while LaBraM-based SPR remains stronger overall with a clearer advantage on 256-channel MAMEM1. SPR therefore offers a general way to reuse pretrained positional knowledge across models, allowing the source of the spatial prior and the receiving signal backbone to be chosen separately.

AuthorsOpenReview ↗

Response to Reviewer mRZ9 (Part 3/5): W3 & Q2

W3 and Q2. How do SPR's positional representations relate to task-specific neurophysiology?

Different EEG tasks may rely on distinct functional brain regions, such as motor and visual areas. Further analysis of task-dependent positional representations and their relationships with functional brain regions would help explain the performance gains of SPR and strengthen the neuroscientific significance of the work.

Furthermore, electrode relationships may depend not only on physical distance but also on functional brain regions and task-specific neural activity.

SPR bridges physical sensor geometry and task-specific functional brain regions by supplying geometry-aligned spatial anchors at initialization while allowing independent electrode-wise fine-tuning to adapt to task-dependent neural activity. To verify this neurophysiologically, we relate model positional sensitivity to task-specific physiological patterns in MAMEM1 (visual stimulation, 11 subjects) and HGD (executed movements, 14 subjects) using separately fine-tuned LaBraM models under leave-one-subject-out evaluation.

For each held-out subject and task class, we compare an EEG physiological map (stimulus-frequency and harmonic SNR for MAMEM1, and movement-related suppression of relative μ/β power against rest for HGD) with a model sensitivity map measuring how sensitive the true-class prediction margin is to each electrode's positional vector. Sensitivity is computed from the absolute gradient with respect to a multiplicative gate on that vector evaluated at a gate value of one while keeping the EEG and model weights unchanged, taking absolute gradients before averaging windows within each trial and then averaging trials of the same true class. Electrode-wise Spearman correlations are averaged within each subject over the five stimulus classes for MAMEM1 or the three movements and two frequency bands for HGD, giving equal weight to subjects and comparing initializations via paired subject-level differences with subject-level bootstrap confidence intervals. The MAMEM1 source controls retain the same geometric remapping with either randomized source vectors or shuffled vector-electrode assignments.

TaskControl initializationControl mean ρSPR mean ρPaired improvement in ρ [95% CI]
MAMEM1Random target0.1140.206+0.092 [0.037, 0.149]
MAMEM1Random source0.0640.206+0.142 [0.095, 0.187]
MAMEM1Shuffled source0.1060.206+0.100 [0.047, 0.158]
HGDRandom target0.1590.263+0.104 [0.045, 0.164]

The physiological interpretation differs appropriately between the two tasks. In MAMEM1, the measured SSVEP response is stronger over occipital than frontal electrodes by an average of 2.11 dB, consistent with the visual origin of the response. SPR's positional sensitivity more closely follows this measured spatial distribution than all three controls. In HGD, the improvement is strongest for β-band suppression, a characteristic correlate of sensorimotor processing. Mean correlation increases from 0.173 to 0.322, with a paired improvement of 0.149 [0.083, 0.221].

A complementary regional perturbation analysis further links the SPR model's predictions to sensorimotor topography. Keeping the EEG and model weights unchanged, we shuffled positional-vector assignments among eight electrodes in each scalp region, averaging the effects over ten permutations per region within each subject. Shuffling positional assignments on HGD reduces balanced accuracy by 4.23 and 4.31 percentage points in the left and right central sensorimotor regions, versus only 0.08 and 0.27 points in frontal and occipital regions, directly linking SPR's correspondence to task-relevant functional topography.

Together, these results connect SPR's positional representations to the physiological patterns relevant to each decoding task. Stronger alignment with visual responses and movement-related rhythms supports the functional relevance of the transferred spatial anchors after task-specific fine-tuning. In particular, the MAMEM1 source controls show that pretrained content and correct electrode correspondence contribute beyond geometry alone, providing a physiological basis for SPR's downstream utility.

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

This paper investigates whether pretrained positional representations can be more effectively reused in EEG foundation models. The authors argue that a model’s ability to accommodate a particular electrode layout does not necessarily mean that it effectively exploits the spatial information learned during pretraining. To address this issue, they propose Spatial Positional Remapping (SPR), which constructs positional representations for new electrode locations through a distance-weighted combination of nearby positional vectors. The authors further combine SPR with Distributionally Robust Optimization (DRO) to form the RADAR framework, examining whether positional reuse provides additional benefits beyond those attributable to data and optimization.

Critical strengths

The paper raises a valuable research question and presents relatively comprehensive experiments supporting the distinction between compatibility with an electrode layout and effective use of pretrained spatial information.

Critical weaknesses

However, the existing experiments do not sufficiently disentangle whether the performance gains arise from pretrained spatial knowledge, geometry-based initialization, or changes in the parameterization and integration of positional representations.

Questions for authors
  1. The paper states that SPR uses learnable projections when necessary to interface with other models, whereas PRVE-PE uses a fixed dimensionality. These different adaptation strategies may introduce an unfairness into the comparison. Could the authors clarify how they ensure a fair comparison between these approaches?
  2. SPR uses both pretrained positional vectors and geometric relationships between electrodes, whereas Random PE removes both sources of information simultaneously. Consequently, SPR’s superior performance over Random PE does not establish which source is responsible for the improvement.
  3. Do the positional vectors need to be assigned to their corresponding electrodes, or would using the same set of pretrained vectors still be beneficial even if they were assigned to incorrect electrodes?
2 comments
AuthorsOpenReview ↗

Response to Reviewer XdMY (Part 1/2): Q1 & Q2

We sincerely appreciate your encouraging assessment of our work and your sharp, insightful questions that go straight to the core of disentangling our empirical gains. Below, we address each question in turn.

Q1. Is the dimensional adaptation comparison fair?

The paper states that SPR uses learnable projections when necessary to interface with other models, whereas PRVE-PE uses a fixed dimensionality. These different adaptation strategies may introduce an unfairness into the comparison. Could the authors clarify how they ensure a fair comparison between these approaches?

We use the adaptation strategy that performs better overall for each positional representation. Our comparison evaluates both SPR and REVE-PE with either fixed dimension selection or a learnable projection updated during fine-tuning. Fixed selection does not freeze the positional encoding itself, whose parameters are still updated during fine-tuning.

To evaluate every subject, our new experiments adopt a full-cohort Leave-One-Subject-Out (LOSO) protocol with 14 folds on HGD (N=14N=14) and 11 folds on MAMEM1 (N=11N=11), where every subject serves once as the held-out test set under an (N−2)(N-2)/1/1 train/validation/test split per fold (12/1/1 on HGD and 9/1/1 on MAMEM1). Each run is repeated across three random seeds (42, 43, and 44) and averaged within each fold. Below we report balanced accuracy as mean ± subject-level SD (%).

DatasetBackboneSPR (fixed)SPR (learned)REVE-PE (fixed)REVE-PE (learned)
HGDCBraMod70.60±6.1971.02±6.3763.16±16.9959.23±19.10
HGDCSBrain70.35±5.0870.89±6.4769.93±7.3069.65±6.72
MAMEM1CBraMod44.44±20.5450.70±19.6421.20±4.3222.18±4.93
MAMEM1CSBrain45.66±20.0649.20±21.4542.43±12.9039.46±13.80
Mean57.7660.4549.1847.63

Our default setup assigns each method its empirically stronger adapter, and SPR outperforms REVE-PE in every setting under both matched fixed selection and matched learnable projection. Specifically, learnable projection improves SPR in all four settings by 2.69 percentage points on average, whereas fixed selection improves REVE-PE by 1.55 points over learnable projection, showing that fixed selection favors rather than handicaps REVE-PE. Even when locked to the exact same adapter, SPR consistently surpasses REVE-PE across all four settings (57.76% vs. 49.18% mean under fixed selection and 60.45% vs. 47.63% under learnable projection), confirming that SPR's advantage does not stem from using a learnable adapter.

Q2. Do the gains come from pretrained vectors or geometric remapping?

SPR uses both pretrained positional vectors and geometric relationships between electrodes, whereas Random PE removes both sources of information simultaneously. Consequently, SPR's superior performance over Random PE does not establish which source is responsible for the improvement.

We added source-table controls retaining the exact same geometric remapping while varying the source vectors. Let W denote the geometric remapping matrix and P the pretrained source table. We compare:

  • SPR (WP): remap pretrained vectors with their correct electrode correspondences.
  • Random-source remapping (WR): replace the pretrained source table with a random table R, retaining the same source geometry and remapping matrix W.
  • Shuffled-source remapping (WΠP): permute the vector-electrode assignments in P, retaining the same vector set and remapping matrix W.

All three initialize an independently fine-tuned target positional table under identical receiving backbones, interfaces, and training settings. Thus, WP versus WR tests the contribution of pretrained vectors beyond geometric interpolation, while WP versus WΠP tests the contribution of their electrode correspondence.

Below we report balanced accuracy as mean ± subject-level SD (%) under the same 14-fold HGD and 11-fold MAMEM1 LOSO protocol.

DatasetBackboneRandom-sourceShuffled-sourceSPR
HGDLaBraM71.54±7.5273.70±6.6676.34±7.16
HGDCBraMod69.00±7.0470.44±6.5271.02±6.37
HGDCSBrain69.95±5.3570.36±5.2170.89±6.47
HGDLUNA66.93±7.9268.03±6.9969.31±8.17
HGDREVE64.71±7.7261.86±11.8768.53±6.71
MAMEM1LaBraM60.20±27.4758.89±24.2864.69±25.98
MAMEM1CBraMod46.87±21.7644.89±19.9350.70±19.64
MAMEM1CSBrain46.76±19.5647.10±19.3949.20±21.45
MAMEM1LUNA32.54±11.9634.13±8.3734.82±11.53
MAMEM1REVE58.79±24.2561.10±24.6762.70±24.65

SPR outperforms both geometry-matched controls in all ten comparisons, showing that its gains extend beyond geometric interpolation alone. Specifically, its remapping makes effective use of both pretrained positional content (WP over WR) and the correspondence between that content and electrode locations (WP over WΠP).

AuthorsOpenReview ↗

Response to Reviewer XdMY (Part 2/2): Q3 & Planned Revisions

Q3. Does the assignment of pretrained vectors to electrodes matter?

Do the positional vectors need to be assigned to their corresponding electrodes, or would using the same set of pretrained vectors still be beneficial even if they were assigned to incorrect electrodes?

Preserving accurate vector-electrode correspondence is essential, as reusing the same pretrained vector set at incorrect electrodes destroys transferable scalp geometry and consistently underperforms SPR in downstream decoding. Three complementary lines of evidence establish this point.

Geometry-matched shuffled-source remapping (WΠP in Q2 above). Permuting electrode assignments while keeping the exact same pretrained vector set and remapping matrix W directly tests whether vector-set statistics alone are beneficial. In Q2, WΠP yields only inconsistent shifts over random-source remapping WR, edging ahead in some settings from shared vector statistics while degrading decoding in others such as MAMEM1 with LaBraM (58.89% vs. 60.20%). Restoring true electrode correspondence in SPR (WP) outperforms WΠP across all ten settings, including gains from 73.70% to 76.34% on HGD and from 58.89% to 64.69% on MAMEM1 with LaBraM.

Unaligned pretrained vectors under trainable and frozen settings (Section 4.3, Tables 4 and 5). In our original 5-fold evaluation, Unaligned Pretrained PE assigns LaBraM vectors by index without geometric alignment (Appendix E.8). With trainable positional embeddings (Table 5), SPR outperforms Unaligned Pretrained PE across most comparable settings, such as 72.7% versus 70.1% on HGD with LaBraM. With frozen positional embeddings (Table 4), SPR widens this margin to 4.4 percentage points (72.7% vs. 68.3%), confirming the direct value of aligned correspondence even without downstream positional updates.

Spatial probing with shuffled assignments (Section 4.5, Appendix F.3, Table 8). Keeping the same set of pretrained vectors and geometric remapping W while permuting only vector-electrode assignments (averaged over 20 permutations under the same source-only linear coordinate probe protocol), mean angular error jumps from 7.18° and 8.36° with SPR to 62.67° and 69.24° with Shuffled on GSN-128 and GSN-256. Correct electrode assignment is therefore indispensable for preserving linearly decodable spatial structure and reliable downstream gains.

Planned Manuscript Revisions

We will incorporate the above experiments and clarifications into the revised manuscript as outlined below.

  • Appendix E.8 and Appendix F. Add the fixed-versus-learned dimensional adaptation ablation comparing SPR and REVE-PE under both matched adapter choices.
  • Section 4.3 and Appendix F.1. Add the geometry-matched random-source (WR) and shuffled-source (WΠP) remapping controls across all five backbones on HGD and MAMEM1, and cross-reference them with our frozen/trainable positional controls (Tables 4 and 5) and linear coordinate probing (Section 4.5, Table 8).

We are finalizing the revised manuscript PDF and will upload it promptly. Thank you again for your insightful feedback and constructive questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revised Manuscript and Core Contributions

We sincerely thank all the reviewers for their constructive feedback and time. We have uploaded the revised manuscript PDF, with all new experiments, analyses, and wording refinements highlighted in blue. Guided by your suggestions, the revision incorporates seven new experimental and neurophysiological studies verifying the generality, causal sources, and functional grounding of our approach.

1. Key Manuscript Revisions

  • Full-cohort LOSO and paired tests (Sections 4.1–4.3, Table 2, Appendices E.1, E.8, F.1). Added 14-fold HGD (N=14N=14) and 11-fold MAMEM1 (N=11N=11) LOSO results across all five backbones with paired tt-tests and Wilcoxon tests, adopting the 78-channel name-matched HGD Native LaBraM baseline.
  • Disentangling content, correspondence, and interfaces (Section 4.3, Table 2, Appendices E.8, F.3, F.8, F.10, Tables 8, 18, 21). Added Random-Source and Shuffled-Source controls (Table 2), Fixed vs. Learned cross-backbone adapter ablations (Table 8), Signal-Spline comparisons (Table 18), and Fixed-W vs. Learned-W ablations alongside the K=4K=4 vs. K=1K=1 discussion (Appendix F.10, Table 21).
  • Cross-donor generalization (Sections 2, 4.3, Appendix F.2, Table 7). Evaluated 512-d EEGPT electrode embeddings as source anchors on CBraMod and CSBrain across four tasks.
  • Task-specific neurophysiology (Section 4.5, Appendix F.6, Tables 13, 14). Added Spearman correlations between positional gate-gradient sensitivity and empirical SSVEP SNR (MAMEM1) or μ/β\mu/\beta suppression (HGD), plus regional scalp permutations across five ROIs.
  • Modular RADAR framing, DRO validation, and prose (Abstract, Sections 1, 3, 4.4, 5, Appendices B, C.3, E.3, F.4, G, Table 11). Clarified decoupled module roles and matched-budget comparisons, added Random DRO vs. Subject Group DRO (Table 11), analyzed AdamW update coupling (Appendix B.1), noted Large/Huge scaling (Appendix G), and streamlined prose to state concrete operations upfront.

2. Core Strengths and Takeaways of Our Work

  • A consequential perspective and principled design beyond layout compatibility. As highlighted across all three reviews (XdMY, mRZ9, 6Xc9), our work establishes that accommodating heterogeneous electrode inputs does not guarantee effective reuse of pretrained spatial priors, formulating pretrained electrode tables as portable spatial anchors across layouts, donors, and heterogeneous architectures. To our knowledge, this is the first systematic study in EEG foundation models to decouple spatial prior transfer from downstream positional learning, revealing why lightweight geometric initialization followed by unconstrained electrode-wise adaptation surpasses more heavily parameterized mappings (such as Learned-W and REVE-PE) and offering actionable design principles to the community.
  • Simple, effective, zero-pretraining spatial transfer that preserves native EEG. SPR requires no extra pretraining, carries clean analytical motivation (mRZ9), and keeps all native channels intact without signal compression (6Xc9). Across layouts and prior donors, it consistently improves mean decoding, achieving statistically significant LOSO gains on electrode-level backbones while showing why unconstrained positional updates matter beyond frozen transfer alone. Linear 3D coordinate probing (6Xc9) and neurophysiological alignment (mRZ9) further corroborate the spatial and functional fidelity of the transferred priors.
  • Designed for real-world downstream adaptation with rich empirical guidance. In real-world downstream applications, practitioners commonly combine multiple adaptation techniques during fine-tuning. RADAR reflects this practical setting as a modular, decoupled framework across representation (SPR), data (STR), and optimization (DRO), verifying that spatial prior reuse continues to deliver non-redundant gains alongside active regularizers rather than substituting for them. Because these modules are not rigidly bundled, our comprehensive experiments provide extensive empirical evidence for the community to reference and guide practical deployment.

Thank you once again for helping us substantially strengthen this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.