BCI Brain Foundation Models

• BCI-integrated brain foundation models are large-capacity neural networks pre-trained in a self-supervised manner on extensive, unlabeled neural recordings to enable versatile decoding and control tasks.
• They employ techniques like neural time series tokenization, transformer-based contextualization, and masked prediction to capture spatio-temporal patterns across heterogeneous EEG and invasive data.
• These models facilitate cross-subject transfer, parameter-efficient adaptation, and real-time deployment while addressing challenges in data standardization, interpretability, and neuroprivacy.

BCI-integrated brain foundation models are large-capacity neural models that are first trained in a self-supervised fashion on massive, unlabeled neural recordings and then adapted to downstream decoding, control, monitoring, or assistive tasks in brain-computer interfaces. In the EEG-based non-invasive BCI context, a brain foundation model is defined as a large-capacity neural network that is first trained in a self-supervised fashion on massive, unlabeled EEG recordings so that generic spatio-temporal neural representations can be re-used across many downstream decoding tasks with only light fine-tuning; recent work extends this pattern from scalp EEG to invasive spike data and from end-to-end decoding to real-time monitoring, cross-subject transfer, and governance-oriented system design (Wu et al., 14 Jul 2025, Jiang et al., 2024, Ogg et al., 2 Jun 2025, Hong et al., 30 Apr 2026).

1. Conceptual scope and research trajectory

Early BCI deep learning systems were typically designed for specific datasets and applications, which limited model scale, perceptual capability, and generalizability. The shift toward brain foundation models is driven by the expectation that self-supervised pretraining on heterogeneous neural corpora can learn subject-agnostic or task-agnostic priors that survive transfer across paradigms, channel layouts, and recording conditions. LaBraM formalized this agenda for EEG by introducing a unified foundation model based on EEG channel patches, vector-quantized neural spectrum prediction, and masked neural code prediction, with pretraining on about 2,500 hours of various types of EEG signals from around 20 datasets (Jiang et al., 2024).

Subsequent work diversified the architectural and operational interpretation of the paradigm. CBraMod argued that full EEG modeling obscures the heterogeneity of spatial and temporal dependencies and proposed separate spatial and temporal attention together with conditional positional encoding (Wang et al., 2024). A HuBERT-style EEG model reframed BCI foundation modeling around minimally pre-processed, eight-channel, one-minute segments for low-profile, real-time usage, and explicitly evaluated not only canonical BCI tasks such as P300 and motor imagery but also participant identity and alpha rhythms (Ogg et al., 2 Jun 2025). DeeperBrain pushed the notion of universality further by distinguishing between end-to-end fine-tuning and frozen probing, arguing that intrinsic universality requires strong frozen-backbone transfer rather than mere fine-tuning elasticity (Wang et al., 5 Jan 2026). UniBCI extended the same logic to invasive BCIs, introducing a unified pretrained model for spike data across species, subjects, brain regions, and behavioral paradigms (Hong et al., 30 Apr 2026).

This progression indicates that “foundation model” in BCI is not restricted to scale alone. In the literature, it encompasses data standardization, channel-robust tokenization, self-supervised objectives, parameter-efficient adaptation, cross-subject transfer, latency constraints, interpretability, and increasingly explicit treatment of neuroprivacy and cognitive liberty.

2. Architectural motifs and self-supervised objectives

Representative architectures converge on tokenization of neural time series, transformer-style contextualization, and pretraining by reconstructive or masked-prediction objectives, but they differ sharply in how they encode EEG structure and BCI constraints (Jiang et al., 2024, Wang et al., 2024, Ogg et al., 2 Jun 2025, Wang et al., 5 Jan 2026, Hong et al., 30 Apr 2026, Fang et al., 18 Oct 2025).

Model	Distinctive mechanism	Pretraining objective
LaBraM	EEG-channel patching with spatio-temporal embeddings and VQ neural tokenizer	Masked EEG channel-patch neural-code prediction
CBraMod	Criss-cross transformer with separate spatial and temporal attention, plus ACPE	Patch-based masked EEG reconstruction
HuBERT-style EEG FM	Six 1D convolutions, 12-layer transformer, k-means pseudo-labels	Two-stage masked prediction over pseudo-classes
DeeperBrain	Volume conduction-aware channel encoding and neurodynamics-aware temporal encoding	Dual objective: MER + NSP
NeurIPT	Crossformer backbone, AAMP, PMoE, 3D electrode PE, IILP	Amplitude-aware masked pretraining
UniBCI	Context-conditioned spatio-temporal tokenization with metadata and IAA	Masked signals reconstruction

LaBraM uses fixed patch length $w=200$ samples, no patch–patch overlap in time, a temporal encoder consisting of three conv–GN–GELU blocks, learnable temporal and spatial embeddings, and a transformer encoder available in Base, Large, and Huge variants. Its tokenizer discretizes patch embeddings by nearest-neighbor search over a codebook with $K=8192$ and $D=64$, and the pretraining loss combines reconstruction of Fourier amplitude and phase with a VQ loss; masked EEG modeling then predicts neural codes under a mask ratio $r=0.5$ and symmetric masking (Jiang et al., 2024).

CBraMod retains patch-based modeling but replaces flattened attention with a criss-cross transformer. Spatial attention operates on “spatial stripes” $E^j \in \mathbb{R}^{C \times d}$, temporal attention on “temporal stripes” $\tilde E^i \in \mathbb{R}^{n \times d}$, and half of the heads are assigned to each branch. Its asymmetric conditional positional encoding is a depth-wise 2D convolution with kernel $(19,7)$, and masked reconstruction uses 30 s EEG segments, 1 s patches, zero mask tokens, and an $\ell_2$ loss on masked patches only (Wang et al., 2024).

The HuBERT-style EEG model begins from raw segments $X \in \mathbb{R}^{C \times T}$ with $C=8$ and $K=8192$0 samples per one-minute segment, applies a stack of six one-dimensional convolutions with kernel sizes $K=8192$1 and strides $K=8192$2, then a 12-layer transformer encoder with $K=8192$3 heads, feed-forward dimensionality 3072, and 96.4 M trainable parameters. Pretraining proceeds in two stages: spectrogram-based clustering with $K=8192$4 and masked prediction, followed by embedding-based clustering with $K=8192$5 and higher-resolution masked prediction (Ogg et al., 2 Jun 2025).

DeeperBrain explicitly embeds neuroscientific priors. Its volume conduction-aware channel encoding smooths electrode coordinates through an exponential-decay kernel over pairwise distances, while neurodynamics-aware temporal encoding uses oscillatory bases over $K=8192$6 Hz and adaptive decay bases over $K=8192$7. Pretraining minimizes

$K=8192$8

where MER reconstructs masked voltages and NSP predicts a 19-dimensional vector of order parameters including relative spectral power, PLV summaries, cross-frequency coupling, and sample entropy (Wang et al., 5 Jan 2026).

NeurIPT emphasizes heterogeneity. Its modified Crossformer alternates temporal self-attention and spatial self-attention, uses $K=8192$9, $D=64$0, merge factors $D=64$1, and replaces FFNs with progressively growing sparse MoE blocks with expert counts $D=64$2. Its amplitude-aware masking selects masked intervals from signal-amplitude percentiles rather than random time intervals, and its spatial encoding uses the real 3D coordinates of electrodes (Fang et al., 18 Oct 2025).

UniBCI adapts the foundation-model pattern to spikes through context-conditioned spatio-temporal tokenization. Raw spike trains are temporally binned, grouped into anatomical areas, linearly embedded, enriched with metadata prompts encoded by a frozen MiniLM-L6-v2, and processed by hierarchical Interval-Area Attention that combines interval linear attention within slots and area-wise sliding-window attention across pooled tokens (Hong et al., 30 Apr 2026).

3. Data scale, standardization, and representation of heterogeneity

The data regime of BCI-integrated foundation models is unusually heterogeneous: scalp EEG varies in channel count, sampling rate, referencing, task design, and signal-to-noise ratio, while invasive corpora vary by species, region, unit count, and behavioral assay. Much of the methodological literature therefore treats standardization as a first-class design problem rather than a preprocessing afterthought (Jiang et al., 2024, Wang et al., 2024, Ogg et al., 2 Jun 2025, Wu et al., 14 Jul 2025, Liu et al., 27 Jul 2025, Hong et al., 30 Apr 2026).

LaBraM assembled 2,534.8 h of EEG from about 20 sources, including motor imagery, emotion, P300, resting-state, artifact, epilepsy, and seizure datasets. All data were bandpass filtered to 0.1–75 Hz, notch filtered at 50 Hz, resampled to 200 Hz, normalized by dividing by 0.1 mV, and aligned to a universal 10–20 set, with missing channels masked by zero embeddings but distinguished by spatial embeddings (Jiang et al., 2024). CBraMod scaled further on the Temple University Hospital EEG Corpus with 69,652 recordings, 14,987 subjects, and 27,062 hours total, standardized to 19 standard 10–20 channels, resampled to 200 Hz, segmented into 30 s non-overlapping samples, and reduced to 1,109,545 usable samples after cleaning (Wang et al., 2024). The HuBERT-style EEG model instead targeted low-latency deployment and deliberately restricted itself to 8 channels from the TUH EEG corpus with 14,979 participants and more than 1.1 k days of data, resampled to 125 Hz and segmented into nonoverlapping 1 min segments without per-segment normalization (Ogg et al., 2 Jun 2025).

Benchmarking work makes the same heterogeneity explicit. AdaBrain-Bench evaluates four public EEG BFMs across 13 public datasets and seven archetypal BCI tasks spanning cognitive assessment, human augmentation, and clinical monitoring, with standardized preprocessing, adaptation strategies, and transfer settings such as cross-subject, multi-subject, and few-shot transfer (Wu et al., 14 Jul 2025). NeurIPT pretrains on 2,200 hours from 15 public EEG corpora and fine-tunes on eight benchmarks after conversion to a 20-channel bipolar “double-banana” montage (Fang et al., 18 Oct 2025). DeeperBrain uses 14 public EEG datasets totaling approximately 17,200 h and 2,438,653 one-second samples (Wang et al., 5 Jan 2026).

The literature also reveals two distinct preprocessing philosophies. One minimizes preprocessing to preserve real-time applicability, as in the eight-channel HuBERT-style model, which applies no per-session z-scoring or artifact rejection (Ogg et al., 2 Jun 2025). The other uses paradigm-specific alignment to suppress nuisance variation before representation learning. MIRepNet’s CLEAN-MI pipeline applies 8–30 Hz filtering, subject screening, a neurophysiologically informed channel template over FC, C, CP, and T regions, inverse-distance interpolation for arbitrary electrode configurations, and Euclidean Alignment to whiten second-order statistics (Liu et al., 27 Jul 2025). This suggests that BCI integration can favor either broad robustness under weak assumptions or stronger inductive bias under a known paradigm.

4. Downstream integration patterns in BCI systems

Once pretrained, brain foundation models are inserted into BCI pipelines through several recurring strategies: full fine-tuning, linear probing, parameter-efficient adaptation, dynamic source selection, frozen-feature heads, and direct deployment in real-time streaming systems (Wu et al., 28 Jul 2025, Shama et al., 29 Jan 2026, Lee et al., 1 Jul 2025, Wu et al., 14 Jul 2025, Zhou et al., 12 Aug 2025).

The HuBERT-style EEG model fine-tunes the Stage 2 encoder under leave-one-participant-out cross-validation with three conditions: full fine-tuning (“Update All”), freezing all layers except the final 256-dim embedding and output (“Update Output Only”), and de novo training. The same latent space is presented as rapidly fine-tunable to a new user’s data, which is explicitly framed as a calibration-acceleration mechanism for low-profile, real-time BCIs (Ogg et al., 2 Jun 2025).

Cross-subject motor imagery decoding introduces a more explicit integration of a pretrained BFM into adaptation logic. Using LaBraM subject embeddings $D=64$3 and $D=64$4, source subjects are selected according to the Cauchy–Schwarz divergence

$D=64$5

with a threshold $D=64$6 such as the 50th percentile of $D=64$7. Selected sources are then used in a multi-source domain adaptation objective that combines feature-level alignment via CS divergence, decision-level alignment via conditional CS divergence, and weighted source classification loss. The total loss is

$D=64$8

This pipeline makes the foundation model an upstream selector and similarity metric, not merely a feature extractor (Wu et al., 28 Jul 2025).

Continuous BCI monitoring provides another integration mode. For cognitive-load estimation, incoming EEG is filtered and resampled to 200 Hz, maintained in a 90 s sliding buffer, split into ten 16 s windows with 50% overlap, and then further chopped into one-second miniblocks for BFM encoding. With frozen encoders and batched one-second processing, one 90 s buffer is converted to features in $D=64$9 s on a standard CPU, without model pruning or quantization, satisfying a real-time latency target of $r=0.5$0 s (Shama et al., 29 Jan 2026).

Adaptation studies also show that integration need not imply full retraining. AdaBrain-Bench standardizes two adaptation strategies—full fine-tuning and linear probing—across cross-subject, multi-subject, and few-shot settings (Wu et al., 14 Jul 2025). The fine-tuning study of large brainwave foundation models pioneers LoRA for LBMs, with

$r=0.5$1

and reports that performance benefits generally emerge when adapting multiple neural network components simultaneously rather than attention alone (Lee et al., 1 Jul 2025). Cross-BCI, although not itself a pretrained foundation model, is directly relevant to deployment because it demonstrates that a lightweight, unified decoder can cover MI, P300, and SSVEP within one architecture, a property repeatedly identified as desirable for practical universal BCI systems (Zhou et al., 12 Aug 2025).

5. Empirical performance across BCI regimes

Empirical results support the utility of pretrained neural representations, but they also expose substantial variation across paradigms, adaptation settings, and efficiency criteria (Wu et al., 14 Jul 2025, Ogg et al., 2 Jun 2025, Wu et al., 28 Jul 2025, Shama et al., 29 Jan 2026, Liu et al., 27 Jul 2025, Wang et al., 5 Jan 2026, Fang et al., 18 Oct 2025).

At benchmark scale, AdaBrain-Bench reports cross-subject macro-average balanced accuracy of 64.61 for LaBraM and 62.66 for CBraMod, compared with 58.12 for Conformer, 56.73 for LDMA, 56.32 for EEGNet, and 55.64 for ST-Tran. In the multi-subject setting, LaBraM reaches 62.87% macro-average B-Acc and CBraMod 61.06%, whereas the best traditional baseline, Conformer, reaches 53.66% (Wu et al., 14 Jul 2025).

Task-specific evidence is similarly favorable. The eight-channel HuBERT-style model reports, on Dataset A, P300 AUROC of 0.69 for full fine-tuning versus 0.66 de novo; on Dataset B, imagery hands vs. feet AUROC of 0.63 for full fine-tuning versus 0.50 de novo; and, across all tasks, pretraining improves over de novo with Wilcoxon $r=0.5$2, while updating all weights outperforms freezing most layers with Wilcoxon $r=0.5$3 (Ogg et al., 2 Jun 2025). In cross-subject motor imagery, BFM-guided source selection plus CS/CCS alignment achieves 86.17 ± 5.88% average accuracy with $r=0.5$4 on BCI IV 2a and 78.41 ± 7.95% with $r=0.5$5 on the GigaDB subset; in a large-pool experiment, mean accuracy is 77.34 ± 7.46% and training time per epoch drops from 13.8 s at 30th percentile selection to 3.1 s at 10th percentile selection (Wu et al., 28 Jul 2025).

Real-time monitoring results are more modest in absolute magnitude but directly operational. In cognitive-load estimation, the best overall configuration is LaBraM features plus a linear head with Pearson correlation $r=0.5$6, outperforming PSD, EEGNet, EEGConformer, and CBraMod-based alternatives in the reported table (Shama et al., 29 Jan 2026). MIRepNet, which argues for paradigm-specific pretraining, reports average accuracies of 81.77 ± 0.27 on BNCI2014001, 81.67 ± 0.26 on BNCI2015001, and 82.36 ± 0.10 on BNCI2014004 using only 30% fine-tuning data, with fewer than 30 trials per class (Liu et al., 27 Jul 2025).

Universal-model claims are further strengthened by transfer without full adaptation. DeeperBrain reports frozen probing performance of 50.96 ± 0.37% balanced accuracy on FACED versus a best baseline of 37.76 ± 0.54, and 56.57 ± 0.73 on PhysioNet-MI versus 53.73 ± 0.39, while keeping the average fine-to-frozen drop at $r=0.5$7 rather than $r=0.5$8 for others (Wang et al., 5 Jan 2026). NeurIPT reports 55.04 balanced accuracy on BCIC-IV-2A, 67.31 on PhysioNetP300, 70.47 on Sleep-EDFx, and 67.61 on TUEV, alongside several task-specific improvements over prior foundation-model baselines (Fang et al., 18 Oct 2025).

Taken together, these results indicate that the practical value of BCI-integrated foundation models depends on the regime being optimized: cross-subject transfer, few-shot calibration, latency, frozen-backbone universality, or raw end-to-end accuracy.

6. Interpretability, governance, controversies, and open problems

Interpretability studies indicate that these models learn physiologically recognizable structure rather than only opaque task labels. In the HuBERT-style EEG model, t-SNE projections reveal clustering by participant and resting-state condition, and eyes-closed trials form tight clusters consistent with alpha-band dominance around 8–12 Hz (Ogg et al., 2 Jun 2025). In cognitive-load estimation, Partition SHAP applied to frozen BFM features shows strong, consistent importance in prefrontal ROIs and parieto-occipital ROIs for LaBraM, with longitudinal changes in which prefrontal SHAP weight increases by approximately 20% from Day 1 to Day 5 while average cognitive-load labels decline from about 0.90 to 0.64 (Shama et al., 29 Jan 2026).

At the same time, the literature contains a substantive controversy over efficiency and present-day capability. A systematic fine-tuning study concludes that state-of-the-art large brainwave foundation models achieve only marginal improvements of 0.9%–1.2% over traditional deep architectures while requiring significantly more parameters, and argues that current architectural and training inefficiencies limit their capabilities. The same work shows that LoRA can greatly reduce trainable parameters without performance degradation and that adapting multiple layer types jointly is more effective than attention-only adaptation (Lee et al., 1 Jul 2025). This directly challenges any simple equation of scale with BCI utility.

A second tension concerns universality versus paradigm specificity. MIRepNet argues that, in practical BCI deployments, the specific paradigm is generally determined prior to data acquisition and that a motor-imagery-specific foundation model with CLEAN-MI preprocessing and hybrid pretraining can substantially outperform generalized EEG models on MI tasks (Liu et al., 27 Jul 2025). DeeperBrain reaches the opposite end of the design space by seeking intrinsic universality through neuroscientific priors that preserve frozen-probe performance across more than ten tasks (Wang et al., 5 Jan 2026). A plausible implication is that future BCI systems may combine a universal pretrained backbone with paradigm-conditioned tokenization, heads, or adaptation rules rather than insisting on a single monolithic notion of “generality.”

Governance work introduces a third axis: fiduciary control. One proposal inserts a fiduciary control layer or “Guardian Model” after the foundation-model core and trains with a compound loss

$r=0.5$9

supplemented by differential privacy, secure multi-party computation, federated learning, immutable logs, dynamic consent interfaces, and external audit structures (Bhattacharjee et al., 18 Jul 2025). In BCI-integrated settings, where raw EEG or other neural signals may be interpreted and acted on in milliseconds, this reframes model design as simultaneously a decoding problem, a systems problem, and a fiduciary problem.

Open directions are recurrent across the literature: richer BCI-style pretraining tasks, hybrid EEG+EMG/EOG or broader multimodal pretraining, on-device distillation and compression, streaming or variable-length inference, explicit modeling of montage shifts, and broader demographic and clinical diversity in pretraining corpora (Ogg et al., 2 Jun 2025, Wu et al., 14 Jul 2025, Hong et al., 30 Apr 2026). The field’s central unresolved question is therefore not whether pretrained neural representations transfer at all—they do—but under what architectural priors, data regimes, adaptation rules, and governance constraints they become reliable enough for plug-and-play neural interfaces.