Large Brain Model (LaBraM) Overview

• LaBraM is a unified EEG foundation model that segments EEG signals into fixed-length channel patches and uses vector-quantized tokenization to build a compact neural vocabulary.
• Its core architecture leverages temporal and spatial embeddings along with masked EEG modeling through Transformer layers to handle arbitrary channel configurations without explicit graph construction.
• LaBraM supports diverse applications—including abnormal detection, emotion recognition, stress classification, and motion artifact removal—by enabling transfer learning and interpretability across heterogeneous datasets.

Large Brain Model (LaBraM) most commonly denotes a unified foundation model for electroencephalography (EEG) introduced to learn generic representations from heterogeneous EEG corpora by segmenting signals into EEG channel patches, training a vector-quantized neural tokenizer through neural spectrum prediction, and pretraining neural Transformers by masked EEG modeling. In subsequent literature, the phrase “Large Brain Model” is also used more broadly: for large-scale, task-capable brain models with multi-level interpretability, for omni-modal brain-decoding systems that align EEG, MEG, and fMRI with a language-model token space, and for brain-aligned language-model programs. A central technical referent in the recent BCI literature, however, is the EEG foundation model introduced in 2024 and then adapted to stress classification, motion artifact removal, and codebook-based architectural refinement (Jiang et al., 2024, Eliasmith et al., 2016, Tang et al., 25 Feb 2026).

1. Definition and design objectives

LaBraM was proposed against a background in which EEG deep learning models were typically dataset- and task-specific because of small labeled datasets, heterogeneous formats, mismatched electrode montages, variable lengths, varied task protocols, and low signal-to-noise ratio. Its stated objective is to build a unified architecture that accepts arbitrary EEG channel configurations and sample lengths, leverages about 2,500 hours of unlabeled EEG from around 20 datasets for cross-dataset self-supervised pretraining, uses a neural tokenizer with vector quantization on spectral targets to create a compact neural vocabulary, and fine-tunes across abnormal detection, event type classification, emotion recognition, and gait prediction (Jiang et al., 2024).

The model’s basic representational move is to treat an EEG recording as $X \in \mathbb{R}^{C \times T}$, split it into fixed-length channel-wise time patches, and represent each patch as a token augmented by temporal and spatial embeddings tied to a universal 10–20 channel set. This allows the model to ingest arbitrary channel subsets without explicit graph construction. The original paper reports a patch window length $w = 200$ samples at 200 Hz, a sequence length capped at 256 patch tokens, and a sample stride of 4 seconds. The tokenizer then learns a codebook over patch representations by regressing Fourier amplitude and phase, rather than raw EEG, because directly reconstructing raw EEG failed to converge due to low SNR and nonstationarity (Jiang et al., 2024).

Subsequent papers describe LaBraM as a Large EEG Model (LEM) or large brainwave foundation model. In that usage, LaBraM is a pretrained backbone supplying transferable EEG representations, while downstream work shifts effort toward curation, preprocessing, and fine-tuning rather than task-specific architecture design. This suggests that, within BCI and EEG modeling, LaBraM functions both as a specific pretrained model family and as a template for data-centric EEG foundation modeling (2505.23042, Barmpas et al., 22 May 2025).

2. Core architecture and pretraining mechanics

In the original formulation, per-channel patch extraction is written as

$$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$

and each patch embedding is augmented by temporal and spatial embeddings: $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$ The temporal encoder is a stack of 1D temporal convolution blocks of the form Conv1D $\rightarrow$ GroupNorm $\rightarrow$ GELU, followed by an encoder-only Transformer with stabilized attention

$$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$

The published Base, Large, and Huge variants differ in depth and width rather than in overall training logic (Jiang et al., 2024).

Variant	Layers / hidden / heads	Parameters
LaBraM-Base	12 / 200 / 10	$\approx 5.8$M
LaBraM-Large	24 / 400 / 16	$\approx 46$M
LaBraM-Huge	48 / 800 / 16	$\approx 369$M

The neural tokenizer uses a codebook

$w = 200$0

with reported hyperparameters $w = 200$1 and $w = 200$2. Quantization uses nearest-neighbor assignment after $w = 200$3-normalization: $w = 200$4 The spectral targets are derived from a discrete Fourier transform. For each patch, the tokenizer regresses amplitude $w = 200$5 and phase $w = 200$6, with a VQ-VAE-style objective

$w = 200$7

Pretraining then proceeds by masked EEG modeling over tokenizer codes. With mask $w = 200$8, masked embeddings are fed through the Transformer, and a linear classifier predicts code indices: $w = 200$9 The final pretraining objective is symmetric masking,

$$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$0

with a default mask ratio $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$1 (Jiang et al., 2024).

The training stack is explicitly cross-dataset. Signals are band-passed to $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$2–$$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$3 Hz, notch-filtered at $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$4 Hz, resampled to 200 Hz, and amplitude-normalized to unit $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$5 mV. Tokenizer training uses AdamW, 100 epochs, cosine learning-rate scheduling, peak learning rate $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$6, batch size 1024, and weight decay $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$7. Masked EEG pretraining uses AdamW, 50 epochs, peak learning rate $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$8, batch size 512, weight decay $$P_{c,i} = X_{c, t_i : t_i + w - 1}, \qquad t_i = (i-1)w + 1,$$9, gradient clipping 3, and EMA weight $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$0 on $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$1 NVIDIA A800 GPUs (Jiang et al., 2024).

3. Scale, datasets, and reported performance

The pretraining corpus is reported as approximately $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$2 hours across about 20 datasets, with examples including BCI Competition IV-1, Grasp & Lift, Inria BCI Challenge, Physionet Motor Imagery, Trujillo 2017 and 2020, Siena Scalp, SPIS Resting, bi2015a, TUAR, TUEP, TUSZ, TUSL, several SEED datasets, and self-collected data. The datasets vary widely in task type, channel count, and sampling rate, which is precisely the heterogeneity the spatial-embedding scheme is designed to absorb (Jiang et al., 2024).

The original evaluation covers four downstream settings. On TUAB abnormal detection, LaBraM-Base reports Balanced Accuracy $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$3, AUC-PR $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$4, and AUROC $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$5, against the best BIOT baseline at $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$6, $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$7, and $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$8, respectively. On TUEV event type classification, LaBraM-Base reports Balanced Accuracy $$e_{c_{i_j},k} \leftarrow e_{c_{i_j},k} + te_k + se_{i_j}.$$9, Kappa $\rightarrow$0, and Weighted F1 $\rightarrow$1, compared with BIOT at $\rightarrow$2, $\rightarrow$3, and $\rightarrow$4. On SEED-V emotion recognition, LaBraM-Base reports accuracy $\rightarrow$5, Kappa $\rightarrow$6, and F1 $\rightarrow$7, improving over BIOT at $\rightarrow$8, $\rightarrow$9, and $\rightarrow$0. On MoBI gait prediction, LaBraM-Base reports Pearson correlation $\rightarrow$1, $\rightarrow$2, and RMSE $\rightarrow$3, with larger models improving to $\rightarrow$4, $\rightarrow$5–$\rightarrow$6, and RMSE $\rightarrow$7–$\rightarrow$8 (Jiang et al., 2024).

Ablation results in the same study define the model’s internal logic. The neural tokenizer improves TUEV Balanced Accuracy from $\rightarrow$9–$$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$0 without tokenizer to $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$1 with LaBraM; a mask ratio of $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$2 is reported as generally strong and stable; symmetric masking improves performance and stability, especially for larger models; removing spatial embeddings degrades TUAB Balanced Accuracy to $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$3 and TUEV Balanced Accuracy to $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$4; and training from scratch causes steep performance drops. Partial fine-tuning of the last 8–12 Transformer blocks often matches or slightly improves over full fine-tuning, whereas linear probing is substantially worse (Jiang et al., 2024).

This scaling behavior was later revisited in “LaBraM++,” which keeps the overall codebook-based foundation-model pipeline but replaces the direct phase loss with a sine/cosine representation: $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$5 The paper argues that direct phase regression is mismatched to the circular topology of phase and reports mean downstream accuracy $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$6 for LaBraM++ against $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$7 for LaBraM under matched settings; the reported task scores are Motor $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$8, Memory $$\mathrm{Attention}(Q,K,V) = \mathrm{softmax}\!\left(\frac{\mathrm{LN}(Q)\mathrm{LN}(K)^\top}{\sqrt{d_{\text{head}}}}\right)V.$$9, Sleep $\approx 5.8$0, Eyes $\approx 5.8$1 for LaBraM++, versus $\approx 5.8$2, $\approx 5.8$3, $\approx 5.8$4, $\approx 5.8$5 for LaBraM (Barmpas et al., 22 May 2025).

4. Fine-tuning, domain transfer, and task-specific extensions

LaBraM was quickly repurposed as a fine-tunable foundation EEG model. In real-world stress classification from resting-state EEG collected during an ongoing graduate class, a fine-tuned LaBraM was trained on data from 18 graduate students using 5-second windows after preprocessing with band-pass filtering between 1 and 50 Hz, ASR, ICA, automatic IC rejection, downsampling to 200 Hz, and overlap-based augmentation of the elevated-stress class only. The best-performing fine-tuned model achieves a balanced accuracy of $\approx 5.8$6 with a 5-second window. Across four random seeds, other best test balanced accuracies are $\approx 5.8$7, $\approx 5.8$8, and $\approx 5.8$9. With fixed splits, average accuracy is $\approx 46$0 with pretraining versus $\approx 46$1 without pretraining, and average accuracy falls from $\approx 46$2 with 30 channels to $\approx 46$3 with 20 random channels and $\approx 46$4 with 11 channels (2505.23042).

A different extension fine-tunes LaBraM for motion-artifact removal by integrating simultaneously recorded inertial measurement unit (IMU) data through correlation attention mapping. In that paper, the base model “vqnsp_encoder_base_decoder_3x200x12” is described as transformer-based, pretrained on 2,500 hours of heterogeneous EEG, and fine-tuned with only 5.9 hours of EEG and IMU recordings, or $\approx 46$5 of the 2500 hours used to train the base model. The fine-tuned system contains approximately 9.2 million parameters, compared with a base model of approximately 8.6 million. It operates on 1-second frames, aligns EEG and IMU features in a shared 64-dimensional latent space, and uses an Artifact Gate Layer to subtract IMU-correlated interference from EEG embeddings before decoding. Across walking and slight running, and across 10-second, 30-second, and 1-minute windows, IMU integration reduces EEG–IMU correlations by $\approx 46$6–$\approx 46$7 versus raw EEG and $\approx 46$8–$\approx 46$9 versus ASR+ICA (Zhang et al., 1 Sep 2025).

LaBraM has also served as a baseline against which more domain-specific EEG pretraining schemes are defined. In silent-speech decoding, the proposed Large Brain LLM (LBLM) explicitly cites LaBraM as a pretrained baseline and contrasts masked-reconstruction EEG pretraining with an autoregressive Future Spectro-Temporal Prediction paradigm. On the reported cross-session silent-speech benchmark, LaBraM-Base reaches $\approx 369$0 word-level accuracy and $\approx 369$1 semantic-level accuracy, whereas the specialized LBLM reaches $\approx 369$2 and $\approx 369$3, respectively. In that comparison, LaBraM functions as a generic EEG foundation model rather than a language-specialized architecture (Zhou et al., 29 Apr 2025).

5. Broader meanings of “Large Brain Model”

Outside EEG foundation modeling, “Large Brain Model” is used in a broader systems-neuroscience sense. BioSpaun, a direct extension of Spaun, is presented as exemplifying this broader ideal. Spaun contains about 2.5 million spiking neurons and roughly 8 billion synaptic connections organized into approximately 20 neuroanatomically mapped areas, receives input through a model eye, produces output with a physically modeled arm, and performs eight cognitive and perceptual-motor tasks without reconfiguration. BioSpaun increases biological plausibility by replacing subsets of leaky integrate-and-fire neurons with conductance-based, multi-compartment neurons, including a reduced pyramidal cell model with 20 compartments grouped into 4 functional regions and 9 voltage-gated channel types with 27 tuned parameters. Crucially, it preserves task performance while enabling direct linkage from sodium-channel blockade via TTX,

$\approx 369$4

to task-specific behavioral disruption. In the digit-recognition intervention, accuracy is robust up to about $\approx 369$5 Na-channel blockage and falls to chance, about $\approx 369$6, by $\approx 369$7 blockage; in an OFC-like recurrent memory circuit, about $\approx 369$8 blockage disrupts counting and list memory through task-state decay (Eliasmith et al., 2016).

A still broader extension appears in omni-modal brain decoding. NOBEL is introduced as a neuro-omni-modal brain-encoding LLM that unifies EEG, MEG, and fMRI within the semantic embedding space of Qwen2.5-Omni-7B. It uses BrainOmni as a shared EEG/MEG encoder, a dual-path fMRI strategy with Evox on raw 4D volumes and domain-specific beta encoders on GLM beta weights, modality-specific 3-layer MLP aligners, and LoRA adaptation with rank $\approx 369$9 and alpha $w = 200$00. Joint multimodal inputs outperform unimodal baselines on paired datasets: on LEMON, NOBEL reports $w = 200$01 versus $w = 200$02 for EEG-only and $w = 200$03 for fMRI-only; on Nat-View, $w = 200$04 versus $w = 200$05 and $w = 200$06; and on SMN4Lang, $w = 200$07 versus $w = 200$08 and $w = 200$09. It also reports $w = 200$10 accuracy on NSD image–brain verification and $w = 200$11 on HAD video–brain verification (Tang et al., 25 Feb 2026).

In language-model alignment work, LaBraM is sometimes treated as a target ideal rather than a fixed EEG architecture. One such paper frames a Large Brain Model as a brain-aligned LLM and reports that associative-memory augmentation and supervised fine-tuning improve LM–brain alignment. Using fMRI from the Narratives dataset, the paper reports that LLaMA-2 achieves higher brain scores than GPT-2, that associative-memory augmentation yields ROI improvements roughly from $w = 200$12 to $w = 200$13, and that instruction tuning on the 1000-sample Association dataset yields approximately $w = 200$14–$w = 200$15 gains in medial temporal lobe regions and approximately $w = 200$16–$w = 200$17 gains in the superior parietal lobule at subject level (2505.13844).

6. Limitations, evaluation debates, and open directions

The original EEG LaBraM already has explicit limitations. Its reported data scale of about $w = 200$18k hours and largest model size of about $w = 200$19M parameters remain below modern LLM and vision-model scales; full fine-tuning is resource-intensive; and the framework is unimodal EEG only. The authors note that some details, such as exact $w = 200$20 for temporal embeddings, padding for short sequences, and exact training time, are not specified. Later LaBraM++ results add a different limitation: although the modified tokenizer lowers pretraining loss and improves downstream accuracy, high-frequency reconstruction remains less accurate than low-frequency trends (Jiang et al., 2024, Barmpas et al., 22 May 2025).

Application papers add further task-level caveats. The classroom stress study uses a small, imbalanced dataset, a single classroom/cohort, self-report DASS labels without detailed thresholding, and splits formed by recordings rather than explicitly enforced subject-level separation; its test set often includes only two elevated-stress subjects, which the authors identify as a source of sensitivity to out-of-distribution samples. The IMU-based artifact-removal study is limited to walking and slight running, 11 participants, and comparisons only to ASR+ICA rather than to other advanced IMU-assisted methods; hardware, optimizer, learning rate, batch size, epochs, random seeds, and training time are not specified in that paper (2505.23042, Zhang et al., 1 Sep 2025).

For brain-alignment interpretations, a separate methodological debate is directly relevant. A critique of “brain score” evaluation shows that shuffled train–test splits can allow trivial temporal structure to inflate apparent LM–brain similarity. In the Pereira fMRI dataset, the Orthogonal Autocorrelated Sequences Model accounts for $w = 200$21 of GPT2-XL variance in EXP1 and $w = 200$22 in EXP2 under shuffled splits; for untrained GPT2-XLU, sentence position plus sentence length account for $w = 200$23 of neural predictivity in EXP1 and $w = 200$24 in EXP2. The paper argues that over-reliance on brain scores can lead to over-interpretations of similarity between LLMs and brains, and recommends contiguous or block-wise cross-validation, explicit control of autocorrelation and positional confounds, and variance partitioning rather than raw brain-score reporting (Feghhi et al., 2024).

Across these strands, the term “Large Brain Model” names a moving research program rather than a single settled artifact. In the narrow EEG sense, it denotes a vector-quantized, masked-modeling foundation model for cross-dataset EEG representation learning. In broader usage, it denotes multi-scale systems that connect molecular interventions to behavior, align multiple brain modalities to shared latent spaces, or seek confound-controlled correspondence between model states and neural data. A plausible implication is that future LaBraM research will be judged less by scale alone than by how well it combines heterogeneous data, preserves modality-specific structure, supports controlled transfer, and yields interpretable claims under rigorous evaluation (Jiang et al., 2024, Eliasmith et al., 2016, Tang et al., 25 Feb 2026).