ScaleBrain: A New Framework for Scale-Consistent EEG Foundation Models

ScaleBrain coordinates multi-scale EEG representations from both input tokenization and reconstruction objectives, achieving top mean balanced accuracy on five BCI benchmarks with only 5.0M parameters.
ScaleBrain is a scale-consistent EEG pretraining framework for BCIs. It addresses unclear feature-scale correspondence and insufficient supervision at some scales by coordinating from both input and objective ends: multi-scale tokenization with shared temporal anchors makes scale structure inherent in the input, while scale-specific reconstruction ensures each resolution gets sufficient supervision. Axial spatio-temporal attention models spatial and temporal context separately. On six public BCI benchmarks against 12 baselines, it ranks first on five and third on one, with only 5.0M parameters.
Title
ScaleBrain: A New Framework for Scale-Consistent EEG Foundation Models
Meta Description
ScaleBrain proposes a scale-consistent pretraining framework for EEG foundation models, achieving alignment between input and target through multi-scale tokenization and scale-specific reconstruction, combined with axial spatio-temporal attention, attaining leading performance on six BCI benchmarks with only 5.0M parameters.
Article Content
EEG foundation models are emerging as an important paradigm in the brain-computer interface (BCI) field. Recent research shows that multi-scale representations based on fixed-length segments can improve EEG representation learning. However, existing frameworks typically construct temporal scales through feature-level operations, which leads to an unclear correspondence between the learned features and the scale characteristics of the raw EEG signal; meanwhile, when self-supervised objectives are inconsistent with multi-scale input representations, information at certain temporal scales may be under-supervised during pretraining. To address these issues, researchers propose ScaleBrain—a scale-consistent pretraining framework that explicitly coordinates multi-scale representations from both the model input and optimization objective ends.

Multi-Scale EEG Tokenization and Scale-Specific Reconstruction
One of ScaleBrain's core innovations is the introduction of multi-scale EEG tokenization. This method represents the same EEG signal through analysis windows of different temporal spans with shared temporal anchors, giving the model input an inherent scale structure. Unlike feature-level operations, this tokenization approach maintains an explicit correspondence between each scale and its corresponding raw signal segment, avoiding implicit mapping between features and scale characteristics.
On this basis, the framework adopts scale-specific reconstruction as the self-supervised objective. The reconstruction task aligns with the multi-scale input structure, ensuring that each temporal resolution receives sufficient supervisory signal. This solves the problem of insufficient supervision at certain scales in previous pretraining, enabling the model to learn dynamic features at different temporal scales more evenly.
The difference between multi-scale tokenization and feature-level operations: Traditional multi-scale EEG representation learning often performs operations such as pooling feature maps, kernel dilation, or cross-scale feature fusion in the deeper layers of the model. These operations occur in latent space, and the scale-transformed features no longer correspond one-to-one with the raw waveforms of a specific temporal window. Multi-scale tokenization moves the scale structure to the input end: for the same EEG segment, multiple token sequences are segmented using time windows of different lengths, with all sequences sharing the same temporal anchors (such as the same start time or center time), thereby ensuring that each token can be traced back to an explicit raw signal segment. The benefit of this approach is that reconstruction objectives can be defined separately for each scale, and supervisory signals are not diluted by feature-level scale mixing.
Axial Spatio-Temporal Attention Mechanism
To effectively fuse multi-scale representations, researchers designed an axial spatio-temporal attention mechanism. This mechanism performs context modeling on the fused multi-scale representations along both spatial and temporal dimensions, capturing spatial correlations between electrodes and temporal dynamic dependencies respectively. This decomposed attention design maintains computational efficiency while enhancing the modeling capability for the spatio-temporal structure of EEG signals.
Computational logic of axial attention: Standard self-attention performs fully-connected attention over all fused tokens, with computation growing quadratically with the number of tokens. Moreover, since the electrode dimension and temporal dimension of EEG have different semantics, direct mixed modeling can easily cause confusion. Axial attention decomposes attention into two steps: first, attention along the spatial axis (electrode dimension) allows information exchange between electrodes at each temporal position; then, attention along the temporal axis allows information exchange between temporal tokens at each electrode position. Both steps have linear complexity, avoiding the quadratic overhead of joint spatio-temporal attention. This decomposition has mature applications in video understanding and high-dimensional image generation. ScaleBrain transfers it to EEG multi-scale representation fusion, which fits well with the natural "electrode × time" grid structure of EEG.
Experimental Performance and Parameter Efficiency
ScaleBrain was extensively evaluated on six public BCI benchmark datasets and compared with 12 representative baseline methods. Results show that the model achieves the highest mean Balanced Accuracy on five datasets and ranks third on the remaining one. Particularly notable is that ScaleBrain contains only 5.0M parameters, demonstrating strong performance alongside good parameter efficiency.
This result indicates that a scale-consistent framework with coordinated design of input and objective can achieve competitive BCI decoding performance without relying on a large number of parameters, providing new ideas for the lightweight and efficient design of EEG foundation models.
Mean Balanced Accuracy: BCI datasets often suffer from class imbalance issues (e.g., unequal numbers of trials across different motor imagery classes), and ordinary accuracy can be dominated by the majority class. Balanced accuracy calculates recall for each class separately and then averages them, providing a fairer reflection of model performance on minority classes. The "mean" across multiple datasets further reduces fluctuation caused by a single data split. ScaleBrain reports this metric on six benchmarks and compares with 12 baselines, indicating that its comparisons are conducted under a class-balanced evaluation protocol, making the conclusions more informative.
One-line Summary
ScaleBrain coordinates multi-scale EEG representations from both input tokenization and reconstruction objective ends, achieving the highest mean balanced accuracy on five BCI benchmarks with only 5.0M parameters.
Paragraph Summary
ScaleBrain is a scale-consistent EEG pretraining framework for brain-computer interfaces. It addresses the issues in existing methods where temporal scales are constructed through feature-level operations, leading to unclear correspondence between features and raw signal scale characteristics, and where self-supervised objectives are inconsistent with multi-scale inputs, causing insufficient supervision at some scales. It coordinates from both the model input and optimization objective ends. At the input end, multi-scale EEG tokenization uses windows of different temporal spans and shared temporal anchors to represent the same signal, making scale structure inherent in the input; at the objective end, scale-specific reconstruction ensures that each temporal resolution receives sufficient supervision. In the fusion stage, axial spatio-temporal attention is introduced to model context along spatial and temporal dimensions respectively. Compared with 12 baselines on six public BCI benchmarks, it achieves the highest mean balanced accuracy on five datasets and ranks third on the other, with the model containing only 5.0M parameters.
Related articles

Gemini 4 Hydrogen Released: A 2.5x Faster Everyday AI Model
Google releases Gemini 4 Hydrogen, claiming 2.5x the speed of Gemini 3.8 Flash, with four-tier thinking control, a million-token context, and Orchestrator multi-model collaboration. API pricing is $1.50 per million input tokens.

Gemini 4 Hydrogen Hands-On: Google's Fastest and Cheapest AI Model
Hands-on with Gemini 4 Hydrogen: the fastest, cheapest model in the Gemini 4 family. We break down the Carbon/Argon/Hydrogen split, API pricing traps, free-tier limits, and the cost trade-offs of four thinking levels—with a real bug-debugging demo.

Gemini 4 Hydrogen vs Sonnet Comparison Test: Uses Only 1% of Weekly Quota, Performance on Par
Hands-on test: Gemini 4 Hydrogen vs Sonnet on the same CRUD task—Hydrogen used only 1% weekly quota with comparable performance and better UI details, while Anthropic's model used 7%. Developer says daily tasks don't require expensive models.