Brain-OF: Brain-Centered Frameworks

• Brain-OF is a term for diverse brain-centered frameworks that organize multimodal neuroimaging data, neural control systems, and analytical pipelines.
• It includes a multimodal foundation model pretraining on fMRI, EEG, and MEG using cross-attention and Sparse Mixture of Experts to enable robust semantic alignment.
• Other implementations address closed-loop prosthetic control, digital twin simulations, and clinical biomarker assessment to enhance neuroscience applications.

Brain-OF appears in current arXiv literature in several distinct senses rather than as a single standardized term. In one explicit usage, it denotes “the first omnifunctional brain foundation model jointly pretrained on fMRI, EEG and MEG” (Guo et al., 26 Feb 2026). In other contexts, it is interpreted as a “Brain-Oriented / Brain-Operating Fusion framework” for intelligent closed-loop neural interfaces (Fares et al., 2022), a “Brain-Operated Framework” for prosthetic control (Basit et al., 23 May 2025), a “brain-organoid-focused” framework for automated mitosis analysis (Awais et al., 2024), or a shorthand for “brain outcome factors/features” in ischemic stroke prognosis (Alhadid et al., 2024). This heterogeneity is substantive: the term recurrently marks architectures, pipelines, or biomarkers in which brain-derived signals, structures, or reserves are the organizing variable.

1. Terminological scope and principal usages

The term does not denote one canonical formalism across the literature. Instead, it names or motivates a family of brain-centered frameworks spanning multimodal representation learning, bidirectional brain–computer interfaces, embodied control, organoid analytics, prognostic biomarkers, and whole-brain simulation.

Usage domain	Representative formulation	Paper
Multimodal foundation model	“An Omnifunctional Foundation Model for fMRI, EEG and MEG”	(Guo et al., 26 Feb 2026)
Hybrid BCI architecture	“Brain-Oriented / Brain-Operating Fusion framework”	(Fares et al., 2022)
Prosthetic control	“Brain-Operated Framework”	(Basit et al., 23 May 2025)
Brain-to-fabrication	“Brain-to-Object Fabrication”	(Zhang et al., 2018)
Brain–swarm control	“Brain-Swarm Interface” as brain-driven control of collectives	(Suresh et al., 2016)
Brain organoid analytics	“brain-organoid-focused” framework	(Awais et al., 2024)
Stroke prognosis	“brain outcome factors/features”	(Alhadid et al., 2024)

A common misconception is that Brain-OF refers only to the 2026 multimodal foundation model. The published record instead shows that the label is used across substantially different methodological layers: end-to-end decoding systems, closed-loop control stacks, dataset-and-benchmark infrastructures, biomarker comparisons, and abstract theories of brain-like information processing. A plausible implication is that Brain-OF functions more as a family resemblance term than as a settled technical standard.

2. Brain-OF as an omnifunctional foundation model

In its most explicit and architecturally concrete sense, Brain-OF is a large-scale brain foundation model jointly pretrained on fMRI, EEG, and MEG, designed to support both unimodal and multimodal inputs within one framework (Guo et al., 26 Feb 2026). The central problem it addresses is the heterogeneity of functional neuroimaging: fMRI has high spatial and low temporal resolution, EEG has high temporal resolution and poor spatial localization, and MEG combines high temporal resolution with somewhat better spatial localization than EEG. Brain-OF treats these modalities as complementary and aggregates them at pretraining time to enlarge the effective corpus and reduce modality-specific over-specialization.

The input signal is represented as $S \in \mathbb{R}^{N \times T}$, segmented into temporal patches and encoded into patch embeddings $X \in \mathbb{R}^{L \times D}$. The Any-Resolution Neural Signal Sampler projects this variable-length sequence into a fixed set of latent tokens $Z \in \mathbb{R}^{C \times D}$ by cross-attention: $$\text{ARNESS}(X,Z) = \text{CrossAttn}(Q = Z,\; KV = [X; Z],\; \text{Attn\_Mask}).$$ This resolves mismatches in channel count, ROI count, and temporal extent while preserving a shared semantic space across modalities (Guo et al., 26 Feb 2026).

The backbone replaces standard self-attention with DINT attention and replaces dense feedforward blocks with a Sparse Mixture of Experts. Shared experts are intended to capture modality-invariant representations, whereas routed experts specialize in modality-specific semantics. Pretraining uses Masked Temporal-Frequency Modeling, which reconstructs both signal-domain and frequency-domain structure: $$\mathcal{L}_{\text{MTFM}} = (1-\alpha) \cdot \mathcal{L}_{\text{SL1}}(S, \hat{S}) + \alpha \cdot \mathcal{L}_{\text{SL1}}(|F_s|, |\hat{F}_s|),$$ with $\alpha = 0.8$ in the reported setup (Guo et al., 26 Feb 2026).

The pretraining corpus comprises 37 publicly available datasets, 32,278 unique participants, 5,870,984 samples, and approximately 2.8 TB of preprocessed data. Three model scales are reported: Brain-OF Base with 47.5M total parameters and about 21.5M active per pass, Brain-OF Large with 331M total and about 150M active, and Brain-OF Huge with 1.7B total and about 500M active. On downstream evaluation over 7 tasks in 11 settings, the model achieves or matches state-of-the-art in 7 of 9 tasks, and scaling from Base to Large to Huge is reported as monotonic on nearly all tasks (Guo et al., 26 Feb 2026).

The multimodal claim is not merely nominal. Brain-OF supports serial ARNESS fusion of paired modalities such as fMRI+MEG or EEG+fMRI, and the reported results show consistent gains on several fusion settings. This suggests that, in this usage, Brain-OF denotes not only a pretrained model but an explicit solution to cross-modality semantic alignment in functional neuroimaging.

3. Brain-OF as a brain-operated and closed-loop control stack

A second major usage treats Brain-OF as a brain-operated framework for perception, actuation, and feedback. In the BI-BCI literature, a natural interpretation is “Brain-Oriented / Brain-Operating Fusion framework,” in which brain-inspired AI, especially spiking neural networks running on neuromorphic hardware, is tightly coupled to invasive or non-invasive neural interfaces in closed loop (Fares et al., 2022). The canonical pipeline is explicitly bidirectional: signal acquisition, preprocessing, feature extraction or encoding, decoding, application layer, encoding or stimulation design, stimulation or feedback, and closed-loop adaptation. In that formulation, Brain-OF is effectively a brain-in-the-loop control system.

BRAVE provides a concrete prosthetic realization of this idea. It combines EEG motor-imagery decoding with automatic speech recognition for mode switching across 3 prosthetic degrees of freedom. Its EEG pipeline uses bandpass filtering at 0.5–45 Hz, ICA for artifact removal, CSP for spatial discrimination, and an ensemble of LSTM, CNN, and Random Forest models fused by a logistic-regression meta-model. The reported classification accuracy is 96% across test subjects, and the system operates in real time with a response latency of 150 ms while using voice commands such as “elbow,” “arm,” and “fingers” to select the active control mode (Basit et al., 23 May 2025). In this setting, Brain-OF denotes an operational stack from scalp EEG to embodied actuation, with a human-in-the-loop correction layer stabilizing online behavior.

Brain2Object extends the same logic from prosthetic control to fabrication. It maps visually evoked EEG to object class, then to a 3D model, then to a 3D printer. Its backend combines multi-class Common Spatial Pattern, a Dynamical Graph Representation that learns channel-to-channel spatial correlations via an adjacency matrix, and a CNN classifier. The reported recognition accuracy is 92.58% on a benchmark dataset and 75.23% on a locally collected dataset, with an online demonstrator reaching 83.3% average accuracy and approximately 2 s total latency (Zhang et al., 2018). Here, Brain-OF becomes a brain-to-object fabrication pipeline rather than a neuroprosthetic controller.

The Brain-Swarm Interface pushes the same idea to collective robotics. It uses EEG-derived Emotiv performance metrics decoded by a two-state Hidden Markov Model to control swarm aggregation versus dispersion, while EOG-like eye movements extracted from headset electrodes specify translational velocity. The system was demonstrated with a swarm of three M3pi robots in the laboratory and a swarm of 128 robots in simulation (Suresh et al., 2016). Across these systems, the recurring structure is low-dimensional brain-derived intent mapped onto a higher-dimensional control policy by a learned or designed intermediate model.

4. Brain-OF as analytical, benchmarking, and inference infrastructure

A third usage centers on scientific analysis rather than direct actuation. In organoid imaging, BOrg is described as a first step toward a “brain-organoid-focused” framework for high-throughput automated analysis of mitosis in human brain organoids (Awais et al., 2024). The dataset contains 262 projected frames and 737 mitotic instances spanning prophase, metaphase, anaphase, and telophase. The annotation strategy uses sparse 3D point annotations projected into 2D, and the benchmark compares object detectors and a DeGPR-derived counting model. Mean projection is reported as the best 3D-to-2D strategy, and YOLOv8 achieves the best balance of precision, recall, and mAP among the tested detectors, while DeGPR++ remains competitive in class-wise MAE (Awais et al., 2024). In this sense, Brain-OF names a domain-specific analysis framework for developmental neuroscience.

For non-invasive BCI evaluation, AdaBrain-Bench defines a standardized benchmark for brain foundation models over 13 datasets and 7 application categories, including cross-subject, multi-subject, and few-shot transfer settings (Wu et al., 14 Jul 2025). It evaluates BIOT, EEGPT, LaBraM, and CBraMod, and introduces a transferability score: $$\text{TS} = \lambda \cdot \frac{P_{\text{prt}} - P_{\text{scr}}}{P_{\text{scr}}} + (1-\lambda)\cdot\frac{P_{\text{prt}} - P_{\text{scr}}}{P_{\text{oracle}} - P_{\text{scr}}}.$$ The benchmark reports LaBraM as the best cross-subject macro-average model at 64.61%, while CBraMod attains the highest average few-shot TS at 0.2761 (Wu et al., 14 Jul 2025). This usage shifts Brain-OF from a single model to an evaluative ecosystem for model selection and transfer analysis.

BrainOOD addresses a different analytical bottleneck: out-of-distribution generalization for graph neural networks on multi-site brain networks. It introduces a feature selector, a structure extractor, and auxiliary losses including an improved Graph Information Bottleneck objective, and it explicitly exploits fixed ROI correspondence across subjects. The paper reports improvements to OOD subjects by up to 8.5% and presents what it calls the first OOD brain network benchmark (Xu et al., 2 Feb 2025). Here Brain-OF is closely aligned with robust graph-based disease inference under site shift.

Brain-optimized inference adds yet another layer: inference-time optimization of reconstructed images against measured fMRI. Starting from a base decoder, it samples candidate reconstructions from a diffusion model, predicts brain activity with an encoding model, and iteratively retains images whose predicted activity best matches the measured pattern. When applied to MindEye, the resulting reconstructions are chosen by human raters 95.62% of the time in a two-way identification task against random reconstructions and are preferred 56.87% of the time over base MindEye reconstructions (Kneeland et al., 2023). Earlier visual areas are reported to converge more slowly and to prefer narrower image distributions than higher-level visual cortex, which the authors interpret as reflecting differences in representational invariance (Kneeland et al., 2023).

5. Brain-OF as clinical biomarker and personalized brain model

In acute ischemic stroke, Brain-OF appears as “brain outcome factors/features,” namely global structural quantities that mediate resilience to lesion burden (Alhadid et al., 2024). The focal comparison is between total brain volume at the time of injury and brain parenchymal fraction, defined as

$$\text{BPF} = \frac{\text{brain volume}}{\text{ICV}}.$$

Using 467 arterial ischemic stroke cases imaged within 48 hours, the study fits matched logistic regressions for 90-day outcome and compares them by Bayesian Information Criterion. Both higher BPF and larger brain volume are associated with favorable functional outcome, but the brain-volume model explains the data better, with $\text{BIC}=501$ versus $\text{BIC}=511$ for the BPF model; the reported $X \in \mathbb{R}^{L \times D}$0 sits at the boundary of “strong” to “very strong” evidence in favor of brain volume (Alhadid et al., 2024). In this clinical usage, Brain-OF is not an algorithmic architecture but a structural reserve construct.

Digital Twin Brain extends Brain-OF into subject-specific simulation and assimilation. DTB simulates whole human-brain-scale spiking neuronal networks up to $X \in \mathbb{R}^{L \times D}$1 neurons and $X \in \mathbb{R}^{L \times D}$2 synapses, with connectivity derived from sMRI, DTI, and PET and with mesoscopic data assimilation against BOLD signals (Lu et al., 2023). The reported scaling experiments show that the system is communication-intensive and memory-access intensive rather than computation-intensive. The full 86B/47.8T configuration reaches a time-to-solution of 65 s per 1 s biological time at 7 Hz average firing rate, 78.8 s at 15 Hz, and 118.8 s at 30 Hz (Lu et al., 2023). In a visual evaluation digital experiment, a mapping learned from biological brain activation to subjective ratings transfers to the digital twin with correlation $X \in \mathbb{R}^{L \times D}$3 and $X \in \mathbb{R}^{L \times D}$4 (Lu et al., 2023). This suggests a Brain-OF interpretation in which the “operating framework” is an assimilated plant-observer system for whole-brain dynamics.

The stroke and DTB usages are methodologically distant, but they share one substantive feature: both elevate global brain state above purely local descriptors. In stroke, lesion burden alone is insufficient without structural reserve. In DTB, isolated local microcircuits are insufficient without subject-specific global connectivity and mesoscopic assimilation.

6. Abstract operating principles and broader theoretical formulations

At the most abstract end, Brain Principles Programming formalizes universal mechanisms of the brain’s work with information and attempts to turn them into algorithms (Vityaev et al., 2022). Its operative mathematical substrate is the probabilistic formal concept. Given a system of probabilistic causal relations $X \in \mathbb{R}^{L \times D}$5, the prediction operator is

$X \in \mathbb{R}^{L \times D}$6

with closure

$X \in \mathbb{R}^{L \times D}$7

A probabilistic formal concept is a pair $X \in \mathbb{R}^{L \times D}$8 such that $X \in \mathbb{R}^{L \times D}$9 and $Z \in \mathbb{R}^{C \times D}$0. The paper then interprets BPP through five principles—generation of complexity, principle of relation, approximation to essence, locality–distributedness, and heaviness—and links them to functional systems, prototypes, causal models, and task-based AGI (Vityaev et al., 2022). In this usage, Brain-OF is an operating framework in the strictest sense: a formal theory of how concepts, contexts, and action plans are generated and stabilized.

A very different theoretical formulation is given by Quantum Brain Dynamics and the Quantum Interpretation of the Brain, which focus on consciousness and memory rather than benchmarked engineering tasks (Iwuh, 2023). There the brain is treated as a quantum field system of electromagnetic and water dipole fields, memory is associated with phase transitions and vacuum states, and consciousness is linked to a global field of condensed evanescent photons overlapping the whole brain tissue in the cranium (Iwuh, 2023). The proposal introduces corticons, spontaneous symmetry breaking, and Nambu–Goldstone modes as organizing concepts. This is not a usage of Brain-OF in the architectural or benchmarking sense, but it belongs to the same broader effort to specify brain-centered operating principles.

A plausible synthesis is that Brain-OF names a family of brain-centered operating frameworks at different explanatory scales. At the engineering end, it denotes multimodal foundation models, closed-loop BCI stacks, prosthetic and swarm controllers, OOD graph analyzers, organoid analysis pipelines, and inference-time reconstruction optimizers. At the clinical end, it can denote brain reserve variables that better explain post-injury outcomes. At the theoretical end, it points toward formal closure operators, functional systems, or even quantum-field descriptions of memory and consciousness. What unifies these otherwise disparate usages is not a single architecture, but a recurring methodological commitment: the brain is treated not as background context for an algorithm, but as the primary state space, control signal, or organizing ontology of the framework itself.