Electroencephalography (EEG) generation is essential for alleviating data scarcity and enabling large scale neural modeling in brain computer interface applications. However, existing flow based approaches assume that every channel and every time segment within a sample shares a single global time progression, overlooking the fact that not all EEG moments are equal. To address this overlooked heterogeneity, we propose an adaptive EEG generation framework built on conditional flow matching. The framework introduces Position-Adaptive Time Scheduling, which tracks per position reconstruction error to modulate a position specific time progress within the flow matching trajectory. It further incorporates Factorized Spatio-Temporal Attention and a frequency aligned multi resolution spectral consistency loss to model inter channel dependencies induced by volume conduction and compensate for the power law spectral bias of EEG, thereby improving the quality of generated signals. Extensive experiments on three EEG datasets with distinct acquisition protocols and task semantics show that our framework consistently outperforms the strongest baseline, reducing TS-FID by up to 62.2% and improving downstream classification accuracy gain by up to 6.77 percentage points. These results suggest that the proposed method represents a promising step toward scalable, high fidelity data augmentation for real world brain computer interface applications.

Abstract visual reasoning benchmarks such as ARC-AGI evaluate the ability to infer generalizable transformation rules from few graphical demonstrations—a capability where current deep learning models severely underperform. Mainstream large language models achieve only 15.8% (DeepSeek-R1) and 34.5% (o3-mini-high) test accuracy. The core reason lies in their static processing of task examples: unlike humans, who iteratively refine their understanding of examples while constructing solutions, these models lack mechanisms for dynamically aligning understanding and solving. We address this gap with the Understanding and Solving Reasoning Loop (USRL) framework. The architecture comprises two explicitly interacting modules: an Understanding Module (UM) that encodes and refines rule representations of examples, and a Solving Module (SM) that generates a draft solution informed by these evolving contexts. Through recurrent interaction, the model iteratively aligns its draft solution with its understanding about task examples continuously. Furthermore, we introduce an adaptive reasoning halting mechanism that autonomously terminates the reasoning loop based on the consistency between the generated draft solution and the rule representations. With only 7M parameters, our model achieves 47.2% accuracy on ARC-AGI-1, significantly outperforming both DeepSeek-R1 and o3-mini-high. This reveals that neurocognitive principles offer an effective pathway for abstract reasoning, with implications extending to compositional generalization and structured problem-solving.

Accurate and robust classification of medical pathology images is pivotal for computer-aided diagnosis. However, the deployment of deep learning models in high-throughput clinical screening faces a fundamental challenge: the trade-off between diagnostic accuracy and computational efficiency. Current lightweight architectures, while reducing parameter complexity through grouped convolutions, often lead to cross-channel information isolation and diminished representational capacity. In this paper, we propose TetraFuse, a novel framework that systematically integrates features from four complementary domains: space, channel, statistics, and frequency. TetraFuse introduces a novel Cross-Channel Dynamic Aggregation (CCDA) paradigm that reconstructs global channel topology with negligible computational overhead, resolving the inter-group isolation issue. To balance perceptual fidelity and efficiency, we design a stage-aware local enhancement mechanism: Local Variance-Guided Enhancer (LVGE) is employed to filter out shallow-stage background noise, while High-Frequency Boundary Injection (HFBI) reinforces deep-stage pathological contours, preventing spatial over-smoothing. Experimental results on the COVID-19, ISIC 2018, and Kvasir datasets confirm that TetraFuse outperforms state-of-the-art (SOTA) methods. Notably, TetraFuse-Tiny achieves a transformative 91.53% reduction in FLOPs compared to ResNet50; on the Kvasir dataset, it achieved an accuracy of 0.926 and an AUC of 0.994 with only 0.345G FLOPs. By combining high representational power with minimal computational demand, TetraFuse offers a scalable solution for large-scale medical image analysis, especially in resource-constrained clinical environments.

Random frame-level data missing is a critical challenge in multimodal sentiment analysis. Existing methods are largely limited to passive completion via single-pass feedforward connections and static cross-modal fusion, which struggle to generate high-quality completed features. However, the brain is not a passive recipient of external information but a dynamic system for active perceptual inference. Its core lies in the dynamic nested recurrents formed by intra-cortical recurrent completion mechanisms and corticothalamic circuits, which iteratively perform perceptual inference. Inspired by this, we propose the Dynamic Nested Recurrent Network (DNRNet). It is the first to introduce recurrent inference into the data completion task, achieving a paradigm shift from passive completion to active perceptual inference. Its local recurrent loop simulates intra-cortical recurrent pattern completion to perform perceptual inference and generate local correction features. The global recurrent loop simulates the modulatory function of the thalamus, calculating modality confidence to dynamically weight and integrate cross-modal information, generating global correction features. The local and global correction features are fused to obtain the completion signal, which is then combined with the input features of the current iteration to serve as the input for the next iteration. Experiments on the MOSI, MOSEI, and SIMS datasets demonstrate that DNRNet achieves an average accuracy improvement of 1.5%–2.0% over baseline models across all missing rates, validating the superiority of the brain-inspired approach in complex missing data scenarios.

Recently, Graph Neural Networks (GNNs) have been widely used in neural decoding due to strong topological feature mining and interpretability. GNNs are heavily based on manually defined brain topology; if there are false connections or noise, it will greatly affect the decoding performance. To address the aforementioned challenges, a series of GNN-based graph topology learning (GTL) methods have received widespread attention due to their ability to automatically optimize brain topology. However, existing GTL methods are usually implemented in a supervised manner and rely on a large amount of annotated data, making it difficult to directly transfer them to different decoding scenarios. Therefore, in this paper, a Brain Topology Inference framework based on Multi-View Contrastive Self-supervised Learning (BTI-MVCSL) is proposed for neural decoding. Specifically, BTI-MVCSL first designs a series of graph learners, which can infer brain topological connections as "learner", generate topology learning objectives as "instructor" from the original fMRI data, and maximize consistency between "instructor" and "learner" to extract the rich information in hidden connections. Furthermore, in order to achieve fully automated topology learning guidance, BTI-MVCSL develops a new self-learning mechanism that can use the "learner"-view brain topology to update the "instructor"-view brain topology during model optimization and further achieves comparative constraints through the "instructor" topology. The proposed BTI-MVCSL has been extensively evaluated in two publicly available fMRI datasets, demonstrating superior performance and revealing potential changes in brain topology under different decoding tasks.

Individuals with severe motor impairments face significant challenges operating conventional interfaces due to reliance on physical movement. While brain-computer interfaces (BCIs) offer alternative control pathways, traditional paradigms (SSVEP/P300) suffer from stimulus dependency, high latency, and cognitive fatigue, limiting real-world deployment. This study presents a novel asynchronous motor imagery-BCI system for hands-free robotic platform navigation, integrating cognitive ergonomics principles to enhance operator experience. We developed an intuitive four-command control paradigm using a cascaded classifier architecture, eliminating dependence on external triggers. Key innovations include: 1) sub-100 ms ultralow-latency pipeline via hybrid feature fusion and ensemble learning; 2) stimulus-independent operation leveraging endogenous sensorimotor cortex activation (mu/beta-band); and 3) cognitive load-optimized interaction with ROS-based neurofeedback and NASA-TLX-validated ergonomic design. Evaluated on BCI Competition IV 2a dataset (nine subjects) and self-collected high-resolution electroencephalography data (seven subjects), the system achieved 72.15% offline classification accuracy and 95.36% online command execution success rate with 97ms mean computational latency. NASA-TLX evaluation revealed a 32% reduction in cognitive workload compared to synchronous paradigms. This work establishes a framework for cognitively enhanced mobility assistance, advancing practical brain-controlled assistive technologies.

Brain-computer interface （BCI） establishes a mapping relationship between external stimuli and internal neural activity in the brain， providing an effective means to understand brain information processing mechanisms and achieve human-machine intelligent interaction. In recent years， foundational models have achieved breakthrough progress in various computer vision tasks， which has also propelled BCIs from task-specific models toward a general intelligence new paradigm. This paper is the first to review the latest research advances of foundational models in neural encoding and decoding for BCIs. It systematically outlines key studies and research trajectories in natural stimulus encoding-decoding， multimodal brain representation learning， and generalization studies. The analysis identifies current challenges in sample size， data heterogeneity， multimodal fusion， and model interpretability. Finally， it highlights future research directions for general-purpose BCIs. This work aims to provide a systematic reference and research insights for building general BCI models capable of handling complex cognitive scenarios.

Efficient storage and transmission of EEG data require high-ratio EEG compression. However, due to their limited capacity in removing statistical redundancy and insufficient focus on complex signal regions, existing convolution-based compression methods struggle to achieve a high compression ratio without sacrificing reconstruction quality. In addition, these methods are constrained by fixed network architectures, lacking the flexibility to adjust compression rates according to application-specific requirements. To address these challenges, we propose a hybrid compression framework that combines variational autoencoder with arithmetic coding. By jointly training, the EEG data are mapped to a more compressible probability distribution, which guides the entropy coding process and substantially enhances the overall compression ratio. In addition, we design a spatial-temporal attention module tailored to EEG characteristics. This module separately extracts and integrates temporal and spatial attention, allowing the network to focus more precisely on complex and informative regions of the signal. Furthermore, by jointly constraining both compression rate and reconstruction error, the proposed method supports adjustable compression ratios. Results on public datasets show that the proposed method maintains the best reconstruction performance even with a compression ratio far exceeding existing methods (100 vs.16).

Artificial intelligence (AI) is fundamentally reshaping societal futures, with its transformative applications driving profound, cross-sectoral change. From surgical robots that enhance medical precision and safety to autonomous vehicles that reconfigure transportation ecosystems, AI is accelerating social progress at an unprecedented pace and scale. Yet current AI technologies face persistent foundational challenges: overreliance on static, context-insensitive rules; limited capacity for dynamic, real-world simulation; and representation models that lack depth and integrative coherence. These limitations constrain adaptability in non-stationary environments, hinder the emergence of holistic intelligence, and obscure the mechanistic essence of true intelligence—constituting a critical barrier to artificial general intelligence (AGI). Dual-Brain Fusion, the deep, bidirectional synergy between the human brain and brain-inspired systems, represents the essential pathway to overcoming this barrier. As nature’s most sophisticated intelligent system, the human brain offers fundamental insights for brain-inspired research through its capacities for multimodal perception, adaptive cognition, and goal-directed decision-making. Although existing research has advanced perception optimization, cognitive enhancement, and decision-making improvement, it remains constrained by a single-layer simulation paradigm of brain functions, incomplete understanding of neural functional mechanisms, and the absence of engineered bidirectional interaction between the two systems. To address these challenges, this paper proposes leveraging the human brain’s perceptual, cognitive, and decision-making functional mechanisms as a unifying bridge, breaking down interdisciplinary barriers to construct a full-chain collaborative innovation paradigm centered on Dual-Brain Fusion, which enables mutual inspiration between the human brain and brain-inspired systems. Within this framework, brain-inspired systems are optimized using principles derived from human brain mechanisms, while empirical feedback from these systems informs refined analysis of brain functional mechanisms, establishing a closed loop of bidirectional empowerment. Looking ahead, the continued advancement of Dual-Brain Fusion will catalyze an intelligence leap grounded in brain functional mechanisms. This paradigm, anchored in brain mechanisms as a source of principled inspiration, drives AI breakthroughs at the level of intelligence itself, unlocks synergistic human–machine co-adaptation, and charts a developmental path toward human–machine symbiosis and co-evolution.

BackgroundMost studies treat visual impairment (VI) as a static, binary exposure. Less is known about whether distinct longitudinal VI patterns are differentially associated with later cognitive and affective trajectories, and whether social vulnerability modifies these associations.ObjectiveTo examine whether longitudinal VI patterns classified by directionality and chronicity are differentially associated with trajectories of cognitive function and depressive symptoms, and whether social vulnerability modifies these associations.MethodsWe analyzed data from 15,393 older adults in the Health and Retirement Study. VI was classified into four patterns (Never, Worsening, Improving, Persistent) based on self-reported visual status in 2006 and 2008. Cognitive function and depressive symptoms were assessed biennially from 2008 to 2020 and 2022, respectively. Linear mixed-effects models estimated trajectories adjusting for demographic, socioeconomic, and health covariates. Three-way interactions tested effect modification by loneliness, living alone, and partner status. Benjamini-Hochberg false discovery rate correction was applied to the three-way interaction analyses.ResultsAll VI groups had lower baseline cognitive scores and higher depressive symptoms than Never VI. The Improving VI group showed the steepest cognitive decline (β = −0.051/wave, 95% CI: −0.100 to −0.002, p = 0.040). Persistent VI had the highest baseline depressive symptoms but showed a relative decline over follow-up (β = −0.035/wave, p = 0.002). After correction for multiple testing, partner status attenuated cognitive decline in the Improving VI group, and partner status, living arrangement, and loneliness modified selected depressive symptom trajectories.ConclusionLongitudinal VI patterns carry distinct prognostic significance for cognitive and affective trajectories. Improving VI was associated with modestly faster subsequent cognitive decline (cumulative difference approximately 0.31 points over 12 years, or 1.1% of the cognitive scale), warranting continued monitoring even after apparent visual improvement. Social vulnerability modified selected VI-related trajectories, supporting the need for future research on integrated sensory and social intervention strategies.

IntroductionElectroencephalogram (EEG) emotion recognition is attracting increasing attention in the field of brain-computer interface due to its strong objectivity and non-forgery. However, cross-subject emotion recognition is complicated by individual variability, limited availability of EEG data, and interference in certain channels during EEG acquisition.MethodsWe propose a novel synergistic Dual Stream Pre-training and Multi-view Consistency Fine-tuning (DSP-MCF) framework. The DSP-MCF is based on a domain generalization architecture. The framework includes a dual stream pre-training stage, wherein the spatiotemporal encoder-decoder network extracts generalized spatiotemporal representations from masked channels and reconstructs EEG features from incomplete data. Then, a multi-view consistency loss function is proposed during the multi-view consistency fine-tuning. This loss function is essential for aligning the distribution of emotion predictions derived from various perspectives, specifically from actual and masked EEG data.ResultsExperimental results demonstrate that the proposed DSP-MCF framework outperforms state-of-the-art methods in cross-subject EEG emotion recognition tasks. The model achieved an accuracy of 89.76% on the SEED dataset and 77.02% on the SEED-IV dataset.DiscussionThe findings indicate that the DSP-MCF framework effectively addresses individual variability and maintains robust performance even under channel loss. By integrating spatiotemporal reconstruction with multi-view consistency, the model provides a reliable solution for handling incomplete or degraded EEG signals in practical BCI applications.

Extracting stable subject-specific features from EEG signals remains challenging due to their entanglement with transient brain states. We propose a universal neural framework that disentangles subject-specific features from state-dependent components in raw EEG signals. Our approach employs a disentanglement module with a cross-reconstruction objective to isolate subject-specific representations. We validate our framework on EEG-based biometric recognition using two public datasets with leave-one-state-out cross-validation. Results demonstrate significant improvements in out-of-distribution identification accuracy across four different backbone models, confirming our method’s universality and plug-and-play capability. This work advances reliable extraction of neural signatures for personalized neurotechnology applications.

Advances in tactile sensing have made contact-rich perception possible, accelerating progress in robotic manipulation, material understanding, and embodied interaction. However, because optical design, elastomer mechanics, and imaging geometry differ substantially across tactile sensors, models trained on known sensor types can suffer an abrupt performance collapse on unknown sensors. To address this problem, we propose the Brain-Inspired Few-Shot Tactile Adaptation (BIFTA) framework; it draws on the brain's rapid sensory adaptation mechanism to adapt a frozen encoder to an unknown tactile sensor from a small labeled support set. BIFTA preserves pretrained representations through dual-view statistical memory, constructs support-conditioned spectral graphs to repair sensor-dependent feature neighborhoods, and applies uncertainty-gated recurrent propagation to strengthen reliable cross-query evidence. Extensive benchmarks across three tactile datasets show that BIFTA substantially improves adaptation to unknown sensors: with only 10% labeled target data on SITR, it raises mean Sparsh accuracy from 6.86% for the frozen source classifier to 87.09%, exceeding the strongest implemented prior comparison by 47.22 percentage points, and these gains generalize across datasets, pretrained backbones, and tactile tasks. These results validate BIFTA for data-efficient adaptation to unknown tactile sensors and offer a promising route toward tactile models that transfer across heterogeneous hardware.

Cognitive state recognition has emerged as a key research topic at the intersection of neuroscience and artificial intelligence. While multi-channel EEG systems achieve high recognition accuracy, their complexity and cost substantially hinder large-scale deployment. In contrast, single-channel EEG has attracted increasing interest due to its portability and accessibility. Nevertheless, its feasibility for reliable cognitive state recognition remains uncertain, as a single electrode records only limited neural activity. To address this issue, we conducted a systematic set of experiments to comprehensively evaluate the potential of single-channel EEG systems in cognitive state assessment. The evaluation encompasses a broad spectrum of widely adopted tasks, including motor imagery, emotion recognition, and workload estimation. We benchmarked 17 representative EEG analysis approaches, ranging from traditional machine learning algorithms to advanced deep learning models, including methods specifically tailored for single-channel EEG. Furthermore, we introduce a novel quantitative metric to assess electrode placement effectiveness, thereby reducing evaluation bias caused by algorithmic differences. Experimental results reveal that single-channel EEG systems still face considerable challenges in cognitive state recognition. Specifically, their accuracy remains significantly lower than that of multi-channel systems, and task performance is highly sensitive to electrode location, with optimal placements exhibiting strong task dependency. These findings suggest that future research should emphasize the development of more effective single-channel EEG analysis techniques and adaptive electrode optimization strategies capable of dynamically adjusting placements according to task requirements.

Category-based attentional selection (CAS) enables the visual system to prioritize objects that share an abstract, semantic label. For example, "tools," "letters," or "animals." Yet how cognitive load and salience processing jointly sculpt this high-level form of attention remains unclear. Here we combined a Majority Function Task (MFT) with a visual Oddball manipulation in a fully crossed 2 (load: low 3:0 vs. high 2:1 ratio) & times; 2 (salience level: standard 80 % vs. Oddball 20 %) & times; 2 (salience relevance: task-relevant vs. task-irrelevant) design. Twenty-nine right-handed adults (24 women; 18-27 yrs) performed 768 trials while BOLD signals were recorded in a 3 T scanner; eye position was concurrently monitored to rule out overt shifts. Inverse-efficiency scores (IES = RT / accuracy) confirmed the expected main effect of load, but also revealed a three-way interaction: under high load, task-relevant Oddballs produced the largest cost (Cohen's d = 0.81), whereas task-irrelevant Oddballs caused a moderate, load-dependent slowdown. This pattern supports a resource-competition account in which maintaining a category template and suppressing conspicuous distractors draw on a common, finite pool. Whole-brain GLM revealed a functional division of effects. Cognitive load (high > low) boosted activity throughout the dorsal control network, including bilateral superior parietal lobule (SPL), dorsal lateral prefrontal cortex (DLPFC) and insula, whereas salience level (Oddball > standard) preferentially recruited ventral salience nodes, including right angular gyrus, bilateral anterior insula and caudate nucleus. By contrast, salience relevance (task-relevant vs. task-irrelevant) produced no reliable univariate clusters, mirroring the absence of a pure relevance main effect in local BOLD amplitude. To test whether relevance information was nonetheless encoded in spatial patterns, we performed multivariate pattern analysis (MVPA). A linear support-vector machine trained on voxels that were jointly responsive to load and salience distinguished the eight experimental conditions with 86.83 % accuracy (t = 73.57, p < .001). Weight-map inspection showed that the right superior occipital/parieto-occipital junction and right pre-central gyrus contributed most strongly but not exclusively, suggesting rPOJ and FEF serve as a convergence hub together with premotor nodes. Thus, although relevance does not manifest as a simple amplitude shift, it is robustly represented in distributed activation patterns and in the connectivity of a posterior occipito-parietal hub, highlighting a pattern-based, network-level code that reconciles the dorsal-ventral division of labor with successful category-based attentional selection. These converging results indicate that CAS operates through a layered priority architecture: dorsal control regions inject goal-related gain, ventral salience regions register statistical deviance, and rPOJ/FEF synergistically re-weights both streams to rebalance priority values when resources are scarce. Taken together, our findings extend priority-map theory into the semantic domain and demonstrate that cognitive load is a key moderator of how salience relevance shapes the competition between dorsal and ventral attention systems. By isolating where (dorsal vs. ventral) and how (pattern vs. amplitude) cognitive load and salience relevance interact, the study refines dual-route models of attention and identifies rPOJ and FEF as pivotal hubs for balancing task demands against environmental conspicuity, that is, a mechanism likely critical for real-world scenarios that call for rapid category-based decisions under pressure.

Neural decoding has always been the cutting-edge neuroscience issue, significant progress has been made in neural decoding with the support of deep learning technology. However, these breakthroughs are based on large-scale fully annotated functional magnetic resonance imaging (fMRI) data, which greatly hinders its further applicability. Recently, foundation models have garnered considerable attention in the realm of natural language processing, computer vision, and multimodal data processing due to their ability to circumvent the need for extensive annotated datasets while achieving notable accuracy gains. Nevertheless, the formulation of effective foundation model approaches tailored for connectivity-based complex spatio-temporal brain networks remains an unresolved challenge. To address these issues, in this paper, we proposed a general Temporal-Aware Graph Self-supervised Contrastive learning framework (TAGSC) for fMRI-based neural decoding. Concretely, it includes three innovative improvements to enhance fMRI-based graph foundation models: (i) a spatio-temporal augmentation strategy considers spatial brain region synergy and temporal information continuity to generate brain spatio-temporal contrastive views; (ii) a temporal-aware feature extractor learns brain spatio-temporal representations, which fully takes into account the continuous consistency of brain state transitions and fetches brain spatio-temporal interaction information from local to global; (iii) a fine-grained consistency loss assists in contrastive optimization from both temporal and spatial perspectives. Extensive evaluation on publicly available fMRI datasets demonstrated the superior performance of the proposed TAGSC and revealed biomarkers related to different states of the brain. To the best of our knowledge, it is among the earliest attempts to employ a spatio-temporal pre-trained model for neural decoding.

Trust decisions involve both social evaluation and uncertainty processing, yet standard trust game paradigms do not fully dissociate trust-specific social computation from more general risk- and value-related processes. In this exploratory whole-brain fMRI study, we examined how objective risk and social distance were associated with trust decisions within a 2 × 2 trust game. Twenty-three adults completed the task, and 20 were included in the fMRI analyses after exclusion for excessive head motion. Behaviorally, trust rates were significantly lower under high than low objective risk, whereas neither the main effect of social distance nor the interaction between objective risk and social distance was significant. Relative to baseline, task performance engaged prefrontal and parietal regions. Compared with distrust decisions, trust decisions were associated with greater activation in prefrontal and visual regions, along with stronger negative activation in the insula. Objective risk was associated with differential activation in temporal, supramarginal, and precentral regions. Under the present manipulation, we did not observe significant neural modulations by social distance. These findings suggest that, in this low-context paradigm, objective risk was a more robust source of behavioral and neural variation than social distance. Given the exploratory design, modest sample size, and the task’s limited ability to separate social trust from generic risk/value processing, the findings should be interpreted cautiously.

Continual learning (CL) for Electroencephalography (EEG) decoding is essential for practical Brain–Computer Interfaces (BCIs), yet it remains challenged by catastrophic forgetting and large inter-subject variability. Replay-based approaches have shown effectiveness, but their real-world deployment faces two critical issues: privacy risks from storing sensitive user data and distribution shift arising from subject variability. To address these challenges, we present the Brain-inspired Memory Consolidation and Attention (BMCA) framework, a new CL paradigm for EEG decoding that draws on neural mechanisms. For privacy preservation, BMCA leverages de-identified class-wise prototypes and synthetic data to replace sensitive raw EEG signals, mitigating privacy risks while retaining the core task-relevant information in the EEG. To mitigate distribution shift, it employs a lightweight architecture with a multi-scale temporal convolution module and a spatial attention mechanism, stabilized by historical momentum smoothing, designed to extract stable and task-relevant EEG features. Extensive experiments on benchmark motor imagery and emotion datasets show that BMCA consistently surpasses state-of-the-art methods under varying memory budgets, with improvements reaching up to 17.26%. By jointly tackling privacy preservation and efficient CL, this work establishes a novel framework that advances the development of practical and reliable BCIs.

Continual Learning (CL) requires balancing plasticity for new tasks with stability for prior knowledge. Parameter-Efficient Fine-Tuning (PEFT) offers a promising direction but often relies on a homogeneous expert module, limiting performance on the stability–plasticity trade-off. Neuroscience, however, shows that the brain addresses this challenge through distinct pathways for general retention and task-specific learning. Motivated by this, we propose Representation-Function Separation Continual Learning (RFSCL), which divides parameter updates into two asymmetric streams: a Representation Stream that captures transferable knowledge and a Function Stream that applies minimal task-specific adaptation. A Dynamic Belief Gating mechanism modulates their contributions based on task dynamics. Theoretically, this separation offers a new perspective on resolving the stability–plasticity dilemma. Empirically, it achieves state-of-the-art results, with a 2.24% absolute gain on the challenging 20-task ImageNet-R benchmark, confirming the practical benefits of our brain-inspired design.