Psychedelics, such as psilocybin, reorganise large-scale brain connectivity, yet how these changes are reflected across electrophysiological (electroencephalogram, EEG) and haemodynamic (functional magnetic resonance imaging, fMRI) networks remains unclear. We present Brain-MGF, a multimodal graph fusion network for joint EEG-fMRI connectivity analysis. For each modality, we construct graphs with partial-correlation edges and Pearson-profile node features, and learn subject-level embeddings via graph convolution. An adaptive softmax gate then fuses modalities with sample-specific weights to capture context-dependent contributions. Using the world's largest single-site psilocybin dataset, PsiConnect, Brain-MGF distinguishes psilocybin from no-psilocybin conditions in meditation and rest. Fusion improves over unimodal and non-adaptive variants, achieving 74.0

Abstract This special issue examines how natural and artificial intelligences (AIs) model the world, and what this modelling reveals about cognition and relationships between life and mind. Rather than adopting a single definition, the collection considers how world models function and emerge in biological and artificial systems, exploring a diverse range of world modelling including causal, self-referential, individual goal-directed, collective and narrative forms. A recurring theme is the extent to which current AI systems trained on vast quantities of data learn the context-sensitive, temporally embedded, value-laden dimensions of world modelling that characterize diverse biological intelligences, or whether their impressive capabilities arise primarily from statistical surface regularities. The contributions also raise broader issues concerning embodiment, complexity, learning architectures and the social and scientific contexts in which world models operate. With this collection, we hope to clarify the conceptual landscape, identify key points of similarity and divergence between natural and artificial minds, and outline questions that may guide future research on the forms of world modelling that support grounded understanding, robust agency and potentially human-like general intelligence. This article is part of the theme issue ‘World models in natural and artificial intelligence’.

BACKGROUND:Despite evidence of group-level differences in striatal morphometry among persons with Huntington's Disease (PwHD), current models of HD progression used for participant selection and assessment of treatment outcomes in clinical trials do not leverage shape information. METHODS:We first validated the capability of a discriminative deep neural network to derive descriptors of shape from all subcortical structures affected by HD, utilizing 2,932 brain scans in 615 PwHD across three longitudinal datasets (TRACK-HD, PREDICT-HD, and IMAGE-HD). We then trained a conditional generative model that used shape descriptors, alongside conventional volumetric, genetic, as well as composite cognitive, motor, and functional features at baseline to predict biomarkers of disease progression at subsequent time points. RESULTS:We observed that the anatomical shapes of subcortical structures, including putamen, lateral ventricle, pallidum, caudate, thalamus, and accumbens, exhibited strong associations with HD progression, as measured by a commonly used prognostic score. Furthermore, within-stage heterogeneity, along the continuum of disease progression, was better captured: when shape descriptors were aggregated using principal component analysis, they showed a high correlation with disease stage (Spearman's correlation: ρ = 0.72), compared to volumetric measurements in cubic millimetres (ρ = 0.45). Finally, incorporating subcortical shape into the generative model improved predictive performance, compared to the same model that relied solely on brain volumes. CONCLUSION:This study demonstrates that subcortical brain shape is associated with HD progression, enables capturing fine-grained within-stage variability, and improves the predictability of characteristic biomarkers. The findings could potentially optimize future clinical trials through more targeted participant recruitment and more objective post-intervention assessments of treatment efficacy.

Depression affects Huntington's disease (HD) gene-expansion carriers at up to four times the general population rate, yet its neurobiological underpinnings remain inadequately understood, potentially contributing to persistent symptoms and suboptimal therapeutic outcomes, despite commonly used interventions. Understanding of the pathophysiological mechanisms remains constrained by fragmented, single-level approaches that do not capture interactions between biological systems. Here, we present an illustrative multilevel, reciprocal pathophysiological framework that integrates evidence across brain network (macro), neurotransmitter (meso), and molecular and systemic (micro) scales that contribute to depression vulnerability in HD. We highlight available evidence suggesting HD-related neurobiological changes contribute to depression. Preferential vulnerability of GABAergic medium spiny neurons in HD may contribute to dysfunction across cortico-basal-ganglia-thalamic circuits, interacting with cellular, molecular, and systemic processes. At the macro level, altered neurocircuitry is represented by structural changes and functional dysconnectivity across brain networks. The meso level represents cellular and neurotransmitter alterations, including glutamatergic, serotonergic, and dopaminergic dysregulation. At the micro level, molecular and systemic alterations include neuroendocrine, immune, neurotrophic, and gut microbiota-brain axis processes. We discuss how this reciprocal framework provides a foundation for generating testable hypotheses of the mechanisms underlying depression in HD and highlight future research priorities, including improved psychiatric phenotyping, longitudinal multimodal study designs, cross-species behavioral harmonization, and integration of behavioral, circuit-level, molecular, and systemic measures.

Understanding how neuronal networks reorganize in response to external stimuli and give rise to behavior is a central challenge in neuroscience and artificial intelligence. However, existing methods often fail to capture the evolving structure of neural connectivity in ways that capture its relationship to behavior, especially in dynamic, uncertain, or high-dimensional settings with sufficient resolution or interpretability. We introduce the Temporal Attention-enhanced Variational Graph Recurrent Neural Network (TAVRNN), a novel framework that models time-varying neuronal connectivity by integrating probabilistic graph learning with temporal attention mechanisms. TAVRNN learns latent dynamics at the single-unit level while maintaining interpretable population-level representations, to identify key connectivity patterns linked to behavior. TAVRNN generalizes across diverse neural systems and modalities, demonstrating state-of-the-art classification and clustering performance. We validate TAVRNN on three diverse datasets: (1) electrophysiological data from a freely behaving rat, (2) primate somatosensory cortex recordings during a reaching task, and (3) biological neurons in the DishBrain platform interacting with a virtual game environment. Our method outperforms state-of-the-art dynamic embedding techniques, revealing previously unreported relationships between adaptive behavior and the evolving topological organization of neural networks. These findings demonstrate that TAVRNN offers a powerful and generalizable approach for modeling neural dynamics across experimental and synthetic biological systems. Its architecture is modality-agnostic and scalable, making it applicable across a wide range of neural recording platforms and behavioral paradigms.

The term _generative model_ is widely used in human neuroimaging; however, its meaning is often left implicit. Prompted by observations and discussions at the 2025 Organization for Human Brain Mapping (OHBM) Annual Meeting, we surveyed members of the neuroimaging community to examine how generative models are defined, used, and evaluated in practice. Responses revealed some agreement on functional criteria — such as a model’s ability to simulate data — alongside marked disagreement about whether specific, widely used methods should be considered generative models. Evaluative priorities also varied across respondents, though out-of-sample generalization and interpretability were consistently emphasized. In light of these findings, we propose a pragmatic working definition of generative modeling, and a short set of reporting commitments intended to make generative claims easier to interpret and evaluate.

Background : While general anaesthesia typically induces unconsciousness, some patients retain the capacity to respond behaviourally to noxious stimulation. Reduction in frontal alpha power has been proposed as a marker of arousal, but its clinical reliability is inconsistent. This exploratory study aimed to identify multichannel EEG spectral and connectivity markers associated with intraoperative behavioural responsiveness around noxious stimulation. Methods : Sixty-four-channel EEG was recorded intraoperatively from seven patients undergoing microlaryngoscopy under propofol anaesthesia accompanied by analgesia. Responsiveness was determined by reactions to verbal commands. Alpha-band spectral features of responders and non-responders were compared pre- and post-noxious stimulation using descriptive statistics. The relationship between spectral power of individual channels and the factors of noxious stimulation and responsiveness was investigated using a linear mixed model (LMM). Event-related synchronization/desynchronization (ERS/ERD) and weighted Symbolic Mutual Information (wSMI) were computed across frequency bands and response categories, with correction for multiple comparisons. Results : Alpha power and global coherence showed minimal changes after stimulation in both responders (n = 3) and non-responders (n = 4). In contrast, volitional (or cognitive) responses to noxious stimulation were reliably associated with increased high-beta and gamma power in sensory-motor and auditory cortices. These responses showed event-related synchronisation in left-central channels, whereas reflexive movements were marked by desynchronisation in same areas. Theta-band activity also differentiated response types: cognitive responses showed suppression, while reflexive responses showed enhancement, particularly over the same regions. Connectivity analysis further revealed that cognitive responses were associated with increased global whole-brain integration, especially in the theta-band linking motor, auditory, and premotor cortices. Reflexive responses, by contrast, were associated with reductions in global brain connectivity. Conclusions : Frontal alpha EEG markers did not reliably index intraoperative responsiveness. Instead, localized high-frequency power increases and enhanced theta-band connectivity more robustly reflected ‘connected consciousness’, a state in which patients remain capable of perceiving and processing sensory inputs despite anaesthesia. Despite the limitations behind small sample size and variability in anaesthetic and patient factors, the surgical setting of the study strengthens the clinical relevance of these multichannel EEG features for guiding intraoperative monitoring and analgesic management before noxious stimulation.

As intelligent systems are developed across diverse substrates - from machine learning models and neuromorphic hardware to in vitro neural cultures - understanding what gives a system agency has become increasingly important. Existing definitions, however, tend to rely on top-down descriptions that are difficult to quantify. We propose a bottom-up framework grounded in a system's information-processing order: the extent to which its transformation of input evolves over time. We identify three orders of information processing. Class I systems are reactive and memoryless, mapping inputs directly to outputs. Class II systems incorporate internal states that provide memory but follow fixed transformation rules. Class III systems are adaptive; their transformation rules themselves change as a function of prior activity. While not sufficient on their own, these dynamics represent necessary informational conditions for genuine agency. This hierarchy offers a measurable, substrate-independent way to identify the informational precursors of agency. We illustrate the framework with neurophysiological and computational examples, including thermostats and receptor-like memristors, and discuss its implications for the ethical and functional evaluation of systems that may exhibit agency.

Most functional magnetic resonance imaging studies rely on estimates of hierarchically organized functional brain networks whose segregation and integration reflect the cognitive and behavioral changes in humans. However, most existing methods for estimating the community structure of networks from both individual and group-level analysis methods do not account for the variability between subjects. In this paper, we develop a new multilayer community detection method based on Bayesian latent block model (LBM). The method can robustly detect the community structure of weighted functional networks with an unknown number of communities at both individual and group levels and retain the variability of the individual networks. For validation, we propose a new community structure-based multivariate Gaussian generative model to simulate synthetic signal. Our simulation study shows that the community memberships estimated by hierarchical Bayesian inference are consistent with the predefined node labels in the generative model. The method is also tested via split-half reproducibility using working memory task fMRI data of 100 unrelated healthy subjects from the Human Connectome Project. Analyses using both synthetic and real data show that our proposed method is more accurate and reliable compared with the commonly used (multilayer) modularity models. The code of this work is available at:https://github.com/LingbinBian/CommuDetectLBM.

While general anaesthesia typically induces unconsciousness, some patients retain the capacity to respond behaviourally to noxious stimulation. Reduced frontal alpha (8–12 Hz) power has been proposed as a marker of arousal, but its clinical reliability remains inconsistent. This exploratory study aimed to identify multichannel EEG spectral and connectivity markers associated with intraoperative behavioural responsiveness around noxious stimulation. Sixty-four-channel EEG was recorded intraoperatively from seven patients undergoing microlaryngoscopy under propofol anaesthesia with analgesia. Responsiveness was assessed via reactions to verbal commands. Alpha-band spectral features of responders and non-responders were compared pre- and post-noxious stimulation using descriptive statistics. Linear mixed models evaluated the relationship between channel-wise spectral power, and the factors of noxious stimulation and responsiveness. Event-related synchronization/desynchronization and connectivity via weighted Symbolic Mutual Information (wSMI) were analysed across frequency bands and response categories, with correction for multiple comparisons. Alpha power and global coherence showed minimal changes following noxious stimulation across patients. In contrast, volitional responses to noxious stimulation were associated with increased high-frequency power in sensory-motor and auditory cortices. These responses showed event-related synchronisation in left-central channels, whereas incoherent movements were marked by desynchronisation in the same areas. Theta-band activity further differentiated response types: cognitive responses showed suppression, while incoherent movements showed enhancement, particularly over the same regions. Cognitive responses were associated with increased global whole-brain integration, especially in the theta-band linking motor, auditory, and premotor cortices. Incoherent movements, by contrast, were associated with reductions in global brain connectivity. None of these patients reported postoperative awareness with recall. These preliminary findings suggest that localised high-frequency power increases and enhanced theta-band connectivity may reflect “connected consciousness,” in which patients retain the capacity to process sensory input despite anaesthesia. As this state may precede awareness with recall, particularly during noxious stimulation, its detection remains a key clinical challenge, underscoring the potential of these multichannel EEG features as markers. Despite limitations in sample size and variability in anaesthetic and patient factors, the surgical setting enhances the clinical relevance of these findings. However, further validation is required before clinical application for intraoperative monitoring and optimisation of analgesia prior to noxious stimulation.

This paper introduces an extension of generalised filtering for online applications. Generalised filtering refers to data assimilation schemes that jointly infer latent states, learn unknown model parameters, and estimate uncertainty in an integrated framework – e.g., estimate state and observation noise – at the same time (i.e., triple estimation). This framework appears across disciplines under different names, including variational Kalman-Bucy filtering in engineering, generalised predictive coding in neuroscience, and Dynamic Expectation Maximisation (DEM) in time-series analysis. Here, we specialise DEM for “online” data assimilation, through a separation of temporal scales. We describe the variational principles and procedures that allow one to assimilate data in a way that allows for a slow updating of parameters and precisions, which contextualise fast Bayesian belief updating about the dynamic hidden states. Using numerical studies, we demonstrate the validity of online DEM (ODEM) using a non-linear – and potentially chaotic – generative model, to show that the ODEM scheme can track the latent states of the generative process, even when its functional form differs fundamentally from the dynamics of the generative model. Framed from a neuro-mimetic predictive coding perspective, ODEM offers a biologically inspired solution to online inference, learning, and uncertainty estimation in dynamic environments.

PsiConnect is a large-scale neuroimaging study designed to investigate context-dependent neural and subjective effects of psilocybin using multimodal neuroimaging. It combines functional, structural, and diffusion-weighted MRI with EEG to examine brain activity in 62 participants before and after a 19 mg dose of psilocybin. The design includes resting-state scans and three naturalistic conditions: guided meditation, music listening, and movie watching. Half of the cohort underwent an 8-week meditation training program, enabling exploration of interactions among meditation, psilocybin, and brain function. fMRI data was obtained through multi-echo fMRI, enhancing signal-to-noise ratio and reducing susceptibility artifacts to improve reliability. A comprehensive battery of behavioural and self-report measures captured acute and longitudinal cognitive and subjective effects, with follow-ups to one year post-administration. The large sample, multimodal imaging, contextual diversity, and behavioural follow-ups enable study of psilocybin-induced brain and behaviour changes with unprecedented comprehensiveness and reliability. Data is curated according to open science principles to ensure accessibility and compatibility with established neuroimaging pipelines, making PsiConnect a valuable, reusable resource for cognitive and computational neuroscience.

In this paper, a novel test-time scaling law for physical artificial intelligence (AI) agents is introduced. This scaling law enables physical AI agents to reason with their world models to generalize in unforeseen scenarios at test time. The derived scaling law is grounded in the first principle of active inference, which equips agents with the general objective to survive in the real world, under which their specific task objectives are subsumed. Active inference achieves this by providing the reasoning to resolve prediction errors that arise when the agent encounters unforeseen situations outside its training distribution, enabling generalization in non-stationary environments. The proposed scaling law captures this by dynamically updating the agent's policy with this reasoning at test time. This policy update is modeled as a soft Bayesian inference process in which beliefs about the policy are updated using the reasoning that reduces expected prediction errors under allowable policies as a likelihood. The resulting posterior policy admits a biological interpretation, recovering the scaling mechanism that engages the brain's basal ganglia and prefrontal cortex at test time. To solve this analytically intractable problem, a variational inference solution minimizing free energy bounds is developed. This solution extends to enable learning beyond training by reinforcing new instances, resolved at test time, in both the policy and world model. Unlike existing scaling laws constrained by model size and training data, the derived solution scales with the continuous real-world experience of a physical AI agent. Simulation results on an autonomous driving task demonstrate that the proposed solution outperforms model-free Q-learning and model-based Bayesian reinforcement learning, achieving robust generalization to unforeseen scenarios while improving inference efficiency by over 36

Psychedelics can profoundly alter consciousness by reorganizing brain connectivity1,2, producing acute experiences that shape lasting psychological change3,4. Psychedelic dynamics are commonly described as desynchronized or entropically disordered5,6, yet the brain organization underlying self-dissolving and boundary-dissolving experiences that participants often report7, and how context shapes that organization8, remain unresolved. To address this, we acquired the largest single-site psychedelic neuroimaging dataset to date. Sixty-two adults underwent functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) during rest and naturalistic stimuli (meditation, music and movie), before and on the day of psilocybin administration (fMRI ~ 80 min post-dose; EEG ~ 150 min post-dose). Half ranked the experience among the most meaningful of their lives7. Here, using machine learning to represent the brain dynamics of each individual as low-dimensional trajectories, we show that psilocybin reorganizes brain activity into structured, context-sensitive patterns that co-vary with the quality of subjective experience, revealing a latent order missed by time-averaged measures. Networks that ordinarily segregate internal and external processing integrated, producing cohesive context-aligned trajectories in participants reporting the felt experience of being continuous with, rather than separate from, the environment, a state we refer to as embeddedness. The strength of this context alignment scaled with both the depth of self-dissolving and boundary-dissolving experience and the next-day mindset change. Our findings recast apparent disorder as latent organization aligned with context, linking neurobiology to subjective experience and behavioural change.

BACKGROUND:Object-location memory impairment in Huntington's disease (HD) occurs from premanifest period and declines as HD progresses, however, pathogenesis of object-location memory is unknown. The striatum and hippocampus are affected in HD, functionally interacting allowing intact object-location memory. OBJECTIVES:The present study investigated if object location memory impairment in premanifest and manifest HD was associated with aberrations in effective connectivity between striatum and hippocampal formation. METHODS:Using dynamic causal modelling, we examined effective connectivity between the striatum and hippocampus and association with object-location memory in 35 HD participants (23 premanifest, 12 early manifest) and 32 controls. RESULTS:HD participants' object-location memory was worse than controls, with performance associated with aberrant effective connectivity from the striatum to hippocampal formation, lower connectivity was associated with poorer object-location memory. Connectivity from hippocampal formation to the striatum was lower in manifest HD. In premanifest HD, connectivity from parahippocampal gyrus to the striatum was stronger and associated with better object-location memory. CONCLUSIONS:Findings raise questions regarding compensatory neural processes in HD and other neurodegenerative diseases, providing pathophysiological evidence that cognitive impairment may be related to connectopathy. © 2025 The Author(s). Movement Disorders published by Wiley Periodicals LLC on behalf of International Parkinson and Movement Disorder Society.

Decades of cross-species research highlight the claustrums extensive bidirectional connectivity with cortical and subcortical regions, implicating it in higher-order cognitive processes requiring synchronized brain states. Psychedelics may disrupt this synchrony by modulating claustro-cortical signaling, reflected by the dissolution of cortical network signatures. Using spectral dynamic causal modeling on resting-state fMRI data from the Human Connectome Project and PsiConnect datasets at 7T and 3T, we provide the first in vivo characterization of claustral effective connectivity with triple networks and subcortical regions in humans, both at rest and under the influence of psilocybin. Claustra displayed widespread bidirectional effective connectivity and a strong inhibitory influence on all target regions. Psilocybin enhanced claustral inhibition of cortical networks while disinhibiting subcortical areas, partially associated with psychedelic subjective effect scores. These findings are consistent with cellular and functional cross-species data, supporting the proposed mechanism of claustro-cortical inhibition in regulating network synchrony, while extending this influence to the subcortex, and revealing hierarchical and hemispheric asymmetries in claustral signaling modulation under psilocybin. ### Competing Interest Statement The authors have declared no competing interest.

This paper offers a road map for the development of scalable aligned artificial intelligence (AI) from first principle descriptions of natural intelligence. In brief, a possible path toward scalable aligned AI rests on enabling artificial agents to learn a good model of the world that includes a good model of our preferences. For this, the main objective is creating agents that learn to represent the world and other agents' world models, a problem that falls under structure learning (also known as causal representation learning or model discovery). We expose the structure learning and alignment problems with this goal in mind, as well as principles to guide us forward, synthesizing various ideas across mathematics, statistics, and cognitive science. We discuss the essential role of core knowledge, information geometry, and model reduction in structure learning and suggest core structural modules to learn a wide range of naturalistic worlds. We then outline a way toward aligned agents through structure learning and theory of mind. As an illustrative example, we mathematically sketch Asimov's laws of robotics, which prescribe agents to act cautiously to minimize the ill-being of other agents. We supplement this example by proposing refined approaches to alignment. These observations may guide the development of artificial intelligence in helping to scale existing, or design new, aligned structure learning systems.