In this paper, the phenomenon generally classified as confirmation bias is formulated on the space of square-root probabilities (or equivalently, using the structures of quantum probability). In this framework, observations are modelled by matrices, rather than random variables on a probability space. In the problem of binary hypothesis testing, an optimal evidence choice minimises the expected error probability. We show that the resulting optimal choice of evidence leads to a confirmation bias, thus revealing a surprising aspect of rationality that encompasses confirmation bias. Specifically, in sequential evidence sampling, the implicit optimality leads to two remarkable evolutionary advantages, namely, (a) the decision maker requires only the smallest memory capacity, and (b) the error probability can be reduced exponentially in sample size. A complementary approach based on the framework of active inference – where the decision maker seeks evidence that provides maximum information – is then considered. The resulting optimal evidence is shown to agree with the one obtained by minimising error probability. Our framework provides an easy-to-implement protocol for an active quantum inference, whereby the optimal evidence choice for making an inference is sought over the space of matrices.

Top-down attention enables selective prioritisation of goal-relevant information, yet how attentional control varies across individuals remains unclear, particularly in relation to sensory processing. Here, we combined magnetoencephalography (MEG) and pupillometry—in older adults performing a spatial cueing task—to examine both shared and subject-specific aspects of top-down attention. Informative spatial cues elicited sustained activity in fronto-parietal and sensory regions, alpha lateralisation, and increased pupil dilation, replicating established neural and physiological markers. Crucially, these group-level patterns were accompanied by systematic intersubject variability: individuals with greater hearing loss showed increased preparatory engagement of precentral and postcentral gyri, and variation in pupil dilation predicted activity in frontal regions. These results reveal that top-down attention arises from heterogeneous network configurations shaped by sensory processing and strategic resource allocation. Overall, our findings underscore the need to move beyond group averages to capture the neural architecture underlying individual differences in top-down attention.

This paper is a first attempt to marry constructive nonlinear control theory techniques with active inference. Specifically, we are interested in the relationship between differential flatness and the design of generative models for use in control settings. We place specific emphasis on the pathwise properties of differentially flat systems that inherit from their definition in terms of successive temporal derivatives and relate this to the use of generalised coordinates of motion in formulating continuous-time generative models in active inference. To illustrate the basic concepts, we appeal to the example of oculomotor control.

Consciousness science faces the challenge of bridging first-person experience with third-person empirical measurements. Neurophenomenology aims to build such `generative passages' connecting the content of experience with behavioural and neuroscientific data. However, the mathematical machinery for such bridges remains underdeveloped. Here we develop a Rosetta Stone hypothesis from predictive processing, where beliefs serve as a central hub connecting phenomenology, behaviour, and neural dynamics. This hinges on a central technical assumption that phenomenology is a function of beliefs. We pursue a conditional approach: if this assumption holds, then certain predictions mathematically follow. We derive predictions for subjective similarity judgements, cognitive metabolic cost, subjective cognitive effort, and time perception. We review the connection between beliefs and neural dynamics to complete the generative passage for neurophenomenology, omitting the connection between beliefs and behaviour as this is already well-documented elsewhere. Testing our predictions will inform the validity of the central assumption connecting beliefs and phenomenology, and advance the neurophenomenology research programme.

As neuroscientific theories of consciousness continue to proliferate, the need to assess their similarities and differences - as well as their predictive and explanatory power - becomes ever more pressing. Recently, a number of structured adversarial collaborations have been devised to test the competing predictions of several candidate theories of consciousness. In this review, we compare and contrast three theories being investigated in one such adversarial collaboration: Integrated Information Theory, Neurorepresentationalism, and Active Inference. We begin by presenting the core claims of each theory, before comparing them in terms of the phenomena they seek to explain, the sorts of explanations they avail, and the methodological strategies they endorse. We then consider some of the inherent challenges of theory-testing, and how adversarial collaboration addresses some of these difficulties. The stage is then set for the empirical work to come: first, we outline the key hypotheses to be tested across a series of multi-site experiments; second, we discuss the kinds of observations that would support or challenge each theory; third, we consider how these theories might assimilate or accommodate such observations. Finally, we show how data harvested across disparate experiments (and their replicates) may be formally integrated to provide a quantitative measure of the evidential support accrued under each theory. Besides orienting the reader to the theoretical foundations of our collaboration, this review aims to provide valuable meta-scientific insights into the mechanics of adversarial collaboration and theory-testing in general - including the way theories may be evaluated in terms of the scientific progress they deliver.

BACKGROUND:Reduced mismatch negativity (MMN) and P300 event-related potential (ERP) components are widely replicated in schizophrenia and are also observed in individuals at clinical high risk for psychosis (CHR-P) who subsequently convert to psychosis. It is unknown whether they reflect changes in excitatory and/or inhibitory synaptic function-both implicated in schizophrenia and considered potential drug targets. METHODS:We analyzed baseline MMN and P300 ERPs from the North American Prodrome Longitudinal Study (NAPLS 2) and asked whether altered synaptic excitation, inhibition, or both could explain amplitude reductions in CHR-P (n = 583). CHR-P participants who converted to psychosis (CHR-converters; n = 77) or remitted by the 24-month follow-up (CHR-remitters; n = 94) were compared on MMN evoked by pitch + duration double-deviant tones and P300 elicited by target tones from passive and active auditory oddball paradigms, respectively. Biophysical modeling was used to infer (excitatory) pyramidal cell and (inhibitory) interneuron function from both MMN and P300 ERPs. RESULTS:MMN and P300 amplitude reductions in future CHR-converters relative to CHR-remitters were best explained by reduced pyramidal cell excitability (posterior probability P > 0.95 of a group-by-condition interaction effect). In simulations, reduced pyramidal cell excitability suppressed deviant and target ERPs. Within CHR-converters, more severe positive symptoms were associated with disinhibition of pyramidal cells (P > 0.99). CONCLUSIONS:Results mirror previous findings in schizophrenia and suggest that reduced pyramidal cell excitability is present at baseline in future CHR-converters, consistent with the hypothesis that hypofunction of pyramidal cells is a primary pathology in schizophrenia rather than a consequence of chronic illness. Positive symptoms among CHR-converters may reflect compensatory downregulation of inhibition.

Major Depressive Disorder (MDD) is highly heterogeneous, limiting treatment efficacy. Despite efforts to delineate patient heterogeneity through subtyping, current approaches remain limited by noise, lack of clinical applicability, and insufficient external validation. Crucially, they focus on subtyping while neglecting staging information (e.g., illness duration). We developed BrainCVAE, a contrastive variational autoencoder, to disentangle MDD-specific neural features. Applying BrainCVAE to fALFF-derived resting-state fMRI from 1590 patients and 1308 controls identified two subtypes: Subtype 1 with hyperactivity in visual, attention, and default mode networks, and Subtype 2 with hypoactivity. Subtypes were validated in 1276 patients across independent centers. Subtype 1 showed superior responsiveness to pharmacological (SSRIs, SNRIs) and non-pharmacological (rTMS) interventions. Cross-sectional analyses revealed subtype-specific differences in DMN profiles across illness duration: Subtype 1 shifted from hyper- to hypoactivity, whereas Subtype 2 remained consistently hypoactive. In an independent dataset, illness duration correlated negatively with symptom reduction (r = -0.5565, 95% CI = (-0.8123, -0.1210), p = 0.0165). Datasets were ethically approved and registered on ClinicalTrials.gov: XJ_QG (NCT05577481, May 24, 2023), SAINT (NCT04653337, Oct 21, 2020), XJ_KG (NCT05544071, May 24, 2023). Integrating subtyping with illness staging bridges neurobiological heterogeneity and disease progression, providing a clinically actionable framework for precision treatment in MDD.

Attractor dynamics are a hallmark of many complex systems, including the brain. Understanding how such self-organizing dynamics emerge from first principles is crucial for advancing our understanding of neuronal computations and the design of artificial intelligence systems. Here we formalize how attractor networks emerge from the free energy principle applied to a universal partitioning of random dynamical systems. Our approach obviates the need for explicitly imposed learning and inference rules and identifies emergent, but efficient and biologically plausible inference and learning dynamics for such self-organizing systems. These result in a collective, multi-level Bayesian active inference process. Attractors on the free energy landscape encode prior beliefs; inference integrates sensory data into posterior beliefs; and learning fine-tunes couplings to minimize long-term surprise. Analytically and via simulations, we establish that the proposed networks favor approximately orthogonalized attractor representations, a consequence of simultaneously optimizing predictive accuracy and model complexity. These attractors efficiently span the input subspace, enhancing generalization and the mutual information between hidden causes and observable effects. Furthermore, while random data presentation leads to symmetric and sparse couplings, sequential data fosters asymmetric couplings and non-equilibrium steady-state dynamics, offering a natural generalization of conventional Boltzmann Machines. Our findings offer a unifying theory of self-organizing attractor networks, providing novel insights for AI and neuroscience.

Minimal Phenomenal Experiences (MPEs) are states of consciousness in which wakefulness is preserved but phenomenal content is low or absent. The Entropic Brain Hypothesis (EBH) is a model of conscious processes that regards the entropy of spontaneous brain activity as a marker of 'phenomenal richness', exemplified by high-content psychedelic experiences (HCPEs). Yet recent human neuroimaging studies of MPEs induced by meditation – and possibly 5-MeO-DMT – suggest that these states, defined by their phenomenological simplicity, also show signs of increased neurophysiological entropy. This presents a conundrum for the EBH: brain entropy is elevated with increased and decreased richness of the phenomenal experience. Here, we put forward the Complex Brain Hypothesis (CBH), which proposes that the richness of experience differentiating MPEs from HCPEs is better indexed by complexity than by entropy. We argue that brain complexity is modulated by the grain of inference through which the brain resolves uncertainty: some HCPEs exemplify a fine-grained regime, in which loosened constraints amplify fluctuations into proliferating content, whereas some MPEs exemplify a coarse-grained regime, in which a simpler model dissolves variety into an experience of 'contentless' awareness. Both regimes can be associated with elevated brain entropy, but they diverge in phenomenology and perturbational signatures. By resolving the entropy-content conundrum, the CBH refines the EBH and highlights MPEs as an important test case for computational theories of consciousness.

This paper offers an active inference narrative that considers deeply-held social attitudes and beliefs in relation to the cognitive concept of zones of bounded surprisal (ZBS) proposed by Manrique and Walker (2023). It is argued that narrow ZBS band-widths characterise the discriminatory minds of people who see themselves as an in-group. They tend not to be indignant or outspoken, but, instead, acquiescent and silent, when they witness poor behaviour of those whom otherwise they regard as members of their in-group. Plausibly, when such behaviour harms other people, as, for instance, in racist or gender-based violence, its perpetrators have a very narrow ZBS band-width: one that likely is a characteristic of dehumanising minds. Because such perpetrators see themselves as members of an in-group with entitlement to control aspects of society, they humiliate or abuse out-groups to which they assign others, and have no compunction about violating their dignity or human rights. We briefly consider policies that could lessen the unwelcome social repercussions of the behaviour of people with discriminatory and dehumanising minds.

We investigate the subjective experience of space around the visual blind spot area, the cortical representation of which is missing feedforward connectivity from one eye. We performed these experiments as part of an adversarial collaboration to test contrasting theories of consciousness; Integrated Information Theory (IIT), Predictive Processing Active Inference (AI), and Predictive Processing Neurorepresentationalism (NREP) accounts. According to the Integrated Information Theory of consciousness, non-activatable retinotopic cortical regions, such as the blind spot region for the ipsilateral eye, create a different cause-effect structure and therefore should contribute differently to the perceived quality of space of activatable retinotopic regions. The two Predictive Processing accounts, in contrast, posit that internal models will accommodate structural deviations around the blindspot based on the available sensory evidence (particulars of this accommodation differ between the two accounts). We present a series of paradigms in which participants evaluate distances and areas that either include the blind spot or not (without stimulating it directly), as well as illusory motion that is either adjacent to the blind spot or not. We model psychometric functions relating perceived and objective space. These models vary in terms of bias and precision according to the experimental conditions (blind spot involved vs. not involved, ipsilateral vs. contralateral eye), making it possible to quantify the potential disruption of subjective spatial extendedness induced by the blind spot. We present simulated results for each experiment corresponding to the predictions of each account, and conclude by discussing challenges and plans for dissemination.

How the brain plans and maintains sequences of future actions remains a central question in systems neuroscience. Studies in the frontal cortex revealed that multiple elements of a sequence are represented simultaneously in separable neural subspaces, challenging classical sequential planning models. Here, we show that these representations emerge naturally under inferential planning, in which sequential actions are inferred from sensory evidence and goals. Using a hierarchical generative model, we reproduce key neural phenomena observed in the primate frontal cortex, including the simultaneous activation of multiple plan elements, the emergence of (almost) orthogonal “memory” subspaces, and their reuse across forward and backward tasks. Our approach provides a mechanistic account of how probabilistic inference over control states produces distributed neural representations of plans. This framework unifies planning, working memory, and motor preparation, and generates predictions about the dynamics of active inference, the role of subspaces, and the impact of uncertainty on sequence processing.

Biological and artificial agents operating in complex environments have to leverage environmental structures to accomplish vital tasks. Recent research across a variety of domains—from the study of animal and human behaviour in different developmental periods and for different tasks, to computational studies of learning—has unveiled many ways in which structures are processed. This gave rise to a burgeoning field of study—structure learning. However, the diversity of phenomena studied, and the different aims and focuses of the researchers, have led to ambiguity and limited consensus on the nature of structure learning and its underlying mechanisms. In this paper we provide a synopsis of illustrative examples of structure learning, introduce the Active Inference Framework (AIF) with a focus on Structure Learning, and discuss points of contact between the two. The Active Inference Framework provides a mechanistic theory which distinguishes three levels of learning: Active Inference, Parametric Learning, and Bayesian Model Selection (a.k.a., Structure Learning), a method for the comparison and selection of models based on model evidence. We argue that when formalised under the Active Inference Framework, Structure Learning provides not only an underlying computational mechanism with aims of ecological validity, but also provides features relevant to computational accounts of structure learning more generally. The unifying aspect of the AIF in terms of having a single objective function for optimising behaviour should not be confounded with the exclusivity of this framework. The integration with other computational accounts is advised.

The vertebrate brain must balance internally generated predictions with constraints of environmental affordances. This balance constitutes a fundamental principle of neural organization that underwrites cortical computation. Using the prosomeric model of the neuraxis, we show how dorsalizing and ventralizing morphogenetic gradients specify excitatory and inhibitory lineages during development, establishing the functional architecture of active affordance. These developmental asymmetries are elaborated through telencephalic expansion, pallial-subpallial integration, and laminar differentiation of the neocortex, as described by the structural model. We demonstrate that motor control emerges within a sensory-predictive architecture due to the alar origin of the telencephalon and that increasing excitatory-inhibitory complementarity within the mammalian neocortex enables selective, context-sensitive action. Subpallial and diencephalic systems provide inhibitory governance over cortical action tendencies, supporting policy evaluation and selection in the framework of active inference. At the base of this hierarchy, the hypothalamus integrates homeostatic and allostatic signals to bias the landscape of affordances, shaping the likelihood of action policies. Together, these findings establish active affordance as a developmental and evolutionary framework linking prosomeric neurodevelopment, cortical architecture, subcortical control, and adaptive behavior. Active inference is thereby situated as the mature cortical expression of a conserved biological solution to acting in an uncertain world.

Episodic memories - declarative memories of past events, characterized by rich spatiotemporal context - play a central role in guiding perception and behaviour. Here, we advance a model that integrates episodic memories within the active inference framework. We describe how episodic memories are incorporated into the generative models used in active inference to support the re-construction, replay and communication of past events. In doing so, we foreground two foundational themes. The first is the message passing in deep temporal models that allow one to actively construct memories of episodes. The second is the communicative aspect of declarative memories, and the way in which one might recount something from one's autobiography. In effect, this means that the message passing that supports episodic memory propagates information about what we have done - or what we would do - given past circumstances to draw inferences about how to communicate those beliefs. Together, these themes emphasise that we are not passive recorders of the things that happen to us. We are active participants in the events we recall and in the telling of stories about them.

Lesion network mapping (LNM) and related techniques have been used in over 200 studies, primarily to test whether anatomically distributed lesions that cause the same symptom fall within a common brain network. A recent article1 challenges the specificity and validity of this technique, suggesting that lesion network maps primarily reflect intrinsic properties of the normative connectome rather than lesion-symptom relationships. However, the data and procedures in van den Heuvel et al. do not reflect those used in most LNM studies. Further, the main conclusions were based on similarity between maps, but similarity does not imply the absence of meaningful differences. In contrast, LNM provides evidence for meaningful differences using specificity testing. Exemplary analyses of 1090 lesion locations from 34 prior LNM studies do not support van den Heuvel's concerns and confirm the lesion-deficit specificity of LNM. While we encourage further methodological investigation, the analyses of van den Heuvel et al. do not invalidate prior LNM findings or future applications.

Building autonomous --- i.e., choosing goals based on one's needs -- and adaptive -- i.e., surviving in ever-changing environments -- agents has been a holy grail of artificial intelligence (AI). A living organism is a prime example of such an agent, offering important lessons about adaptive autonomy. Here, we focus on interoception, a process of monitoring one's internal environment to keep it within certain bounds, which underwrites the survival of an organism. To develop AI with interoception, we need to factorize the state variables representing internal environments from external environments and adopt life-inspired mathematical properties of internal environment states. This paper offers a new perspective on how interoception can help build autonomous and adaptive agents by integrating the legacy of cybernetics with recent advances in theories of life, reinforcement learning, and neuroscience.

To encode information efficiently, our perceptual system should detect when situations are unpredictable (that is, informative) and modulate brain dynamics to prepare for encoding. Under uncertainty, there is an increased need to generate predictions about upcoming information, a process that has been proposed to require coordinated activity between the hippocampus and neocortex. Here we show, with direct recordings from the human hippocampus and visual cortex, that after exposure to unpredictable visual stimulus streams, hippocampal ripple activity increases in frequency and duration before stimulus presentation. Prestimulus hippocampal ripples suppress changes in visual cortex gamma activity associated with uncertainty and modulate poststimulus prediction error gamma responses in higher-level visual cortex to surprising stimuli. We reveal a function of hippocampal ripples in facilitating the propagation of visual stimuli based on the expected information gain. These results, therefore, link hippocampal ripples with predictive coding accounts of neuronal message passing and precision-weighted prediction errors, revealing a mechanism relevant for perceptual synthesis and subsequent memory encoding.