{"paper_id":"628bdebf-2cba-4bcb-b297-9321699d11e7","body_text":"Fragmentation and Multithreading of \nExperience in the Default-Mode Network \n \nFahd Yazin1, Gargi Majumdar2, Neil Bramley1 & Paul Hoffman1 \n \n1 – School of Philosophy, Psychology & Language Sciences, University of Edinburgh \n2 – University of Hamburg \n \nCorresponding authors: Fahd Yazin, Paul Hoffman \n \nEmail: fahd7yazin@gmail.com, p.hoffman@ed.ac.uk  \n \n \nAbstract \n \nReliance on internal predictive models of the world is central to many theories of human cognition. Yet it \nis unknown whether humans acquired multiple separate internal models, each evolved for a specific \ndomain, or maintain a globally unified representation. Using fMRI, we show that during naturalistic \nexperiences (during movie watching or narrative listening), adult participants selectively engage three \ntopographically distinct midline prefrontal cortical regions, for different forms of predictions. Regions \nresponded selectively to abstract spatial, referential (social), and temporal domains during model updates \nimplying separate representations for each. Prediction -error-driven neural transitions in these regions, \nindicative of model updates, preceded subjective belief changes in a domain -specific manner. We find \nthese parallel top -down predictions are unified and selectively integrated with sensory streams in the \nPrecuneus, shaping participants' ongoing experience. Results generalized across sensory modalities and \ncontent, suggesting humans recruit abstract, modular predictive models for both vision and language. Our \nresults highlight a key feature of human world modeling: fragmenting information into abstract domains \nbefore global integration. \n \nIntroduction \nIn our lives, we encounter a wide range of situation s with complex and ever-changing properties \n– spatial, temporal and social. To understand and predict future events, we form internal models \nof our experiences. A well-tuned internal model 1–4 allows us to interact with the world optimally \nby generating future predictions 5,6. Understanding how these models are structured and \nrepresented are central questions in cognitive neuroscience . A hallmark of cortical computation \nis the prevalence of functionally specialized modules geared towards  processing specific kinds \nof information 7,8. This may be because different types of environmental variables r equire \ndifferent kinds of inductive biases 9 for efficient computation. For example, relational prope rties \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nlike reference frames are better captured by graph structures  whereas, temporal characteristics \nof action sequences might require a more sequential structure.  Different domains thus may \nnecessitate different prior constraints , and successful generali zation requires getting these \nconstraints right 10. Organizing world knowledge efficiently11–13(e.g. cognitive map ) might \ntherefore require a modular approach to internal model construction, leveraging  domain-\nappropriate inductive biases.  Introducing such modular principles has also been shown to \nenhance artificial learning systems14,15. Yet it is unknown whether humans acquired an assembly \nof many separable and highly specialized models, each representing a world domain, or a more \nunified global representation. Modular representation of internal models through parallelization of \ndomains would, however, create a coordination problem:  how are the contents of these distinct \nmodels unified to provide coherent behavior, let alone our integrated experience of the current \nproperties of the world16? \n \nThere are strong computational reasons to assume that humans possess multiple distinct \ncognitive maps11 or model spaces17, specialized for particular domains of the world. In this study, \nwe introduce three such models, tuned to different domains of the world: states, agents and \nactions (illustrated in Fig 1a). First, navigating a complex world requires an accurate \nrepresentation of one's current environment. However, it may not be possible to obtain (or \nobserve) all the variables required for this, requiring strong background information ( e.g., \nmemory). A mapping between prior knowledge and observed sensory information un derlies the \ninference of the environmental state. These states provide abstract contexts to situate events \nand are crucial for accurate future state predictions and learning;  their absence can bias \ninference18.  \n \nRepresenting these abstract Spatial or State models, is necessary but insufficient for a full world \nmodel. This is because states are populated by other people (or generally agents) who each \nadopt a different reference frame, and thus form distinct goals and perspectives. Modelling these \nreference frames are crucial to represent the mental states, beliefs and intentions of other agents, \nourselves and any interactions. These Agent models, may then be represented quite separatel y \nto state representations. This facilitates perspective taking and simplifies joint inference across \nvarious combinations of states and agents. In group settings, that form a large portion of human \nlife, accurate representation of relational properties of each agent to oneself and others are key. \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nThird, for a given State and Agent -reference frame, the space of (abstract) transitions or paths \none might take through them are vast. Thus, separable representation of temporally abstract \nactions, or Action mode ls allows the mapping of previously learned action path onto newly \nlearned states or agents.  In sum, we suggest three core abstract domains of cognitive \nrepresentation are needed – States, Agents, and Actions. Each operate s on different sources of \ninformation about the world. They depend on one another, but generate distinct predictions about \nthe unfolding environment. The errors in each domain demand a fundamentally different kind of \nupdate, requiring a modular architecture. Thus, a neural specialization for these domains allows \nfast and flexible inference7,8 to the near -infinite permutations of contex ts, people and plans that \nwe encounter in our lives.  \n \nDifferent sectors of midline prefrontal cortex are sensitive to such properties of world, making this \ncortical territory a likely site fo r specialized world models (Fig 1 a). State model estimation have \nbeen extensively shown to be centered around the ventromedial prefrontal cortex 19, w hich \nencodes them as cognitive maps 20 or low -dimensional schemas 21. This region is also heavily \nassociated with reward processing, however newer perspectives suggest these effects stem from \na more general function of state estimation 22. Sitting dorsal and anterior to vmPFC, the \nanteromedial prefrontal cortex is heavily involved in social cognition, theory of mind 23,24, social \nhierarchy learning25 and goal processing 26. All of these activities require referential modeling27,28 \nand computing goals (self/others). Further dorsal and posterior, the dorsomedial prefrontal cortex \nis critical for (high -level) action planning 29, strategic decis ions30, and formulating hierarchical \nplans31, all of which fall under the notion of modeling temporal properties over longer time -scales. \nTaken together, it is appealing to position this triumvirate of regions as the core model space \nwithin the midline prefrontal cortex. These also form the anterior nod es of the Default -Mode \nNetwork (DMN), which is thought to process internal models of experience12,13,16,32.  \n \nAs mentioned above, a modular architecture creates the challenge of integrating different \npredictions together into a unified format and merging  top-down priors with sensory data. \nPrecuneus, the core node of DMN 33 is in a strategic position to meet this demand due to its \nhypothesized role in global integration 34, interfacing with other cortical networks and being a \nsensory hub 16. Lesions to this region  often lead s to integratory deficits 35,36. We propose the \nPrecuneus is where distinct prefrontal predictions relatin g to states, agents and actions are \ncombined. Integrating these with sensory data allows the brain to maintain a coherent, unified, \nand up-to-date model of its physical and social environment. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nIn this study, we test this proposal: that the midline PFC regi ons operate as a partitioned domain-\nspecialized, tripartite model space. This system generates top-down predictions from each model \nin a parallel, independent manner. We focus on p rediction errors; moments when a n aspect of \none's current world model is inaccurate and thus  updated. One would expect brain regions \nsupporting a particular domain to show increased activity during such updates. Folding this notion \nonto three distinct model s means different kinds of errors trigger updates to unique parts of the \nworld model. We hypothesize that these predictions are then integrated in the Precuneus, thus \noutfitting the DMN with a modularization within its prefrontal sectors , allowing for hierarchical \ncomputation within the network.  \n \nWe explore three fundamental questions about internal model representation using fMRI data \ncollected while participants watched a short movie, where all three domains are intermingled. We \ncollected behavioral data from a different sample of participants watching the movie to determine \nwhen people generally  updated their beliefs about states (movie situations), agents (movie \ncharacters) and actions (what transpired). Using these updates as predictors of neural activity, we \ninvestigated whether humans possess a single global representation of the world model or a \nmodular, domain -specific organization. Next, through hidden Markov modeling, we explored \nwhether region -specific neural transitions coincided with  subjective belief changes. Third, we \ntested whether these domain-specific predictions are integrated within  core DMN by analysing \nshared connectivity profiles. Finally, w e replicated ou r key findings in a second dataset that \ninvolved a different sensory modality (a spoken story) and level of emotional content. Our results \nspecifically outline how the human prefrontal cortex performs domain -specific world modeling \nand, more generally, how the DMN integrates these to shape our subjective experience. \n \nResults \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nFig 1: Fragmentation of World Models. a. Any experience can be divided into models of states (abstract \ncontexts), agents (others' beliefs/goals), and actions (temporal paths through state space). The midline prefrontal \ncortex can be viewed as assembling such a model space, configuring the best model for each world abstraction. It \ngenerates top -down predictions in a tripartite organization, with domain -specific belief updates recruiting each \nregion selectively. In this example, upon visiting a friend's home for lunch, you notice their ki tchen is a mess. This \ncontrasts with your expectation , prompting a state update from 'clean' (State 2) to 'messy' (State 1). You observe \nthe friend's apparent unhappiness during cleaning, possibly due to not offering help (Agent frame 1) changing your \nrepresentation of their mood (Agent frame 2). Consequentially you consider alternative dining options than eating \nin, e.g. grabbing food from a nearby food truck (Action path 1) or restaurant (Action path 2). The experience itself \nappears fused but its deeper compositionality is implicit in the narrative structure of human experience (and later \nmemories). b. Design schematic for obtaining belief update time -courses by aggregating reported updates over \nmultiple participants. c. Smoothed, group -level belief update time-courses peaking when participants signaled \ntheir predictions were being updated in each domain.  (Inset) Movie montages  show examples of scenes during \ndifferent domain-updates. d. Interrater reliability in update time- courses computed through split-half correlation. \n \nOur approach was to use fMRI data from young participants (n=111) passively watching a \nshort movie “Bang you’re dead”, taken from the Cam -CAN project 37. We obtained \ncontinuous belief -update (prediction updates) time -courses for this movie from separate \ngroups watching online and use this for our analysis of the neuroimaging participants' BOLD \ndata. Later, we generalize the main results to a separate cohort listening to a story 38 (n=52) \nwhile undergoing scanning, using the same methods.  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nMeasuring Belief Updates in State, Agent and Action models \n \nOur analysis approach is to identify changes in neural activity, patterns and connectivity \nand, assess whether these  coincide with subjective markers of internal model updates of \nmovie content in each domain. We first probed when belief updates typically occurred in \neach domain, during the movie. Participants from three independent groups watched the \nmovie and pressed a button whenever they felt their beliefs were updated in one of the three \ndomains (Fig 1b) . Briefly, these were contextual updates 39 (for States; n = 18), belief \nupdates about people (for Agents; n=21) and belief change due to an action taken , that \ncould affect the trajectory of the movie (for Actions; n=19) (see Supplementary table 1 for \ncomplete instructions). Update time-courses from individual participants were combined and \nsmoothed to give a continuous time -course (Figure 1c).  Higher values indicate a greater \nproportion of individuals watching the m ovie marked a belief update at that point.  These \ntime-courses indicate  that the predictions (and the experience) of each domain fluctuate \nconsiderably throughout this movie.  By collecting these ratings in separate groups of \nparticipants to those who provided the fMRI data, we ensured that the fMRI data reflected \nprocessing of a fully naturalistic experience, free from instructional , meta -monitoring and \nresponding effect s. Group-level ratings were only weakly correlated with one another \n(States vs Agents, r = 0.31, Agents vs Actions, r = 0.29, Actions vs States, r = 0.28). This \nindicates that models in different domains were being updated at different times during \nmovie-watching. We also conducted  a split-half correlation analysis for each rating, which  \nindicated generally good levels of agreement between participants in the timing of updates \n(Fig 1d) (States r = 0.80, Agents r = 0.49, Actions r = 0.57, see Methods). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nTopographically distinct Internal Models in the Midline \nPrefrontal Cortex \n \n \nFig 2: Fragmentation of Predictions in Prefrontal Cortex. a. Whole brain maps (p<0.05 FDR corrected) show \na topographically distinct segmentation in the midline prefrontal nodes responding to revisions of predictions in \nActions (top), Agents (middle) and States (bottom) domains. Warm colours indicate positive effects for each domain \nrelative to the other two, and cool colours denote negative effects. b. ROI analysis confirming a domain -specificity \nwithin these regions. c. Intersubject correlation (ISC) rev eals that, in  each region, activity is more synchronized \nacross participants during updates in that regions’ preferred domain. \n \nOur neuroimaging analysis began by identifying activations that were parametrically \nmodulated by each domain -specific belief upd ate. Updates involve revision and \nreconfiguration of the current internal model. Thus, these are periods where we would \nexpect heightened processing demand when regions representing domain -specific model \ncontent should show increased activity. To test this , all three smoothed update probability \ntime-courses (Fig 1c) were used to simultaneously predict neural activity. We then \ncontrasted effects of each domain against the other two, allowing us to tease apart whether \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nthe regions were particularly involved in each type of belief update. \nWhole brain maps (Fig 2a) revealed a topographically distinct activation profile in the \nmidline prefrontal sector. vmPFC activity was most strongly correlated with State updates, \namPFC with Agent updates and dmPFC with  Action updates, in line with our hypothesized \nmodel space. A further ROI -level analysis (Fig 2b) was conducted on the effects of each \npredictor (beta values) in each region, using anatomically defined ROIs derived from the \nBrainnetome atlas40 (Supplementary Figure 1). This  analysis showed that the BOLD \nresponse to State updates was significantly higher in vmPFC than either amPFC or dmPFC \n(vmPFC > amPFC, t = 8.2548, p -value = 3.693e-13, d = 0.78, vmPFC>dmPFC, t = 5.1039, \np-value = 1.404e -06, d = 0.48), while effects of Agent updates was highest in amPFC \n(amPFC > vmPFC, t = 4.0458, p-value = 9.718e-05, d =0.38, amPFC > dmPFC, t = 4.9594, \np-value = 2.591e-06, d = 0.47) and  Action updates highest in dmPFC (dmPFC > vmPFC, t \n= 5.0828, p-value = 1.536e-06, d =0.48, dmPFC > amPFC, t =3.5133, p-value = 0.0006434, \nd = 0.33). These results suggest a topogr aphically distinct pattern of effects within the PFC \nfor different kinds of model updates. That is, different PFC regions responded to different \ntypes of prediction updates during naturalistic experience. The scanning cohort had no \ninstructions to watch the movie in any special way. Yet they showed our predicted \nseparation in the PFC when revisions of beliefs occurred (as indicated in the ratings  of \nindependent groups of participants).  \n \nDomain-specific increase in Prefrontal Shared Activity \nduring prediction updates \n \nThe activation effects suggest that specific PFC regions show heightened processing in \nresponse to domain-specific belief updates. But this does not confirm that these effects \nare truly driven by the movie stimulus alone  (rather than stimulus-unrelated internally \ndriven thoughts). In naturalistic neuroimaging paradigms, synchronization in the temporal \nprofile of activity across participants is often used as evidence that a region is engaging in \nstimulus-driven processing41. If these regions  are indeed generating top -down predictions \nthen periods of belief update should  also have particularly high levels of synchronization, \ni.e., increased shared response across participants. This is because all participants should \nupdate or calibrate their predictions in response to events in the movie in similar ways. \nTo tackle these arguments rigorously, we took a principled approach to the update ratings, \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nthat stayed constant in all further analyses unless specified otherwise. From the group - \naveraged time course, we only took the updates that crossed a specified threshold (θ), for \neach rating. We used θ as 2 SD here  (see Method for details). The underlying logic being, \nsince model updates are a subjective judgement, there is no consensus as to what exactly  \nis an objective update (no matter how many raters are present). At the same time, an \nupdate that is most prevalent across raters would have the highest likelihood to be present \nduring the experience shifts ( i.e., prediction updates) of the movie -watching cohort. This \nensures that 'base rate' of updates for each domain are respected and , only those updates \nthat were widely shared across the cohort are analyzed. \nWe then calculated the intersubject correlation (ISC) values 41 for these update periods, in \nour three PFC regions. ISC values were obtained by constructing a 7 TR (scan) window \naround these time points. These update segments were concatenated and ISC computed \non this one long segment, for each participant. \nWe predicte d that in each PFC region, ISC would be higher during its preferred domain’s \nupdates relative to other domains. We used a hypothesis driven Bayesian hierarchical \nregression42,43 to test this (see Methods). \nWe found strong evidence (Fig 2c) for these regions being synchronized in a domain \nspecific-manner across participants. In State updates, ISC showed strong eviden ce in \nfavor of vmPFC showing higher ISC than  dmPFC (vmPFC>dmPFC Estimate = 0.16, 95% \nCI 0.11-0.21, BFfor >150, P = 0.99) and amPFC (vmPFC>amPFC, Estimate = 0.03, 95% \nCI -0.01 - 0.08, BFfor = 7.35 P = 0.88). Similarly, in Agent updates, amPFC had more ISC \nthan vmPFC (amPFC>vmPFC, Estimate = 0.03 95% CI -0.02 - 0.07, BFfor = 4.81, P = \n0.83) and dmPFC (amPFC>dmPFC, Estimate = 0.04, 95% CI 0 - 0.09, BFfor = 15.39, P \n= 0.94). Finally, in Actions updates, there was evidence for dmPFC being higher than \nvmPFC (dmPFC>vmPFC, Estimate = 0.11, 95% CI 0.06 - 0.17, BFfor >150, P = 0.99) and \namPFC (dmPFC>amPFC, Estimate = 0.04, 95% CI -0.01-0.09, BFfor = 12.02, P = 0.92). \n \nThese results indicate a topographically tripartite profile in midline PFC that showed \ndomain-specific increase in activation and shared -response during moments of stimulus - \ndriven belief updates. \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nPrefrontal Neural Transitions track Experienced Belief \nupdates \n \n \n \nFig 3: Discrete Neural Transitions Precede Subjectively Experienced Belief Updates. \na. Using Hidden Markov Models, we identified neural transition points within an ROI during the movie (top row). \nWe then compared the timing of these transitions with the belief updates by counting how many updates \noccurred within different time windows (TRs /scans) following these neural transitions. This varied from 0 to 8 \nscans. As the window length increases (light grey to black), more belief updates are naturally included. However, \nthe critical prediction is that each region’s transition  windows will contain more updates from its preferred \ndomain than from the other domains. b. Proportion of belief updates that fall within neural transition windows in \neach PFC region, for various window sizes. (Top) dmPFC neural transitions are most closely aligned with  Action \nmodel updates, while amPFC (middle) and vmPFC (bottom) transitions capt ure more Agent and State updates \nrespectively. (Asterisks denote window sizes where p<0.05) \nThe previous analyses revealed increased activity when participants updated their \npredictions and consequently showed an increased shared-response in the PFC. \nUpdating one's current internal models would involve rebuilding the model representations  \non the fly  leading to newer interpretations and experience of the ongoing stimuli. If such a \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nmodel rebuilding occurs, then each update should also be associated with susta ined shifts \nto the prefrontal neural dynamics of the ongoing experience. The signatures of this could be \nextracted from the BOLD latent dynamics/transitions. To test this, we used Hidden Markov \nmodels (HMMs) to identify transitions in the neural activation  states of each PFC region. \nWe then tested  how well these transitions aligned with belief updates in the domains of \nState, Agent and Action. \n \nHidden Markov models are well -suited to tackling such problems in cognitive  neuroscience \nand has been successfully applied onto naturalistic stimuli 12,13,43. We deployed a HMM (see \nMethods) designed to be highly resistant to the timing of participant update variability. We \ndid this by exhaustively bootstrapping and model -fitting across participants over a range of \npossible latent states and, selecting the most statistically efficient number of states. The \nfinal HMM, configured with this number of states, was then estimated multiple times, \naveraging the latent state transition time points producing a transition ti me-course for each \nROI. We only used the most reliable and consistent update points for a rigorous \ncomparison (see Methods). We aimed to identify which type of belief update most closely \naligned with neural state transitions in each PFC region. To do this,  we counted the belief \nupdates occurring immediately after a neural transition, using various TR /scan window \nsizes (Figure 3a). \n \nFig 3b shows the proportion of belief updates that occurred immediately after neural state \ntransitions in each PFC region, for a range of temporal window sizes. Strong domain - \nspecificity was observed. vmPFC transitions captured State updates more than Agent or \nAction updates, amPFC was most attuned to Agent updates and dmPFC showed a \npreference for Action updates. These effects w ere largely consistent across the size of the \ntemporal windows used, but tended to be statistically significant when using longer  \ntemporal windows. This  suggests that the subjective experience of an update to \npredictions occurs sometime after the neural model is reconfigured. \nThese results emphasize that discrete shifts in the prefrontal neural dynamics coincided \nwith updated model predictions , preceding the experienced belief updates. Three \nseparate internal models  in the prefrontal cortex  appear to mediate these shifts in three \nkey domains during unguided naturalistic experience. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nPrecuneus selectively integrates updated Prefrontal \nRepresentations \n \n \nFig 4: Prioritized Integration of Prefrontal Predictions within the Precuneus. \na. Posterior estimates of different forms of functional integration. Precuneus had more shared functional \nintegration (ISFC) during the belief updates with each of the domain - specific PFC regions. Bayes factors showed \nmore evidence (see text) for such a within-DMN functional integration than an integration within the PFC ( Within \nPFC) or with Hippocampus  (PFC-Hippocampus). b. Pattern-level Integration. Inter -subject pattern correlation \n(ISPC) was computed by correlating voxel patterns at each time point in  the movie across participants. This \nprovides a time -course of shared patterns for different  ROIs. c. During moments of updates, pattern \nrepresentations in the PFC suggests new predictions. An  increased correlation of the time-course between a \nPFC region and another ROI (here Precuneus), during the update suggests functional coupling on the \nrepresentational level. If an ROI has high similarity selectively with each of the PFC region for its domain -update, \nthen it integrates new prefrontal representations in  a prioritized, multithreaded manner. Threading here refers to \nswitching between multiple prefrontal prediction threads. d. Bar plots show correlation strength of each region’s \nISPC timecourse during updates with that of the Precuneus. (Top) dmPFC displaying higher similarity with during \nAction updates than other regions. (Middle) amPFC showing more similarity than other two during Agent updates \nand (bottom) vmPFC showing similarly specificity with Precuneus during State updates \nSo far, we have provided evidence  consistent with our hypothesis that  internal world \nmodels relating to States, Agents and Actions are represented in distinct regions of PFC. \nHow do prediction threads from these  simultaneous yet spatially distinct systems get \nintegrated and distributed globally? \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nThere are two distinct subproblems here. First, prefrontal model contents must be unified \nto form a global representation of the current state of the world. This can occur withi n the \nPFC, or outside it. Second, these integrated representations must be used to constrain \nprocessing of incoming sensory information in other parts of the cortex. \n \nFunctional integration of updated predictions across domains may happen in several  \nways. First, a decentralized manner within the Prefrontal cortex, suggesting coordination \ndriven by anatomical proximity. Second, centralized integration with the Hippocampus, \nleveraging its extensive connectivity 44. Third, network integration in DMN 16,45, with the \nPrecuneus playing a pivotal role due to its integrative dynamics 16,33,35 and network \ncentrality34. \n \nTo explore the se potential integration mechanisms, we utilized intersubject -functional \nconnectivity45,46 (ISFC) analysis between domain -specific PFC regions and other ROIs \nduring belief updates. ISFC is particularly effective in naturalistic paradigms as it remov es \nstimulus-unrelated connectivity influences. This allowed us to measure and compare how \nupdated prefrontal representations integrate. Using a similar approach to the earlier ISC \nanalysis, we assessed the correlation between each PFC region and other regions during \nupdate periods for its preferred domain (e.g., during State updates for vmPFC). We \ncorrelated each PFC region with (a) Hippocampus, (b) Precuneus (PCN) and (c) other PFC \nregions. Employing a Bayesian hierarchical regression, we  then compared posterior \nevidence for each integration hypothesis: within- PFC, PFC-Hippocampus and PFC-PCN.  \n \nWe found evidence (Fig 4 a) for high er functional integration between PFC nodes and \nPrecuneus than the other two forms of integration during updates. State updates showed, \nvmPFC having more integration with PCN  than the other regions  (PCN-vmPFC>HPC-\nvmPFC Estimate = 0.03 95% CI 0.01 - 0.05, BFf or = 216.39, P = 0.99 & PCN -\nvmPFC>within-PFC Estimate = 0.2  95% CI 0.18 - 0.22, BFfor >1000, P = 0.99). Similarly, \nin Agent updates, amPFC showed more  evidence of integrating with PCN (PCN-\namPFC>HPC-amPFC Estimate = 0.06 95% CI 0.04 - 0.08, BFfor >1000, P = 0.99 & \nPCN-amPFC>within-PFC Estimate = 0.16 95% CI  0.14 - 0.18, BFfor >1000, P = 0.99). \nFinally, Action updates had dmPFC displaying more integration with PCN (PCN -\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\ndmPFC>HPC-dmPFC Estimate = 0.08 95% CI 0.06 - 0.1, BFfor >1000, P = 0.99 & PCN-\ndmPFC>within-PFC Estimate = 0.09 95% CI 0.07 - 0.11, BFfor >1000, P = 0.99). \nMoreover, the model could account for a fairly large variance in the data (Bayes adjusted \nR2 = 0.85). \n \n \nThe above results suggest prefrontal representations are integrated with in the Precuneus. \nThis is not surprising given the high connectivity between these regions.  However, it is still \nunknown how exactly this integration is carried out within the DMN. Specifically, in our \ndesign, how are these updated top-down representations unified and integrated with sensory \ndata? The Precuneus is in a position to achieve this due to its anatomical proximity with \nvisual cortex and functional coupling with the prefrontal sectors. \nGiven the parallel nature of the domains the Precuneus likely ha s access to all running \nprefrontal prediction threads. This is helpful if sensory evidence required only one domain \nto be updated, while (mostly) keeping the other two untouched. For instance, upon finding a \nrestaurant closed, one might update the action strategy without changing the reference \nframe of wanting food. We hypothesize that integration in the Precuneus follows similar \nlogic, prioritizing updates to representations with in the relevant domain of the PFC. Since \nthis involves accessing and switching between multiple running prediction threads, we term \nthis Multithreaded integration. This interpretation  allows us to directly test if such a form \nof functional integration is occurring on the neural representations during the updates. \nTo test this, we utilized intersubject spatial pattern correlation 47 (ISPC). This measure  is \nthe spatial equivalent of the ISC measure used earlier. It indexes the degree to which \nthe pattern of activation across the voxels in a region is similar across participants in \ntime (Fig 4b) . By computing ISPC in Precuneus for each TR, we can construct a time -\ncourse of shared patterns across participants. High ISPC values indicate similar shared \n(stimulus-driven) neural representations. During update time -points this suggests updated \nprediction representations in the PFC. If Precuneus performs a global integration role, then \nwe would expect its ISPC dynamics to resemble those seen in different PFC regions in a \ndomain-specific manner (Fig 4c). Around State u pdates, we would expect the Precuneus  \nISPC timecourse to be aligned to that of vmPFC, since these regions should be \nengaged in processing State-related changes. For Agent updates, it should be most similar \nto amPFC and for Action updates, to dmPFC. Th us, we tested the correlation between two \nregions' ISPC values during belief updates of each type. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nTo examine this, we computed ISPC timecourses for all regions and computed \ncorrelations between them for different segments of the movie. During State updates, \nvmPFC ISPC (Fig 4 d bottom) was more correlated with Precuneus ISPC than amPFC \nand dmPFC (p < 0.001). For Ag ent updates, amPFC ISPC (Fig 4 d middle) displayed \nstronger correlations with Precuneus ISPC than vmPFC and dmPFC (p < 0.001). Action \nupdates elicited higher dmPFC ISPC (Fig 4 d top) correlations with Precuneus ISPC than \nvmPFC and amPFC (p < 0.001).  \n \nThese results indicate that the representational dynamics of precuneus resembles that of \ndifferent PFC regions at different points during the movie, with the resemblance \ndetermined by which domain is currently engaged in model updating. To determine the \nspecificity of this result, we performed similar analyses comparing PFC regions with visual \ncortex, Hippocampus and with other parts of the DMN: Angular Gyrus, Middle Temporal \nGyrus, Retrosplenial Cortex and Posterior Cingulate (Suppl Fig 7-12). None of these \nregions showed the same domain-specific changes in ISPC correlations with PFC \nregions, suggesting that the global unification of multi-domain prefrontal predictions here is \nspecific to Precuneus. \n \n \n \nPrecuneus integrates updated Representations with \nSensory Regions \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nFig 5: Sensory regions synchronize more with Precuneus than PFC during belief updates \nIntersubject functional connectivity (ISFC) between whole -brain voxels and a. Precuneus seed, and b. domain-\nspecific PFC seed (for its domain). Values suggest an increased visual cortex functional connectivity with \nPrecuneus than PFC. Maps are visualized at (r > 0.1, p<0.001). (Top) Action updates (Middle) Agent updates \n(Bottom) State updates. \nThe above r esults suggest that Precuneus selectively unifies the updated prefrontal \nrepresentations. We hypothesize that it integrates these predictions to form a unified \nworld model. This is then used to influence and constrain processing in sensory and \nassociative regions, also shaping the ongoing subjective experience of the movie. No \nother node of the DMN, nor key sensory regions, showed the same pattern of domain - \nselective similarity with PFC . Thus, the Precuneus appears to be in a unique position to \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nintegrate top-down predictions with the sensory stimuli processed throughout the cortex \nduring these updates. \nTo test this idea, we took the  time periods when  updates occurred using the same \napproach as the preceding analysis  and performed a seed -based ISFC 41,46 analysis for \neach domain. \n \nConnectivity maps (Fig 5a) show activity in the Precuneus was correlated with \nsensory regions (Visual Cortex) during updates in each domain. In contrast, domain-\nspecific Prefrontal nodes  showed less connectivity with visual regions (Fig 5b) . Whether \nseeing in the Precuneus or PFC , coupling with a unique set of networks specific to a \ndomain occurred; Hippocampal/Parahippocampal regions during States, Tempor o-Parietal \njunction and Anterior Temporal lobe for Agents, in addition to other  heteromodal and \nassociative regions. When combined with the previous analysis,  these results indicate \nPrecuneus is highly coupled with both  the prefrontal top-down predicting r egions and \nbottom-up sensory information. This suggests a role for integrating both of these, during \nbelief updates, shaping the ongoing subjective experience. \nThe results so far are broadly consistent with our conjecture. The midline PFC represents \nthe wo rld in a modular way, fragmented into three domains, actively generating and \nadapting predictions of it. These separate classes of predictions are then unified and \nintegrated with sensory regions by the Precuneus. Such a network -level process hints to \nthe Precuneus as a hub having access to the integrated form of prefrontal predictions. \nThus, this region could potentially be an important neural correlate of unified subjective \nexperience. \n \n \n \nIntegrated representations in the Precuneus track ongoing \nsubjective experience \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nFig 6: Precuneus Unifies Fragmented Predictions into Global Experience. a. Neural transitions analysis \napplied on Precuneus and PFC regions included transitions linked with periods of heightened  Arousal, a measure \nof global experience. Arousal shifts was captured much more than each of the belief updates. PFC regions do not \ncapture Arousal more than their specialized domains. b. Correlation between Arousal  and Group-averaged ISC \ntime-courses showing Precuneus having larger correlations (r = 0.75), than dmPFC ( r =0.22), vmPFC ( r =0.58) \nand amPFC (r =0.60) c. Correlation between Arousal and participant -level ISC time-courses. Precuneus had more \ncorrelation than each of the PFC subregions. d. Relation between Pred ictions and Experience. Scatterplot shows \naverage functional connectivity between PFC and Precuneus (integration of prefrontal predictions across all \ndomains) during updates correlated with the whole -movie Precuneus ISC (movie shared experience). Dots  \nrepresent individual participants. \n \nIntegration of top -down predictions with bottom -up sensory information is key for a unified \ncurrent model of the world. Since the Precuneus connectivity seems to suggest this  \nintegration occurs here, we predicted that this region would have a unified representation of \nthe movie experience. \nPrevious studies have observed that this region had similar representations  which were  \nrelatively higher than other cortical regions, during movie watching and subsequent recall, for \nboth w ithin and across participants 49. This suggests the experiential changes due to the \nmovie might be reflected in i ts neural dynamics. We used emotional Arousal ratings as our \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nmeasure of overall movie experience. People experience high arousal around moments \nwhen they have high uncertainty that change their understanding of a situation43. Therefore, \nwe used levels of emotional arousal as a proxy for the degree to which participants are \nengaged with the unified experience of the movie. \n \nFirst, we repeated our bootstrapped HMM analysis, inv estigating whether neural state \ntransitions coincide with times when participants experience high Arousal. We predicted  \nthat Precuneus would show neural shifts linked with high arousal moments, while  \ntransitions in the prefrontal regions would be more spec ific to their respective domains (as \nshown previously). Other than setting the threshold θ to 1 SD for Arousal  (due to no \npoints surviving at 2 SD threshold we previously used; see Methods), the exact same \nprocedure was applied here. \n \nWe observed that periods of high arousal captured a strikingly large proportion of neural \nstate transitions in the Precuneus, more than periods of domain -specific belief updates  \n(Fig 6a, top) . This difference was significant throughout all the temporal windows used to \nmodel updates/high arousal. Crucially, none of the prefrontal regions showed transitions \nthat coincided with Arousal in the same way. Instead, each PFC region’s transitions \ncoincided with updates in its specific domain  (Fig 6a) . The precuneus effect persisted \nwhen we used an alternative definition of high arousal (times where arousal was greater \nthan mean arousal, rather than more th an 1 standard deviation higher than the mean) \n(Suppl Fig 6). \n \nNext, we obtained a dynamic intersubject  correlation time -course41 (sISC) for Precuneus \nand compared the Arousal time series to this. This analysis uses a sliding window to \ncompute ISC at each point in time, thus providing temporal information about shared, \nmovie-driven activity. If a region is tra cking the unified experience, then it should correlate  \nmore with Arousal,  compared to regions carrying only the fragmented experience. \nPrecuneus showed more correlation with Arousal than each of the Prefrontal subregions  on \nparticipant-level (Fig 6c) (PCN vs vmPFC t = 2.8342, p-value = 0.005467, d = 0.27, PCN \nvs amPFC t = 2.4467, p-value = 0.016, d = 0.23, PCN vs dmPFC t = 5.1409, \np-value = 1.198e -06 d = 0.49). Group -level time courses (Fig 6 b, top ) showed strong \ncorrelations between Precuneus (r = 0.75, p = 1.62e-31) and Arousal, more than each of \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nthe prefrontal regions individually  (Fig 6b) suggesting its representations are unified and \ncovaries strongly with subjective experience. (vmPFC r = 0.58, p =1.29e -16, amPFC r \n=0.60, p =1.12e-17, dmPFC r =0.22, p = 3.95e-03). \n \nSimilar Predictions Accompany Similar Experience \n \nWe further explored the relationship between the integration of prefrontal predictions in \nthe Precuneus and individuals' shared experiences of the movie. This was prompted by \nfindings that suggest a close relationship between Precuneus ISC and experienced \narousal dynamics. Specifically, we hypothesized that individuals with similar integration of \nprefrontal predictions in the Precuneus would share more similar experiences than those \nwith different integration profiles. To test this, we correlated the average ISFC b etween the \nPrecuneus and each PFC region during updates with whole-movie Precuneus ISC (Fig \n6d), which showed a significant correlation (r = 0.54, p = 1.4e -09). While various unknown \nfactors may influence shaping experience, placing this alongside our broa der results \nsuggest that similar levels of prediction integration are associated with comparable shared \nexperience. \nThe picture emerging suggests that the Precuneus unifies prefrontal predictions with \nbottom-up sensory data into a coherent, continuously ev olving experience during \nunguided natural settings. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nGeneralization to Spoken Narratives, Input Modality and \nEmotional Content \n \n \n \nFig 7: Fragmentation and Multithreading is independent of sensory modality or content. a. Whole brain \nmaps (p<0.05 FDR corrected) show a topographically distinct activation  in the midline PFC to revisions of \npredictions in States (bottom), Agents (middle) and Actions domains (top) during spoken narrative processing. b. \nDomain-specific integration between Precuneus with vmPFC during States and amPFC during  Agents, but not \nwith dmPFC during Actions (compare with Fig 4d). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \nAre these results limited to this movie's content? More importantly, is this result limited to \nthe visual modality? To test this, we replicated our two central results - modular prefrontal \npredictions and integration with the Precuneus - on a different cohort process ing a \nspoken narrative. The narrative was similar in duration to the movie  (8min vs 9min) but \ninvolved dramatically different content and emotional -valence (humorous c .f. the \nsuspenseful content of  the movie). Most importantly, it was presented as spoken audio. \nThis allows us to assess generalizability of our claims across people, content, sensory \nmodality and emotional salience (and a different pre-processing pipeline, see Methods). \nWe first tested for  activity covarying with State, Agent and Action belief updates. Whole \nbrain maps show remarkably similar prefrontal fragmentation while listening to a spoken \nnarrative, suggesting that these prefrontal modules fragment different types of experience \nin a highly consistent manner (Fig 7a). \nWe then compared whether shared patterns during updates in the PFC (via ISPC time - \ncourses) showed domain-specific alignment with the Precuneus ISPC. We found \nevidence of multithreaded integration in States (p = 0.01) and  Agents (p = 0.001), but not  \nin Actions (p = 0.225). This suggests integrated predictions in the core DMN are modality - \nagnostic and emotion -neutral i.e., abstract (Fig 7b). unlike in the movie data, here we did  \nnot find that precuneus ISPC was most correla ted with dmPFC during Action updates. This \ncould be because there were fewer characters in this story (two, one of whom narrates). \nThis may put less demands to model potential courses of action independently of the \nagents making it harder to detect effects in this domain. \n \nOverall, this replication addresses potential limitations of using a movie stimulus and \nprovides for an independent validation of some of the main results. In doing so, this \ncements our claim that humans utilize a set of modular predictiv e models in vision and \nlanguage inference during general world modelling. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nDiscussion \n \nInductive biases offer useful and computationally advantageous prior knowledge in \nstructuring internal models 10. By analyzing prediction updates across different domains \nduring naturalistic experience, we uncovered how humans might utilize such biases to \nrepresent different internal models. We suggest that humans model the world by \npartitioning it into three distinct domains, within the PFC. Each model occupie s a \ntopographically distinct portion in the midline PFC. Our analyses of fMRI movie -watching \ndata suggest that t hese three parallel neuronal syst ems adaptively guide predictions for \neach domain; namely, States, Agents and Actions. We found evidence that these top-down \npredictions are then unified in the Precuneus, the posterior hub  of DMN. We propose that \nPrecuneus continually integrates  top-down predictions with bottom-up sensory information \nto form and update the current model of the world. These results also generalized from a \nmovie to spoken narrative with very different content. This illustrates how the DMN \ncontains modular representations of abstract predictive models. \nOur results support our proposal that the joint world modelling process is divided into \ndedicated modules of States (Spatial),  Agents (Referential) and  Actions (Temporal) \nmodels. Domain activation profiles distinctly mapped to a ventral-dorsal gradient in the \nmidline PFC. These roles align with insights from various lines of work 18–30. However, to \nour knowledge they have not been previously integrated  into a unified th eoretical \nframework, localising them to prefrontal cortex. First, the vmPFC, traditionally associated \nwith reward learning and decision -making, appears to play a broader role in context -based \ninference. In our study, the vmPFC responded to context changes within the movie. In  \nmore goal-directed situations, relevant states might relate to task instructions or the reward \nvalue of different stimuli. This supports the theory that State estimation is a core \nfunction of the vmPFC 18,19,22. We generalize this notion into a model space of States \nencoded within this region. Here vmPFC not only tracks but also generates predictions \nabout various States in the environment , updating these predictions as necessary to \nnavigate experiences. \n \nThe amPFC plays a crucial role in  various forms of complex social cognition23–25. Central to \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nthese functions is the ability to construct reference frame models of Agents, enabling \ngeneralizations to new or familiar individuals across varying contexts. As a social species, \nour understanding  of the world would be dangerously incomplete without having robust \nmodels of the people around us. This allows us to anticipate their emotions and behaviour \naccurately. In the present study, amPFC activity was coupled with updates in beliefs about \nthe characters in the movie. However, amPFC is also highly engaged in reasoning about our \nown (future) mental states, suggesting that agent models also guide interpretation of  our \nown motivations and behaviours 16,26–28. It’s important to reiterate, that agent and state - \nbased predictions are often orthogonal. Individuals exhibit personality traits  that are stable \nacross various contexts, and environments possess characteristics that remain consistent \nregardless of the inhabitants. This orthogonality makes it computationally sensible to code \nState and Agent predictions separately. \n \nManaging a vast state space requires abstracting ways of transitioning, or paths across it. \nThis allows us to  navigate through various states to achieve different goals. Modeling \ntemporal properties that evolve over extended periods is crucial for this. Th e dmPFC, our \nAction model  space, plays a critical role in strategic decision -making30, hierarchical \nplanning31, and compressing action sequences over time 29. Th ese functions are vital for \nencoding and inference through abstract  Action models, where specific actions trigger \nparticular paths or sequences.  These models are built and represented separately to \nunderlying reference frames (coded by agent models) or the  contexts (coded by state \nmodels) in which they occur. In our study, a change in State or an Agent's behavio ur \ntriggers an update in the possible trajectories within the inferred story. This requires \nadjustments to the predicted 'paths' across States or fu ture agent behaviors. The ability to \ngeneralize actions provides significant advantages in adapting to new goals and \ncompositionally reusing model components elsewhere, a key aspect of human flexibility. \n \n \nOur data make a case for top -down predictions also arising in a parallel, distributed manner \nakin to bottom -up sensory pathways. Taken together, our data indicate that  PFC is a core \nregion from which top -down model predictions can arise. Importantly, it also suggests that \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nspecialization across this region is a simple yet flexible adaptation used by  the brain. This \nstrategy processes the continuous and incredibly high-dimensional world  by 'carving' it \ninto distinct domains and separately computing predictions in each50. \n \nIf our world models are represented across three modular systems, then why does our \nsubjective experience of the world not feel similarly fragmented? Our results are consistent \nwith the idea  that Precuneus unifies  the prefrontal predictions, integrating them with \nsensory data. This is not only a core node of DMN  (of which Precuneus is perhaps the \ncentral node) , but structural 34 and functional connectivity 33 data show that this region  \ninterfaces between sensory regions  and the PFC. It also acts as a connecting hub \nbetween various cortical networks 33,34. Therefore, it is well -placed to play such an \nintegrative role. The shared representations in this region were selectively aligned to that \nof each domain - specific PFC region, during their corresponding updates. Such a form of \nprioritized multithreaded integration of prediction threads was observed to be unique to \nPrecuneus, compared to a host of other regions. Studies have shown the brain could \nimplement multithreading structurally 51 and that dopamine might be functionally integrating \nmultiple threads of reward prediction errors52. Maintaining complex unified representations \nlikely requires such parallel neural architectures53 and multithreadedness can be seen as \nan adaptation to  distributed errors. This leads to robust interareal communication, further \nbolstering this region's increasing evidence in global integration16,36,46,48. \n \nWe also ruled out other  forms of regional integration such as within -prefrontal and with \nHippocampus. Consistent evidence emerged for the Precuneus, whose activity was \nattuned to discrete shifts and continuous cascades of the unified experience. Conversely, \nthe PFC was only selec tive to domain -specific shifts of experience. We found that \nactivation dynamics in precuneus aligned with ratings of emotional arousal, which index \ntemporally evolving, emotion -laden engagement with stimuli 43,54. The usually high \ncorrelation values observed in this region across subjects in studies indicate shared \nrepresentations during shared experiences 49, a proxy for the stimulus -driven states of \nexperience. Such a fun ctional manner of integrating predictions into experience might also \nunderpin neural correlates of consciousness. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nA dominant question in consciousness research is to adjudicate between various neural \ntheories of conscious processing. Currently, a major the oretical debate is whether it is \nprefrontal or posterior parietal zones that mediate access to conscious representations 55. \nOur results suggests that the answer might be a holistic gathering of both PFC and the     \nPrecuneus. Perhaps, PFC is required but ultimately generates an incomplete, coarse -grain \nexperience while parietal integration is critical for the final view. This was in line with our \nresults where individuals with similar prefrontal predictions integrated with the Precuneus  \nhad similar shared global experience. As  a consequence, it becomes difficult to falsify \ncompeting theories that have neural implementation shared with each of these regions. \nIndeed, such an ambiguous conclusion was observed in a recent adversarial experiment56 \nwhich pit these two theories against each another.  \n \nAlternatively, multiple, concurrent streams  of consciousness are central to some \nphilosophical theories of consciousness 57. Here different neural modules can have 'control' \nat different times. Implementing any theory into neuronal machinery to be called as a neural \ncorrelate of consciousness (NCC), requires satisfying several different criteria 58. One such \ncriteria is the differentiation between global and local contents. Fragmented prefrontal \nrepresentations and their eventual integration within the Precuneus might be seen as a way \nof differentiating these. Another constraint is that the NCC should be a systematically \nspecific form of conscious  processing, rather than an arbitrary or spurious neural \nassociation. In our framework, domain -specificity of these modules (e.g. conscious \nupdates to contexts vs people) satisfies such a requirement. Predictive processing \nframeworks centered on different cortical networks seems to be a promising avenue to \nexplore here. Despite domain - specificity, these regions still responded in a remarkably \nsimilar manner during updates to perceptually and emotionally different input. This suggests \nthat these representations are separated from the concrete textures of the sens es, \nsomething the DMN is in a legitimate position to fulfill. \n \nA rich literature of cognitive 12,13,16,43,45,46,49,59 and clinical studies 35,36,60 supports the role of \nDMN in higher-order human cognition. Although classically seen as task-negative, this \nnetwork is implicated in a variety of cognitive activities associated with subjectivity, such as \nmind-wandering, creative thought, self -related processing and mental time travel. These \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\ntasks are inferential in  nature and possibly involve construction of rich internal models of \nexperience. Episodic memories of expe rience are thought to be compartmentalized \nthrough event segmentations, which are functionally driven by prediction errors 39,61. Our \ndata are consistent with this broad view and provide new insights into the underlying \nmechanisms. The discrete neural transitions observed in the midline prefrontal sectors \noffer novel investigative links here. One question is whether different aspects  of episodic \nmemories (people vs places) are encoded in different ways 48. Fragmentation of \nexperience may be seen as a possible reason behind  consequential clinical accounts like \nblind-sight, spatial neglect 62, dissociative consciousness disorders, and in extreme cases \ncommissurotomy-related phenomenona63. Inability to integrate these prefro ntal predictions \ncan offer a fresh perspective in examining psychiatric conditions with independent (and \noften rebellious) 'conscious' entities within.  Indeed 'misintegration' by Precuneus, the hub \nof DMN, are well reported in clinical studies underlying r elated phenomena 35,36. Finally, \nthere is a computational formalism of seeing resting-state DMN activity as the prior models \nencoded in the cortex, under Bayesian frameworks 4,64. The present study is suggestive of \nthe dynamic processes and structural constraints by which these priors are updated as an \nexperience unfolds. \n \nFrom a methodological perspective, one strength of our study is that o ur neuroimaging \nparticipants were not give n any specific cognitive task to perform while experiencing the \nstory. Having them explicitly provide conscious ratings of their updates would have \nchanged their experience, evoking metacognitive/response-related neural signatures and \nprecluding a fully natural experience of the movie . Our design was specifically aimed to \ndetect naturally occurring predictive changes rather than perceptual changes, viewed \nthrough the lens of an individual's internal model. This approach  minimizes instructional \neffects and o ffered a window into the nature of  our internal models. Like most \nneuroimaging studies, most of our analysis are correlational, exploiting the model \ngeneration and updating processes that occur spontaneously during a naturalistic \nstimulus. In future works it will be important to exert more experimental control over the \nnature and timing of such processes, in order to validate our findings. That said, our \nreplication of the main results across two settings supports the value of studying internal \nmodels using naturalistic neuroimaging, where a suite of specific an alytic techniques has \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nbeen established in recent years. \n \nTo summarize, we  claim that humans model the world by fragmenting it into different \ndomains first  – states, agents and actions . Each of these internal models, potentially \nleveraging different kinds of inductive biases, are represented along a functionally distinct \ntopography in the PFC. Predictions threads from PFC are integrated within the Precuneus, \nenabling the DMN to shape an d transform these into a unified subjective experience. Such \nparallel, modular representations highlight the inevitability of distributed processing in the \nbrain. Through the broad framework of predictive coding and ecologically rich designs \nwe hope to have offered a novel and unificatory account of various phenomena associated \nwith the DMN, capturing its possible role in general world modelling. \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nMethods \nStimuli \nWe examined the behavioral and neuroimaging responses of participants to two \nnaturalistic stimuli –a movie and a spoken narrative  - sourced from two different datasets. \nThe movie was an edited 8 -minute excerpt from Alfred Hitchcock’s “Bang! You’re Dead!”, \nobtained from the Cam-CAN dataset37. In short, this movie involves a boy discovering \nloaded gun and pulling the trigger at various unsuspecting people. It involves various  shifts \nin context and various beliefs about the characters, and their possible actions at each \njuncture of the plot. For the narrative, we used an audio clip (9 min 7 sec)  of \"It’s Not the \nFall that Gets You\" derive d from the publicly available “Narratives” dataset 38. It concerns a \nself-narrated account of a person trying to date at a skydiving academy and the various \nbloopers that occur. Both stimuli were chosen because they are linked with publicly \navailable fMRI data from their large samples and are well -studied in the existing literature. \nMost of our analyses used the movie stimulus, while the data from the narrative served as a \nreplication dataset with different modality and content, establishing the generalization of key \nresults. \n \n \nParticipants \nOur study comprised of 8 experimental groups – one fMRI and three behavioral for each \nstimulus modality, with no overlap between them. \n129 participants provided ratings of various aspects of the stimuli. 58 UK -based \nparticipants (21 -35 years, Mean: 27.6, SD = 3.3, 47% females) were recruited via the \nonline platform Prolific to provide continuous belief update ratings, distributed into 3 \nseparate groups – States (n=18), Agents (n=21) and Actions (n=19). Additionally, \ncontinuous ratings of Arousal (n=17) were obtained from a previous study43. \nParticipants were demographically matched with the fMRI  participants (UK residents aged \n21-35). All participants were native English speakers and had normal or corrected -to- \nnormal vision and hearing. All participants reported no previous history of watching the \nmovie. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nSimilarly, prediction ratings on the narrative were collected from 3 independent groups \n(21-35 years, Mean: 28.5, SD = 3.5, 44% females) of participants using Prolific, for States \n(n=16), Agents (n=18) and Actions (n=20). The participants reported no previous history of \nlistening to the story. \n \nThe online rating studies were approved by University of Edinburgh School of Philosophy, \nPsychology & Language Sciences Research Ethics Committee and all participants gave \nwritten informed consent. \n \n \nNeuroimaging data sources \nfMRI data for the movie were obtained from the healthy population -derived Cam -CAN \ncohort (N=135, Age:18-35 years). \nParticipants exceeding a maximum framewise displacement of 1 mm or angular rotation > \n1.5° were excluded from subsequent analyses (see fMRI preprocessing) leading to a \nsample of 111 participants (µage  = 28.5 ± 4.93, 63 female) for further analyses, as \ndescribed in a previous study 43. Functional MRI scans were acquired on a 3T Siemens \nTIM Trio system, using a T2* -weighted multi-echo pulse sequence with a TR of 2470 ms, \nand multiple TEs of 9.4 ms, 21.2 ms, 33 ms, 45 ms, 57 ms, a flip angle of 78°, and 32  \naxial slices. The movie -watching session lasted for 8 minutes and 13 seconds  (493 \nseconds), yielding 193 TRs, of which we discarded the first 4 volumes from all analyses. \nThe narrative dataset 38 consisted of preprocessed (fMRIPrep) 3 -T (Siemens Magnetom \nSkyra) fMRI T2* weighted BOLD responses (TR 1500ms) from 52 participants (18 -29 \nyears, µage  = 28.5 ± 4.93, 31 female). Functional BOLD images were acquired in an \ninterleaved fashion using gradient -echo echo -planar imaging (EPI) with an in -plane \nacceleration factor of 2 using GRAPPA: TR/TE = 1500/28 ms, flip angle = 64°, bandwidth = \n1445 Hz/Px, in-plane resolution = 3 × 3 mm, slice thickness = 4 mm, matrix size = 64 × 64, \nFoV = 192 × 192 mm, 27 axial slices with roughly full brain coverage and no gap, anterior – \nposterior phase encoding, prescan normalization and fat suppression. The functional \nscanning session included 400 TRs, totaling 600 seconds of acquisition time. \n \n \nBelief Update Time-courses \nThe stimuli were presented to the participants using Testable (tes table.org). Before the  \nstart of the experiment, participants were given instructions that included a definition of \nthe type of beliefs they were being asked to monitor (State/Agent/Action). An example of a \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nsituation in which they would need to update those beliefs (from a scenario unrelated to \nthe stimuli) was also included. Participants were asked to focus on the stimuli \n(movie/narrative) and press a key immediately when they felt their beliefs in the domain of \ninterest had been updated. Full i nstructions given to the participants for all three domains \ncan be found in Suppl Table 1. \nFor each participant, this yielded a binarized time -course indicating the points at which  \nthey signaled that they had experienced a belief update. The precise timing  of updates \nvaried between participants. Even when two participants experienced an update in \nresponse to the same event in the story, they might respond at slightly different times due \nto differences in their speed of responding, attentiveness and threshol d for deciding an \nupdate has occurred. To accommodate this variability, we created a 5 second update \nwindow around each belief update point, generating a boxcar timeseries for each \nparticipant with a value of 1 for the temporal regions at which they experi enced updates \nand 0 for all other times. We then averaged these time -series over participants. The  \ngroup-mean time -course then represents, for each point during the movie, the likelihood \nthat a person watching the movie would be experiencing an update. Group-averaged \ntime-courses for each domain were then smoothed using local regression smoothing  \n(loess in R, a non - parametric approach) with a 10% span which was kept the same for \nboth movie and story. These smoothed ratings were downsampled to match the fMRI \nBOLD time-course, providing an update probability for each image acquired during movie - \nwatching (see Figure 1C). \nWe computed the correlation between all three update timecourses to check for any \nstrong (r>0.5) coupling between any two domains, which might render the neural analysis \nless feasible. Additionally, an inter -rater consistency (IRC) analysis was performed to \nassess the reliability of ratings. For each domain, the participant pool was randomly split \ninto two subgroups. Each subgroup’s time -courses were smoothed, averaged and their \ngroup-mean correlated with the other subgroup's mean. This process was repeated 100 \ntimes for each domain in both movie and narrative. \nThe smoothed update probability time -courses were used to predict BOLD activation  in \nunivariate analyses. Other analyses (HMM, ISC, ISFC, ISPC; see below) required \ndiscrete estimates of periods when updates occurred. To generate these, for each \ndomain, we applied a threshold θ to the smoothed group -averaged timeseries. Time \nperiods whe re update probability exceeded θ were set to 1 and rest to 0. Where a \nsequence of multiple TRs /scans exceeded θ, we selected the final TR /scan in the \nsequence to represent the update. We adopted a θ of 2 SD above the mean update \nprobability for all movie update time-courses and θ of 1 SD above the mean value for the  \nArousal ratings \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n(since no timepoints were 2 SD higher than the mean). Update timecourses for the \nnarrative used a θ of 1 SD as there was less variation in the update timecourses for th is \nstimulus. \n \n \nArousal Ratings \nThe continuous ratings of Arousal were collected as part of a previous study 43. Participants \nwere instructed to watch the movie while continuously rating it with respect to their \nemotional intensity (Arousal) using their mouse on a vertical slider ranging from -1 to 1. \n \n \nfMRI preprocessing \nThe fMRI data for the movie were processed using SPM12 using a standard  \npreprocessing pipeline consisting of slice time correction, realignment, co -registration, \nnormalization and smoothing (6mm x 6mm x 9mm FWHM), as described in a previous \nstudy43. To account for motion artifacts, six parameters capturing translation and rotation \n(X-, Y -, Z -displacements, and pitch, roll, yaw) were removed from the functional data \nthrough least -squares regress ion. Participants exceeding a maximum framewise \ndisplacement of 1 mm or angular rotation > 1.5° were excluded from subsequent  \nanalyses. The processed functional data then underwent voxel -wise detrending and was \nsubjected to a band -pass filter between 0.01 and 0.1 Hz, implemented using a second - \norder Butterworth filter. \nThe narrative data we obtained were already preprocessed using fMRIPrep 38. These \nimages were smoothed with a 8mm x 8mm x 8mm FWHM kernel for the GLM. \nFor all the voxel pattern analyses, unsmoothed voxel-wise BOLD time series were used. \n \n \nRegion of Interest (ROI) definition \nAll ROI masks were defined from the Brainnetome  Atlas40. We created vmPFC ROI by \ncombining 6 subregions in the Orbital Gyrus region. S imilarly, amPFC and dmPFC ROIs \nwere created by combining 6 and 4 (non -overlapping) subregions from the Superior \nFrontal Gyrus. Precuneus (PCN) ROI was constructed by combining all 8  subregions in \nthe Precuneus region. The four primary ROI masks (PFC and PC N) pertaining to our \noverarching hypothesis, are visualised in Supplementary Fig 1. Other \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nregions used in control analyses (Hippocampus, Visual Cortex, Posterior Cingulate Cortex, \nRetrosplenial Cortex, Angular Gyrus and Middle Temporal Gyrus ) were similarly \nconstructed from various subregions as used in a previous study43, and was mainly used as \ncontrol to establish specificity of effects. \n \n \nGeneral Linear Model (GLM) \nTo estimate the neural responses corresponding to the State, Agent and Action updates, \nwe modelled the whole -brain movie viewing BOLD timeseries by fitting  a general linear \nmodel (GLM). It included a regressor modeling the temporal dy namics of the entire movie \nas multiple events with 1 second duration, convolved with the canonical hemodynamic \nresponse function (HRF) to model the expected BOLD response. Additionally, the GLM \nincorporated the three group - averaged smoothed timeseries of the prediction updates \nas parametric modulators (mean-centered). This allowed us to assess how fluctuations in \nthese prediction time - courses modulated brain activity throughout the movie. Six head -\nmotion parameters defined by the realignment were added to the model as nuisance \nregressors. \nWe assessed 3 statistical contrast maps – States>Agents+Actions, \nAgents>Actions+States and Actions>States+Agents  with weights [1 -0.5 -0.5] for each \nmap. All contrast images were obtained at participant level and a group -level random - \neffects analysis was conducted. Thereafter, we thresholded the statistical maps at q  < 0.05 \nFDR with an extent threshold of 25 voxels (k =  25) which was performed using NeuroElf  \n(http://neuroelf.net). \nThe same GLM structure was employed for the Narrative data. \n \n \nIntersubject Correlation Analysis (ISC) during \nUpdates \nFor each domain (State. Agent and Action), we first constructed a 7 scan(TR) window \naround the update points previously identified i.e. 3 scan before and after update scans, \nwithout any overlapping segments. Thereafter these segments were extracted and \nconcatenated and used to compute a leave -one-out ISC on each of the concatenated \ntimecourses in each PFC ROI. To do this, we computed the average BOLD timecourses \nin each ROI and calculated the correlation of these timecourses for each participant with \nmean of all other participants (minus the selected participant). This method ensured any \nshared neural activity was a result of stimuli -driven processing than any idiosyncratic \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nnoise. The State update timecourses consisted of 63 scans, Agents had 28 scans and \nAction had 35 scans. \n \nBayesian Regression for ISC \nISC values of each PFC region during the updates were tested in a Bayesian hierarchical \nregression model using participant’s ISC values as the dependent variable. This modelled \nthe effects of Domain (States, Agents, Actions), ROI (vmPFC, amPFC, dmPFC) and their \ninteraction on ISC. It included random effects of participants including random slopes for \nthe fixed effects to capture participant-level variance to these terms. \nThe model equation was therefore Y ij=β0 + β1(Domain)ij+ β2(ROI)ij+ β3(Domain×ROI)ij+ γ0j+ \nγ1j(Domain)ij+ γ2j(ROI)ij+ εij Where: \nYij is the ISC outcome for subject j in observation i. β0 is the overall intercept. \n \nβ1, β 2, and β 3 are the fixed effects coefficients for Domain, ROI, and their interaction, \nrespectively. \nγ0j is the random intercept for subject j. \n \nγ1j and γ 2j are the random slopes for the effects of Condition and ROI within subject j, \nrespectively. \nεij is the residual error for subject j in observation i. \n \nPosterior estimation was done using the “brms” package in R through Hamiltonian Monte \nCarlo for 2 chains of 6000 samples with the first 500 discarded from each. Model \nconvergence was assessed through R-hat statistics, which were found to be \n~1.00, sufficiently large estimated sample size for stable posterior estima tes, and by \nvisually inspecting the chains for convergence and large autocorrelations. Posterior \npredictive checks for model validation were conducted by simulating 1,000 samples from \nthe posterior and by fitting with the empirical data (Suppl Fig 2, 3). \nWe were mainly interested in testing the domain-sensitiveness of each of these three ROIs \nfor their respective domain. That is if vmPFC has higher ISC than dmPFC and amPFC \nduring State updates, amPFC higher ISC than vmPFC and dmPFC during Agent updates \nand dmPFC higher ISC than vmPFC and amPFC during  Action updates. For this, we \ncomputed Bayes Factors for (and against) these hypotheses and their associated  \nposterior probabilities. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n \n \nHidden Markov Modelling \nHidden Markov Models  (HMMs) have been exte nsively used to compute neural state \ndynamics from BOLD data, and have been particularly powerful in analysing naturalistic \nstimuli that involve the maintenance of temporally extended internal states 12,13,43. Our \ngoal was to \nestimate the timing of BOLD hidden latent transitions in each region, and compare that to \nthe belief update points, and contrast it across the neural regions and type of belief \nupdates. We used R package depmix6 for constructing the HMMs. \nWe aimed to implement a group -level HMM. After extracting the ROI BOLD time series for \neach subject, we fit our bootstrapped HMM as follows. Although HMMs are usually fitted \nby specifying a number of discrete neural states and then selecting based on some \npenalizing information criteria, we face a pr oblem of intersubject variability here. That is, \ndifferent subjects would have variability in update timing for each domain, on each ROI. \nThis meant that the transition points of group -level HMM should be robust to this factor. \nFirst, we determined the num ber of neural states present in each ROI. For each ROI, we \nrandomly removed 10 subjects from the dataset (~10% of data) and estimated HMMs over \na range of pre -specified latent states from 2 - 9 (for a total of 8 pre - specified states), from \nwhich the best-fitting number of states was adjudicated using Bayesian Information Criteria \n(BIC). This was repeated 30 times with 10 random subjects removed in each run, for a total \nof 300 'subjects' removed or ~thrice the full dataset  (n=111). This resulted in a vector of \nthe best fitting number of states for each HMM run (i.e. 30 such values, one for each \nrun). We took the median as the number of states for the final analysis. \nNext, using this median number of states, we fit the group -level HMM and obtained the \nposterior trajectories using the Viterbi algorithm. Since parameter estimation in HMMs \n(Expectation Maximization algorithm) is inherently stochastic (unless one seeds it), it \noutputs the transitions between the states with slight temporal differences. To e nsure \nrobustness against this, we ran this final HMM (using the median number of states) 10 \ntimes, and obtained an averaged transition time course, which indicates the probability of \nstate transitions at each point in time (0 -1). Higher values suggest the transition point is \nconsistent across the 10 runs, and hence robust for this ROI. This time course was then \nbinarized by subjecting it to a threshold of 2 standard deviation from its mean (exactly like \nthe belief update time courses), thus identifying scan s in which neural state transitions \noccurred. This allowed us to not incorporate the individual variability in neural transitions \nin a conservative and quantifiable manner, as well as accounting for the stochasticity in in \ntransition points which is a cons equence of Viterbi algorithm, (rather than just seeding it \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nwith an arbitrary seed). Thus, our bootstrapped HMM ensured 1) most reliable number of \nlatent states are acquired, 2) with the most consistent transition times obtained for each \nROI, 3) while prese rving the update dynamics of each ROI individually, in the most \nstatistically efficient manner. Across run similarity and optimal states distributions can be \nseen in Suppl Fig 4 & 5. \n \n \nTo investigate the degree to which belief updates were coincident with neural state shifts \nidentified by the HMM, we constructed forward windows of duration t scans beginning at \neach neural shift in the final HMM state transition time series. We then counted what \nproportion of belief updates of each type occurred within the s can windows following HMM \ntransitions (see Fig 3a).  We repeated this process for windows of size 0 to 8 scans to \nensure that results were robust across different temporal windows. \nTo assess significance, we used non -parametric permutation testing. We repeated each \nanalysis 10,000 times, each time randomly shuffling the time -series of experiential shifts \n(belief updates), while keeping the neural shifts constant. This yielded a distr ibution of \nresults under the null hypothesis that the timing of belief updates was unrelated to neural \nshifts. We were then able to assess where our observed result fell within the null \ndistribution and assign a p -value accordingly. For example, a observed  result more \nextreme than 95% of null results would receive a p-value of 0.05. \nWe used this method to test the hypothesis that HMM neural transitions were more likely \nto coincide with updates in a region’s preferred domain than in the other two domains. Th e \ntest statistic here was conditional on the criterion above (i.e., p_Rate1>p_Rate2 & \np_Rate1>p_Rate3, where Rate1 is the domain-specific Rating to each ROI while other \ntwo are the 'nonspecific' ratings). \nThe exact same approach was later deployed for the HMM involving Arousal as well, \nwherein the hypothesis then becomes p_Arousal > p_Rate1, p_Arousal > p_Rate2, \np_Arousal > p_Rate3 to underscore the specialization to the integrated form (i.e.,  Arousal) \nthan the fragmented form. \n \n \nIntersubject Functional Connectivity (ISFC) Analysis \nfor functional integration \nIntersubject functional connectivity (ISFC) is measure of the correlation of the BOLD time - \nseries in a region in one participant to the group -averaged BOLD time -series (without this \nparticipant) to a second region. It is particularly useful in naturalistic stimuli du e to its \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nrobustness to stimulus -unrelated processing (i.e. noise). We first used this method to \ninvestigate functional connectivity profiles between the three PFC regions with Precuneus, \nHippocampus as well as within them during periods of State, Agent and Action updates. \nWe computed the ISFC employing a similar approach to that of the ISC, by concatenating \nthe 7-scan windows around the update timepoints (3.5 -scan before and after) for each of \nState, Agent and Action. This allowed us to assess ISFC between pairs of regions in period \nwhere each type of updates was occurring. For each participant, ISFC between a pair of \nROIs (i and j) was computed by taking their concatenated BOLD timecourse for the \nselected updates for ROI i and correlating it with average ROI j time-series for the other \nparticipants. \nFor Prefrontal Integration, we averaged the ISFC between the three PFC ROIs. This \ninvolved averaging the ISFC between vmPFC -amPFC and vmPFC -dmPFC during State \nupdates, amPFC-vmPFC and amPFC -dmPFC during Agent updat es, and dmPFC -amPFC \nand dmPFC-vmPFC during Action updates. \nFor Hippocampal and Precuneus Integration, this was computed with the hypothesized \ndomain-specific ROI – with vmPFC during States, amPFC during Agents and dmPFC \nduring Actions. \n \n \nBayesian Regression for ISFC \nThe Bayesian Regression for ISFC integration during updates was constructed and \nexecuted in the exact manner as for the ISC (and with the same MCMC settings), with the \nthree Domains being replaced by PCN -PFC, PFC-HPC and Within -PFC Integration a s the \nthree 'Integration' terms to be tested. \nWe were mainly interested in hypothesis testing of which form of functional integration \nwas higher in each domain. That is if PFC -Precuneus has higher ISFC than PFC - \nHippocampus and Within -PFC during each of St ate, Agent and Action updates. We \ncomputed Bayes Factors for (and against) this hypothesis and their associated posterior \nprobabilities. \n \n \nIntersubject Pattern Correlation Analysis \nThis analysis used ISPC to compare the timing of across -subject alignment of voxel \npatterns between Precuneus and PFC regions. \nWe first obtained the ISPC (intersubject spatial pattern correlations) from the unsmoothed \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nBOLD time -series. For each participant , this was achieved by extracting the multivoxel \npattern vector at each time point, and correlating that with the group -mean (sans this \nparticipant), for a given ROI. This results in a time -course for this participant which reflects \nthe spatially correlate d pattern dynamics for that region. From this, for each participant  \n(and for each ROI) we selected three different subsets of this timeseries. In each domain, \nwe identified update points as described earlier and constructed a 9 -scan window around \nthem (4.5-scan before and after) resulting in as many segments as there were updates. \nWe extracted data from these segments and concatenated them to give ISPC timecourses \nfor the periods of State, Agent and Action updates. To compare how similar the ISPC was \nacross two regions, we then computed the correlation of these timecourses across \npairs of regions (e.g., computing Precuneus -vmPFC ISPC correlation for each update \ntype). \nThe main ISPC analysis was done comparing each of prefrontal nodes with Precuneus. \nTo establish the specificity of the results, PFC regions were also compared with a number \nof other ROIs (VC, HPC, AG, MTL, RSC & PCC as shown in Suppl Fig 7 - 12). \nWe inferred multithreading  if a region showed domain-selective increase of simila rity in \nISPCs with the PFC. This similarity was obtained by computing the relative increase in \ncorrelation of a prefrontal region during its prefered updates (e.g. ROI - vmPFC ISPC \ncorrelation is higher during State updates than ROI -amPFC/dmPFC). Prediction s were \ntested by computing contrasts between conditions, of the same form as those used in the \noriginal whole -brain GLM. For example, for States this was achieved by comparing \nvmPFCStates - ((vmPFCAgents+vmPFCActions)/2) with amPFCStates - \n((amPFCAgents+amPFCActions)/2) and dmPFCStates -\n((dmPFCAgents+dmPFCActions)/2). These contrasts compute the relative increase in ISPC \ncorrelation with the Precuneus for State updates and test whether this is higher for vmPFC  \nthan the other two PFC regions. Similar logic was applied in Agents and Actions \nSignificance was assessed via non -parametric permutation tests. We shuffled the ISPC \ntime-courses of the prefrontal nodes, breaking any temporal associations within these, \nwhile keeping the Precuneus (or control ROI) fixed. \nThe above -mentioned contrasts were computed in the permuted data (for 1000 \npermutations) to generate a distribution of expected values under the null hypothesis.  As \nbefore, a p- value for the contrast was calculated using the proportion of permuted values \nthat exceeded the true value.  \n \nSeed-based ISFC \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nWe used this method to investigate ISFC between the Precuneus and other brain regions \nduring periods of State, Agent and Action updates. Precuneus BOLD time -series were  \nextracted from the same ROI mask  we used in all other analyses and compared to the \ngroup-average time -series of every other voxel in the brain. We used the AAL atlas to  \nobtain a whole-brain mask. For computational tractability, voxels were resampled at 6mm \nx 6mm x 6mm, such that the seed  time course was compared to 6814 voxels throughout  \nthe brain. We computed the ISFC for periods of State, Agent and Action updates by \nemploying a similar approach to that of the ISC, i.e., concatenating the 7 -scan windows \n(3.5-scan before and after) around the update timepoints of each type. For each participant, \nISFC between Precuneus  and each voxel was computed by correlating the participant’s \nPrecuneus BOLD time -series with the group -averaged voxel time -series (i.e. Precuneus to \nvoxel). We then used the same approach to compute seed based ISFC for each of the PFC \nregions. vmPFC was used as the seed in State updates, amPFC was used as the seed i n \nAgent updates and dmPFC was used as the seed in Action updates \nIndividual participant’s correlation maps were averaged to give a group -level correlation \nmap. These whole brain ISFC maps were visualized at p < 0.001 (with a threshold of \nr>0.1). \n \n \nCorrelations between Subjective Experience and \nPredictions \nFor estimating  the correlations between prediction integrations and experience, we \nobtained the correlations between Precuneus and each of the domain specific PFC \nregions during the respective updates (from the ISFC analysis), and then averaged these \nto get a single val ue for prediction integration strength between PFC and Precuneus \nduring the updates. This measure was then correlated with the whole movie ISC values \nof Precuneus. The dynamic variant of the latter was shown to be highly correlated with \nArousal, making it a proxy of the shared, unified conscious experience. \n \nReplication of key findings in spoken narrative \ndataset \n \nWe sought to replicate the two key findings observed in the movie - modular predictions \nand multithreaded integration. For this, we used the narr ative data. Belief update time -\ncourses were obtained from 3 independent groups of participants from Prolific, for States \n(n=16), Agents (n=18) and Actions (n=20). The exact same instructions were used (aside \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nfrom changing the 'watching a short movie’ part in the instructions to 'listening to a short \nstory') \n \nSimilar to the movie, 5 second update window around each update point were used \ngenerating a boxcar time-series for each participant, which were averaged and smoothed \nin the same manner. \n \nThe three domain-specific belief updates were used as predictors of the narrative neural \ndata \"It’s Not the Fall that Gets You\" (9 min 7s). The exact same GLM structure used for \nthe movie was deployed for this with three resulting contrast maps \n(States>Agents+Actions, Agents>Actions+States and Actions>States+Agents). Statistical \nmaps were thresholded at q < 0.05 FDR with an extent threshold of 25 voxels (k = 25) \nwhich was performed using NeuroElf. \n \nFor the multithreading/ISPC analysis, we obtained ISPC from the unsmoothe d BOLD \ndata. We obtained the update -time points akin to the movie as well with one difference. \nThe narrative used a θ of 1 SD threshold as there was less variation in the update \ntimecourses for this stimulus. We constructed the same 9 -scan window around th ese \nupdate points as in the movie. Relative correlations between PCN ROI with that of \ndomain-specific PFC regions were conducted and significance assessed via non -\nparametric permutation tests where we shuffled the ISPC time -courses of the prefrontal \nnodes, breaking any temporal associations within these, while keeping the Precuneus (or \ncontrol ROI) fixed. \n \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nReferences \n \n1. Daw, N. D., Gershman, S. J., Seymour, B., Dayan, P. & Dolan, R. J. Model -Based Influences on \nHumans’ Choices and Striatal Prediction Errors. Neuron 69, 1204–1215 (2011). \n2. Gläscher, J., Daw, N., Dayan, P. & O’Doherty, J. P. States versus Rewards: Dissociable Neural \nPrediction Error Signals Underlying Model-Based and Model-Free Reinforcement Learning. Neuron 66, \n585–595 (2010). \n3. Lee, T. S. & Mumford, D. Hierarchical Bayesian in ference in the visual cortex. JOSA A 20, 1434–1448 \n(2003). \n4. Fiser, J., Berkes, P., Orbán, G. & Lengyel, M. Statistically optimal perception and learning: from \nbehavior to neural representations. Trends Cogn. Sci. 14, 119–130 (2010). \n5. Aitchison, L. & Le ngyel, M. With or without you: predictive coding and Bayesian inference in the brain. \nCurr. Opin. Neurobiol. 46, 219–227 (2017). \n6. Clark, A. Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behav. \nBrain Sci. 36, 181–204 (2013). \n7. Bertolero, M. A., Yeo, B. T. T. & D’Esposito, M. The modular and integrative functional architecture of \nthe human brain. Proc. Natl. Acad. Sci. 112, E6798–E6807 (2015). \n8. Meunier, D., Lambiotte, R. & Bullmore, E. T. Modular and Hierarchically Modular Organization of Brain \nNetworks. Front. Neurosci. 4, (2010). \n9. Zador, A. M. A critique of pure learning and what artificial neural networks can learn from animal brains. \nNat. Commun. 10, 3770 (2019). \n10. How to Grow a Mind: Statistics, Structure, and Abstraction | Science. \nhttps://www.science.org/doi/10.1126/science.1192788. \n11. Behrens, T. E. J. et al.  What Is a Cognitive M ap? Organizing Knowledge for Flexible Behavior. \nNeuron 100, 490–509 (2018). \n12. Baldassano, C. et al. Discovering Event Structure in Continuous Narrative Perception and Memory. \nNeuron 95, 709-721.e5 (2017). \n13. Baldassano, C., Hasson, U. & Norman, K. A. Re presentation of Real-World Event Schemas during \nNarrative Perception. J. Neurosci. 38, 9689–9699 (2018). \n14. Zhang, R., Pitkow, X. & Angelaki, D. E. Inductive biases of neural network modularity in spatial \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nnavigation. Sci. Adv. 10, eadk1256 (2024). \n15. D’Ambrosio, D. B. et al.  Achieving Human Level Competitive Robot Table Tennis. Preprint at \nhttps://doi.org/10.48550/arXiv.2408.03906 (2024). \n16. Yeshurun, Y., Nguyen, M. & Hasson, U. The default mode network: where the idiosyncratic self \nmeets the shared social world. Nat. Rev. Neurosci. 22, 181–192 (2021). \n17. FitzGerald, T. H. B., Dolan, R. J. & Friston, K. J. Model averaging, optimal inference, and habit \nformation. Front. Hum. Neurosci. 8, (2014). \n18. Expectancy-related changes in firing of dopamine neurons  depend on orbitofrontal cortex | Nature \nNeuroscience. https://www.nature.com/articles/nn.2957. \n19. Schuck, N. W., Cai, M. B., Wilson, R. C. & Niv, Y. Human Orbitofrontal Cortex Represents a \nCognitive Map of State Space. Neuron 91, 1402–1412 (2016). \n20. Constantinescu, A. O., O’Reilly, J. X. & Behrens, T. E. J. Organizing conceptual knowledge in \nhumans with a gridlike code. Science 352, 1464–1468 (2016). \n21. Mack, M. L., Preston, A. R. & Love, B. C. Ventromedial prefrontal cortex compression during \nconcept learning. Nat. Commun. 11, 46 (2020). \n22. Schuck, N. W., Wilson, R. & Niv, Y. A state representation for reinforcement learning and decision -\nmaking in the orbitofrontal cortex. in Goal-directed decision making: Computations and neural circuits  \n259–278 (Elsevier Academic Press, San Diego, CA, US, 2 018). doi:10.1016/B978 -0-12-812098-\n9.00012-7. \n23. Jenkins, A. C. & Mitchell, J. P. Mentalizing under Uncertainty: Dissociated Neural Responses to \nAmbiguous and Unambiguous Mental State Inferences. Cereb. Cortex N. Y. NY 20, 404–410 (2010). \n24. van Veluw, S. J. & Chance, S. A. Differentiating between self and others: an ALE meta -analysis of \nfMRI studies of self-recognition and theory of mind. Brain Imaging Behav. 8, 24–38 (2014). \n25. Park, S. A., Miller, D. S. & Boorman, E. D. Inferences on a multidimensiona l social hierarchy use a \ngrid-like code. Nat. Neurosci. 24, 1292–1301 (2021). \n26. D’Argembeau, A. et al. The Neural Basis of Personal Goal Processing When Envisioning Future \nEvents. J. Cogn. Neurosci. 22, 1701–1713 (2010). \n27. Muysers, H. et al.  A persiste nt prefrontal reference frame across time and task rules. Nat. \nCommun. 15, 2115 (2024). \n28. Gusnard, D. A., Akbudak, E., Shulman, G. L. & Raichle, M. E. Medial prefrontal cortex and self -\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nreferential mental activity: Relation to a default mode of brain func tion. Proc. Natl. Acad. Sci. 98, 4259–\n4264 (2001). \n29. Ostlund, S. B., Winterbauer, N. E. & Balleine, B. W. Evidence of Action Sequence Chunking in \nGoal-Directed Instrumental Conditioning and Its Dependence on the Dorsomedial Prefrontal Cortex. J. \nNeurosci. 29, 8280–8287 (2009). \n30. Venkatraman, V., Rosati, A. G., Taren, A. A. & Huettel, S. A. Resolving Response, Decision, and \nStrategic Control: Evidence for a Functional Topography in Dorsomedial Prefrontal Cortex. J. Neurosci. \n29, 13158–13164 (2009). \n31. Ribas-Fernandes, J. J. F. et al. A Neural Signature of Hierarchical Reinforcement Learning. Neuron \n71, 370–379 (2011). \n32. Margulies, D. S. et al. Situating the default-mode network along a principal gradient of macroscale \ncortical organization. Proc. Natl. Acad. Sci. 113, 12574–12579 (2016). \n33. Utevsky, A. V., Smith, D. V. & Huettel, S. A. Precuneus Is a Functional Core of the Default -Mode \nNetwork. J. Neurosci. 34, 932–940 (2014). \n34. Yamaguchi, A. & Jitsuishi, T. Structural connectivity of the precuneus and its relation to resting -\nstate networks. Neurosci. Res. (2023) doi:10.1016/j.neures.2023.12.004. \n35. Cavanna, A. E. & Trimble, M. R. The precuneus: a review of its functional anatomy and \nbehavioural correlates. Brain 129, 564–583 (2006). \n36. Kumral, E., Bayam, F. E. & Özdemir, H. N . Cognitive and Behavioral Disorders in Patients with \nPrecuneal Infarcts. Eur. Neurol. 84, 157–167 (2021). \n37. Shafto, M. A. et al.  The Cambridge Centre for Ageing and Neuroscience (Cam -CAN) study \nprotocol: a cross -sectional, lifespan, multidisciplinary ex amination of healthy cognitive ageing. BMC \nNeurol. 14, 204 (2014). \n38. Nastase, S. A. et al. The “Narratives” fMRI dataset for evaluating models of naturalistic language \ncomprehension. Sci. Data 8, 250 (2021). \n39. Yazin, F., Das, M., Banerjee, A. & Roy, D.  Contextual prediction errors reorganize naturalistic \nepisodic memories in time. Sci. Rep. 11, 12364 (2021). \n40. Fan, L. et al.  The Human Brainnetome Atlas: A New Brain Atlas Based on Connectional \nArchitecture. Cereb. Cortex 26, 3508–3526 (2016). \n41. Nastase, S. A., Gazzola, V., Hasson, U. & Keysers, C. Measuring shared responses across \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\nsubjects using intersubject correlation. Soc. Cogn. Affect. Neurosci. 14, 667–685 (2019). \n42. Chen, G. et al. An integrative Bayesian approach to matrix -based analysis in ne uroimaging. Hum. \nBrain Mapp. 40, 4072–4090 (2019). \n43. Majumdar, G., Yazin, F., Banerjee, A. & Roy, D. Emotion dynamics as hierarchical Bayesian \ninference in time. Cereb. Cortex 33, 3750–3772 (2023). \n44. Eichenbaum, H. Prefrontal–hippocampal interactions in episodic memory. Nat. Rev. Neurosci. 18, \n547–558 (2017). \n45. Brandman, T., Malach, R. & Simony, E. The surprising role of the default mode network in \nnaturalistic perception. Commun. Biol. 4, 1–9 (2021). \n46. Simony, E. et al.  Dynamic reconfiguration of t he default mode network during narrative \ncomprehension. Nat. Commun. 7, 12141 (2016). \n47. Meshulam, M. et al. Neural alignment predicts learning outcomes in students taking an introduction \nto computer science course. Nat. Commun. 12, 1922 (2021). \n48. Silson, E. H., Steel, A., Kidder, A., Gilmore, A. W. & Baker, C. I. Distinct subdivisions of human \nmedial parietal cortex support recollection of people and places. eLife 8, e47391 (2019). \n49. Chen, J. et al. Shared memories reveal shared structure in neural activity across individuals. Nat. \nNeurosci. 20, 115–125 (2017). \n50. Parr, T., Sajid, N. & Friston, K. J. Modules or Mean-Fields? Entropy Basel Switz. 22, 552 (2020). \n51. Senden, M., Goebel, R. & Deco, G. Structural connectivity allows for multi -threading during rest: \nthe structure of the cortex leads to efficient alternation between resting state exploratory behavior and \ndefault mode processing. NeuroImage 60, 2274–2284 (2012). \n52. Takahashi, Y. K. et al.  Dopaminergic prediction errors in the ventral tegmental area reflect a \nmultithreaded predictive model. Nat. Neurosci. 26, 830–839 (2023). \n53. Griffa, A. et al. Evidence for increased parallel information transmission in hu man brain networks \ncompared to macaques and male mice. Nat. Commun. 14, 8216 (2023). \n54. Dmochowski, J. P., Sajda, P., Dias, J. & Parra, L. C. Correlated Components of Ongoing EEG \nPoint to Emotionally Laden Attention – A Possible Marker of Engagement? Front. Hum. Neurosci.  6, \n(2012). \n55. Seth, A. K. & Bayne, T. Theories of consciousness. Nat. Rev. Neurosci. 23, 439–452 (2022). \n56. Consortium, C. et al. An adversarial collaboration to critically evaluate theories of consciousness. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint \n\n2023.06.23.546249 Preprint at https://doi.org/10.1101/2023.06.23.546249 (2023). \n57. Dennett, D. C. Consciousness Explained. xiii, 511 (Little, Brown and Co, New York, NY, US, 1991). \n58. Hohwy, J. & Seth, A. Predictive processing as a systematic basis for identifying the neural \ncorrelates of consciousness. Philos. Mind Sci. 1, (2020). \n59. Patel, T., Morales, M., Pickering, M. J. & Hoffman, P. A common neural code for meaning in \ndiscourse production and comprehension. NeuroImage 279, 120295 (2023). \n60. Vanhaudenhuyse, A. et al. Default network connectivity reflects the level of consciousness in non -\ncommunicative brain-damaged patients. Brain 133, 161–171 (2010). \n61. Zacks, J. M. & Swallow, K. M. Event Segmentation. Curr. Dir. Psychol. Sci. 16, 80–84 (2007). \n62. Corbetta, M. & Shulman, G . L. SPATIAL NEGLECT AND ATTENTION NETWORKS. Annu. Rev. \nNeurosci. 34, 569–599 (2011). \n63. de Haan, E. H. F. et al.  Split-Brain: What We Know Now and Why This is Important for \nUnderstanding Consciousness. Neuropsychol. Rev. 30, 224–233 (2020). \n64. Pezzulo, G., Zorzi, M. & Corbetta, M. The secret life of predictive brains: what’s spontaneous \nactivity for? Trends Cogn. Sci. 25, 730–743 (2021). \n \n \n \n \nAcknowledgements \n \n \nWe'd like to thank Rob McIntosh, Chris Lucas, Chris Summerfield, Karl Friston, Joszef Fiser, Mihalyi \nBanyal and Adam Koblinger for helpful discussions. The authors would like to dedicate this to the memory \nof Daniel Dennett. PH was supported by a BBSRC grant (BB/T004444/1). For the purpose of open access, \nthe author has applied a Creative Commons Attribution (CC BY) license to any Author Accepted \nManuscript version arising from this submission. \n \n  \nEthics declarations \n  \nThe authors declare no competing interests. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 25, 2024. ; https://doi.org/10.1101/2024.10.24.620113doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}