Fragmentation and multithreading of experience in the default-mode network

Humans form internal models to understand and predict future events in complex, dynamic environments. A central question is whether these models are unified or composed of multiple specialized modules with distinct inductive biases. The authors propose that world knowledge is efficiently organized through modular internal models tailored to different domains, enabling flexible inference. They introduce three abstract domains: States (contextual situation inference), Agents (reference frames for others' beliefs, intentions and goals, including self), and Actions (temporally abstract transitions or paths through state space). They hypothesize that these domains are supported by distinct midline prefrontal regions—vmPFC for state estimation, amPFC for social/reference-frame predictions, and dmPFC for action trajectory predictions—and that their outputs are integrated into a unified experience by the Precuneus, a core node of the Default-Mode Network. The study focuses on prediction errors that trigger model updates, and tests whether domain-specific belief updates map to distinct PFC activations and neural transitions, and whether the Precuneus selectively integrates these predictions with sensory information to maintain a coherent, evolving representation of the current environment.

Prior work suggests functional specialization within the midline PFC: vmPFC encodes cognitive maps and schemas and is implicated in state estimation beyond reward processing; amPFC supports social cognition, theory of mind, goal and hierarchy learning, consistent with referential modeling of agent-specific frames; dmPFC is implicated in high-level action planning, strategic decision-making, hierarchical reinforcement learning and action sequence chunking. The Default-Mode Network (DMN), with the Precuneus as a central hub, is linked to internal model processing, global integration, and interfacing between prefrontal and sensory systems; lesions to the Precuneus produce integratory deficits. Event segmentation and predictive processing literatures suggest prediction errors reorganize episodic memories and drive segmentation during naturalistic stimuli. Prior naturalistic fMRI studies using intersubject correlation/connectivity and HMMs reveal stimulus-driven shared responses, event structure discovery, and DMN reconfiguration during narrative comprehension. These strands motivate a modular predictive architecture within PFC and an integrative role for the Precuneus within DMN.

Main dataset: fMRI from the Cam-CAN project recorded while young adults (n=111; 18–35 years; TR=2470 ms multi-echo) watched an edited 8:13 min Hitchcock movie (“Bang! You're Dead”; 193 TRs; first 4 volumes discarded). Separate online groups reported subjective belief updates while watching the same movie: States (n=18), Agents (n=21), Actions (n=19); continuous Arousal ratings were obtained from a prior study (n=17). For each domain, binary button-press time-courses were boxcar-expanded with 5 s windows and averaged; group-mean update probabilities were smoothed (loess, 10% span) and downsampled to BOLD TRs. Reliability was assessed via split-half correlations with Spearman–Brown correction; weak cross-domain correlations ensured separability. Update events were discretized by thresholding group-mean time-courses: movie θ=2 SD above mean (domain updates), arousal θ=1 SD; narrative replication θ=1.2 SD. Non-update control points were selected from low-rating TRs at least one TR away from any update and matched in count. Preprocessing: SPM12 pipeline (slice timing, realignment, coregistration, normalization, smoothing 6×6×9 mm), motion regression, detrending, band-pass 0.01–0.1 Hz; participants exceeding FD >1 mm or rotation >1.5° excluded. ROIs: vmPFC, amPFC, dmPFC (Brainnetome atlas subregions), Precuneus (all 8 BN subregions), plus control ROIs (hippocampus, visual cortex, PCC, RSC, AG, MTG). Analyses: (1) Whole-brain GLM with parametric modulators (smoothed update probabilities) and contrasts States > Agents+Actions, Agents > Actions+States, Actions > States+Agents (q<0.05 FDR, k≥25). ROI-level Bayesian hierarchical regression tested domain×ROI specificity of GLM betas. (2) ISC: leave-one-out intersubject correlation computed on concatenated 7-TR windows around update points for each domain and ROI; Bayesian regression tested domain-specific increases in shared response. (3) HMM: group-level bootstrapped Hidden Markov Models per ROI to detect robust neural state transitions; median number of states chosen via BIC across 30 resampling runs (2–9 states); final HMM run 10 times, transitions averaged and binarized at 2 SD; alignment tested by counting belief updates within forward windows (1–8 TRs) after transitions; significance via 10,000 phase-randomization permutations preserving autocorrelation. (4) ISFC: intersubject functional connectivity between PFC ROIs and Precuneus, hippocampus, and among PFC nodes during domain-specific updates; Bayesian regression compared coupling forms; update vs non-update analyses assessed specificity. Seed-based whole-brain ISFC used Precuneus and domain-specific PFC seeds; group maps FDR p<0.001, visualized at r>0.1. (5) ISPC: intersubject spatial pattern correlations from unsmoothed data; 9-TR windows around updates; correlation of Precuneus ISPC with each PFC ROI ISPC during each domain’s updates tested multithreaded representational integration; permutations (1,000 phase-randomizations) assessed significance. (6) Dynamic ISC (sISC): 21-TR sliding windows for Precuneus and PFC; correlation with Arousal time-course assessed tracking of unified experience. Replication dataset: spoken narrative (“It’s Not the Fall that Gets You”, 9:07 min; n=52; TR=1500 ms; fMRIPrep preprocessed; GLM structure matched; update threshold θ=1.2 SD); ISPC multithreading and seed-based ISFC repeated to test generalization across modality and content.

- Domain-specific PFC activation: Whole-brain GLM revealed topographically distinct modulation in midline PFC: vmPFC tracked State updates; amPFC tracked Agent updates; dmPFC tracked Action updates. ROI Bayesian regression confirmed domain specificity (e.g., State updates: vmPFC > dmPFC Estimate=0.83, 95% CI 0.57–1.09; vmPFC > amPFC Estimate=1.22, 95% CI 0.96–1.48; BF>150). Agent updates: amPFC > dmPFC Estimate=0.93; amPFC > vmPFC Estimate=0.84; BF>150. Action updates: dmPFC > vmPFC Estimate=1.00; dmPFC > amPFC Estimate=0.47; BF>150. - Shared-response increases during updates: ISC rose in a domain-specific manner (e.g., State: vmPFC > dmPFC ISC Estimate=0.16, 95% CI 0.11–0.21, BF>150; vmPFC > amPFC Estimate=0.03, BF=7.35). Agents: amPFC > dmPFC Estimate=0.04, BF=15.39. Actions: dmPFC > vmPFC Estimate=0.11, BF>150. - Neural transitions align with belief updates: HMM-derived neural state transitions in vmPFC, amPFC, dmPFC were most closely followed by updates in their preferred domains across window sizes (1–8 TRs); permutation tests indicated significant domain specificity. - Precuneus integrates updated predictions: During updates, ISFC between domain-specific PFC nodes and Precuneus exceeded PFC–Hippocampus and within-PFC coupling (e.g., States: PCN–vmPFC > HPC–vmPFC Estimate=0.03, 95% CI 0.01–0.05, BF=216; PCN–vmPFC > within-PFC Estimate=0.20, 95% CI 0.18–0.22, BF>1000). Agents: PCN–amPFC > HPC–amPFC Estimate=0.06, BF>1000. Actions: PCN–dmPFC > HPC–dmPFC Estimate=0.08, BF>1000. Update vs non-update: coupling was stronger during updates (States Update–Nonupdate Estimate=0.36, 95% CI 0.30–0.42, BF>1000). - Multithreaded representational integration: Precuneus ISPC time-course correlated most with vmPFC during State updates, amPFC during Agent updates, and dmPFC during Action updates (all p<0.001), indicating prioritized, domain-selective integration; Action correlations were negative overall but least negative with dmPFC, suggesting subregional specialization within Precuneus. - Sensory integration: Seed-based ISFC showed Precuneus coupling with visual cortex during movie updates (stronger than PFC seeds), consistent with integrating top-down predictions with bottom-up visual input; this coupling was absent in the audio narrative, supporting modality-appropriate integration. - Unified experience tracking: Precuneus dynamic ISC correlated strongly with Arousal (group r=0.75, p=1.62e−31), exceeding vmPFC (r=0.58), amPFC (r=0.60), dmPFC (r=0.22); at the participant level, Precuneus ISC–Arousal correlations exceeded each PFC region. Individuals with stronger update-time PFC–Precuneus ISFC had greater whole-movie Precuneus ISC (r=0.54, p=1.4e−09), an effect absent at non-update times. - Generalization: Core findings replicated in the spoken narrative dataset (n=52): modular PFC activation maps (vmPFC/States, amPFC/Agents, dmPFC/Actions) and domain-specific Precuneus–PFC ISPC alignments (all p<0.001).

Findings support a modular architecture of internal models in midline PFC, partitioned into States, Agents and Actions, each performing distinct predictive operations during naturalistic experience. Prediction-error-driven updates triggered domain-specific activation increases and neural state transitions in vmPFC, amPFC and dmPFC, respectively. The Precuneus, a central DMN hub, selectively integrated updated predictions across domains and interfaced them with relevant sensory systems (visual cortex in the movie), yielding a coherent, unified representation that tracked subjective arousal. This multithreaded integration suggests parallel top-down prediction streams are unified in posterior DMN to shape ongoing experience, reconciling how subjective experience is continuous despite underlying modular computations. The results dovetail with literatures on cognitive maps, event segmentation, social reference frames, and action planning, and offer a network-centric account linking PFC specialization and Precuneus integration. Implications include a possible bridge in debates on conscious processing loci (prefrontal vs posterior), a mechanistic account of DMN’s role in memory and narrative comprehension, and a framework for understanding clinical phenomena involving integratory deficits. Subregional patterns (e.g., dorsal vs ventral Precuneus differences) hint at second-order specialization in integration of different model components. Replication across modalities indicates these predictive models are abstract and not tied to specific sensory textures.

The study provides convergent evidence that human world modeling is fragmented into three abstract domains—States, Agents and Actions—represented along a distinct topography in midline PFC. Prediction updates in these domains elicit domain-specific activation and neural state transitions, and are selectively unified within the Precuneus, which integrates top-down predictions with sensory input to sustain a coherent, evolving experience. This architecture generalizes across visual and auditory narratives, supporting abstract, modality-agnostic predictive models within the DMN. The work advances a unifying account of DMN function in naturalistic cognition and suggests avenues for future research: fine-grained parcellation and functional-gradient mapping of the Precuneus; designs that manipulate and temporally isolate domain-specific updates; artificial stimuli to orthogonalize domains without sacrificing ecological validity; and testing directionality and causality of PFC–Precuneus information flow.

The naturalistic design precluded measuring individual participants’ precise update timings during scanning, limiting analyses to group-level update probabilities. Most analyses are correlational; causal directionality between PFC and Precuneus cannot be established. Domain updates were not fully independent and sometimes overlapped (~25%), potentially conflating effects despite control analyses. Seed-based connectivity findings are modality-dependent (e.g., strong Precuneus–visual coupling in the movie but not in the audio narrative), and auditory cortex coupling was not observed, possibly due to intermediate lexical-semantic transformations. Negative ISPC correlations for Action updates suggest complexity and subregional heterogeneity within Precuneus requiring finer mapping. Future controlled experiments are needed to validate and disambiguate mechanisms.