Pith. sign in

REVIEW 4 major objections 4 minor 52 references

A three-layer active-inference model reproduces expert–novice meditation dynamics by treating thoughtseeds as latent causes and meta-awareness as a gating signal.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 00:55 UTC pith:JX6CHQR6

load-bearing objection A polished and transparent demonstration architecture, but the headline 'reproduces empirical findings' is circular: the expert/novice contrasts are written into the prior tables before the run. the 4 major comments →

arxiv 2607.14833 v1 pith:JX6CHQR6 submitted 2026-07-16 q-bio.NC

Thoughtseeds as Latent Causes: A Dual-Process Computational Phenomenology of Focused-Attention Meditation

Prakash Chandra Kavi , Daniel Ari Friedman , Gustavo Patow This is my paper
classification q-bio.NC
keywords computational phenomenologyactive inferencefocused-attention meditationthoughtseedsmeta-awarenessmind-wanderingglobal workspacedual-process cognition
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that focused-attention meditation can be simulated as a hierarchy of prediction and control—an active-inference process—rather than described only in words or fMRI correlations. Its central move is to treat inner mental content as a small set of latent causes, called thoughtseeds, that sit between fast physiological network activity and a slow, capacity-limited metacognitive monitor. The proposed architecture reproduces the four canonical meditation states and, after training, yields expert and novice profiles that match empirical findings: experts stay on the breath longer, wander less, recover through a clean sequence, and suppress default-mode activity. If the model is right, it provides an implementation-level bridge from first-person experience to objective neurophysiological measures, and makes meta-awareness a computable signal rather than a private sensation.

Core claim

Four canonical states of focused-attention meditation—breath focus, mind-wandering, meta-awareness, redirect attention—are modeled as attractors of a three-layer hierarchical generative model. A physiological substrate evolves over attentional networks; a low-dimensional set of learned latent causes ('thoughtseeds') summarizes that substrate, proposes actions by minimizing expected free energy, and sends predictions downward; a metacognitive monitor combines habit priors with policy evidence, with meta-awareness as a gated divergence between the two. Variational EM reproduces expert/novice patterns: longer breath-focus dwell, less mind-wandering, a recovery loop, and default-mode suppression

What carries the argument

The load-bearing mechanism is the nested Markov-blanket hierarchy. Layer 1 is a switching multivariate Ornstein-Uhlenbeck process over four attentional networks; its attractors define each meditation regime, and it is the generative process that receives no learning. Layer 2 is a low-dimensional latent space of 'thoughtseeds'—learned latent causes of mental content—which reconstructs Layer 1 activity, infers a discrete state belief, evaluates stay/switch policies via expected free energy, and issues descending predictions. Layer 3 implements the workspace bottleneck: meta-awareness is an ignition signal computed as the divergence between policy evidence and a learned habit prior, multiplied

Load-bearing premise

The load-bearing premise is that the expert–novice contrast is placed into the simulation before it runs: the model's hand-set settings for network strengths, state durations, and which state follows which are chosen to match the very expert/novice differences the simulation then reports; if those settings were made identical, the reported contrast would mostly disappear.

What would settle it

Equalize all phenotype-specific parameter settings across the two simulated groups and re-run the same training protocol; if expert/novice differences persist, the hand-set priors are not the cause, and if they vanish, the architecture alone does not produce the contrast. A complementary empirical check is to estimate dwell/transition parameters from individual meditation fMRI runs and compare them with the hand-set expert/novice tables.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The four canonical meditation states are not arbitrary labels but controllable attractors of one hierarchy, so changing a single layer should shift dwell times and transitions across all states.
  • Meta-awareness becomes an operational, computable quantity that can be correlated with breath-count accuracy, experience-sampling reports, or neurophysiological markers.
  • Expert–novice differences reduce to a small set of learnable parameters, suggesting meditation training could be quantified as movement in this parameter space.
  • Because automatic content generation and executive gating live in different layers, the model offers a concrete mechanism for why mind-wandering is hard to notice.
  • The framework extends beyond meditation to any setting where attention lapses, spontaneous thought, and metacognitive recovery interact.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension is to sever the Layer 3 gate while keeping Layer 2 intact: the model predicts the expert's clean recovery loop should degrade to novice-like diffuse switching even though content inference still works.
  • The paper leaves implicit that novice-to-expert change is primarily a change in embodied attractor structure and learning rate, not in metacognitive priors—if so, training interventions should target attentional stability and interoception before metacognitive strategy.
  • A decisive test the authors have not run is to estimate dwell-time and transition priors from individual practitioners' data and feed them into the same architecture; if expert/novice differences then persist without phenotype-specific settings, the hand-set priors would be validated as empirical measurements rather than assumptions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a three-layer active-inference architecture for focused-attention meditation: L1 is a multivariate Ornstein–Uhlenbeck process over four attentional networks, L2 implements a low-dimensional 'thoughtseed' generative model (System 1), and L3 implements a Global Neuronal Workspace-style metacognitive monitor (System 2). The authors report simulations of expert and novice phenotypes that, they claim, reproduce empirical findings on dwell times, transition patterns, network activation profiles, and latent-space geometry across the four canonical states (breath focus, mind-wandering, meta-awareness, redirect attention). The mathematical specification is explicit in the Supplementary Material, including equations and parameter tables, and the code is released.

Significance. If the central claim were established, the model would offer an implementation-level account linking first-person phenomenology, attentional regulation, and large-scale network dynamics. The paper is also valuable for its explicit mathematical scaffolding, the attempt to formalize meta-awareness as a gated policy-prior divergence, and its release of simulation code. However, the central validation claim is not supported: the principal expert/novice contrasts are effectively written into phenotype-specific priors before the simulation runs, and the one-run-per-phenotype design provides no uncertainty quantification. The architecture is coherent as a computational sketch, but the reported results are not a test of that architecture.

major comments (4)
  1. [§4 / Tables S4–S6 / Eqs. (S2.1), (S4.5)] The expert/novice differences reported as simulation results are direct consequences of phenotype-specific priors. Section 4 states that 'phenotype differences live only in L1 attractors, L1 coupling/stiffness, dwell/transition priors, and learning rate'. Eq. (S2.1) uses μx(s) from Table S4; S1.2 samples dwell durations from Table S5; Eq. (S4.5) builds the policy posterior from a dwell-aware prior whose switch probabilities come from Table S6. The longer expert Breath-Focus dwell and shorter Mind-Wandering dwell in Fig. 3B, and the MW→MA / RA→BF recovery loop in Fig. 3C, are therefore direct consequences of the priors, not emergent results of free-energy minimization or the L3 GNW gate. The abstract and Section 4.1 present these as reproductions of empirical findings; as designed, the simulation cannot validate that claim.
  2. [§S1.5] The discrete regime label st is supplied to L2 as an external conditioning variable: 'the discrete regime label st is also supplied to L2 as an external conditioning variable by the simulation controller.' The model is therefore told which of the four canonical states it is in; the state sequence is not inferred from L1 network dynamics. This weakens the claimed 'tractable link between first-person phenomenology and objective neurophysiological measures', because the central phenomenological variable is an input rather than an inference or an action outcome.
  3. [§3 / Table S2 / §5.1] Only one run per phenotype is reported (seed 42; 12,000 steps). There are no error bars in Figs. 3–5, no sensitivity analysis over the prior tables, and no ablation that equalizes Tables S4–S6 across phenotypes. Section 5.1 acknowledges that dwell ranges 'should be calibrated against empirical measures', but no calibration, model comparison, or quantitative fit to the cited empirical studies is provided. Because the priors already encode the expected differences, the observed contrast is not informative about the three-layer architecture.
  4. [Fig. 3C / Table S6] Fig. 3C reports expert MW→MA = 1.00, MA→RA = 1.00, and RA→BF = 1.00, whereas Table S6 sets the corresponding prior probabilities to 0.75, 0.85, and 0.80. This discrepancy is unexplained: if the priors are the cause of the recovery loop, the simulated transition matrix should track Table S6; if the loop is learned, the mechanism should be explained. Either way, the single-run transition matrix is uninterpretable without uncertainty quantification.
minor comments (4)
  1. [Fig. 3C] The caption says 'self-transitions excluded', but some rows in the displayed matrices do not sum to unity (e.g., novice MW: 0.29 + 0.43 + 0.29 = 1.01). The normalization should be clarified.
  2. [Eqs. (S3.4), (S6.1)] The E-step free-energy objective in Eq. (S3.4) contains only accuracy and complexity terms, while the training loss in Eq. (S6.1) adds forward surprisal and a recognition term. The relationship between the E-step objective and the optimized loss should be stated explicitly.
  3. [§4.1.1] The sentence 'Profiles are emergent time-averaged activations; config defines attractor predictions' is unclear; 'config' appears to refer to configuration files, but this is not explained.
  4. [§2.1] The conceptual parallels with pratītyasamutpāda, ālaya-vijñāna, and related constructs are interesting but not operationalized; consider separating motivational framing from formal definitions.

Circularity Check

3 steps flagged

Expert/novice contrasts are written into phenotype-specific priors, so the reported 'reproduction' is an input–output tautology.

specific steps
  1. fitted input called prediction [Section 4 / §S1.2 / Tables S5–S6 / Eq. (S4.5)]
    "States are piecewise-constant with dwell durations sampled from Table S5. Once dwell elapses, a regime transition is attempted with probability ˜pt = max{1−q(πt = stay),1/N dwell}, ... the successor regime is sampled from q(πt) over switch candidates (see §S1.4, Eq. (S4.5)). ... Table S5: Breath Focus (BF) Novice 10–18, Expert 15–25; Mind Wandering (MW) Novice 15–25, Expert 10–18. Table S6: Expert MW→MA 0.75, RA→BF 0.80; Novice MW→MA 0.60, RA→BF 0.40."

    Fig. 3B's longer expert BF dwell and shorter expert MW dwell, and Fig. 3C's clean expert MW→MA→RA→BF recovery loop, are obtained by sampling dwell durations from Table S5 and by routing Table S6's transition probabilities through pdwell(switch-to s′) = dt2·P(s′|s) in Eq. (S4.5). The phenotype-specific differences reported as empirical consistencies are therefore the same values already inserted as priors.

  2. fitted input called prediction [§S1.2 Eq. (S2.1) / Table S4 / Fig. 3A]
    "dxt =−Θ(s t) [xt −µ x(st)] dt+σ x dWt,(S2.1) where Wt is a 4D Wiener process, µx(s) is the state-conditioned attractor (the fixed point of the OU drift for state s; Table S4). ... Table S4: Breath Focus (BF): Novice DMN 0.50 DAN 0.58 FPN 0.60; Expert DMN 0.35 DAN 0.65 FPN 0.70."

    L1 is the generative process and receives no M-step updates, yet its fixed points are Table S4's phenotype-specific attractors. Fig. 3A's 'stronger DMN suppression, higher DAN/FPN' expert profiles and higher novice DMN in MW are, up to coupling/noise/clipping, exactly these prescribed µx(s) values. The reported network-activation 'findings' are thus direct renderings of Table S4, not emergent results of free-energy minimization.

  3. self definitional [§S1.3 Eqs. (S3.1), (S3.2), (S3.5) / Table S7 / Figs. 4–5]
    "dzt =−Θz(zt −µz(st)) dt+σz dWt (S3.1) ... p(zt |st) = N(zt;µz(st), σ2z I) (S3.2) ... q(st) = softmax(− ∥zt −µz(s)∥2/τz) (S3.5)."

    The latent thoughtseed zt is evolved toward µz(st) by Eq. (S3.1), and the state belief q(st) is defined as a softmax over distances to the same µz(s) set in Eq. (S3.5). Consequently the L2 state separation, the L2 thoughtseed trajectories, and the PCA structure in Figs. 4–5 are self-confirming displays of Table S7's prior distances rather than independent inferences. The external supply of st (§S1.5) further ensures the model is conditioned on the very labels whose empirical frequencies are fixed by Tables S5–S6.

full rationale

The architecture itself is not circular: the EM updates of encoder/decoder/forward model and the L3 policy equation are specified in their own right, and the thoughtseeds concept is not justified solely by self-citation. The circularity lies in the central empirical claim. Section 4 states that 'Phenotype differences live only in L1 attractors, L1 coupling/stiffness, dwell/transition priors, and learning rate; L2/L3 priors are shared.' Eq. (S2.1) makes Table S4's attractors the fixed points of L1, so Fig. 3A's network profiles are injected means. §S1.2 samples dwell durations from Table S5, and Eq. (S4.5) feeds Table S6's transition priors into pdwell, so Fig. 3B/C's dwell times and recovery loop are prescribed. Eqs. (S3.1) and (S3.5) make L2's latent trajectories and state beliefs self-confirm Table S7, while §S1.5 states that 'the discrete regime label st is also supplied to L2 as an external conditioning variable by the simulation controller.' The paper's own Section 5.1 concedes that parameter choices 'should be calibrated against empirical measures' but provides no such calibration, model comparison, or uncertainty quantification (one run, seed 42). With equalized priors, the reported expert/novice contrast would largely vanish. The result is therefore forced by construction for the key empirical differences, although nontrivial shared L2/L3 learning machinery remains, so the score is 8 rather than 10.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The central result depends on a large set of hand-set, phenotype-specific parameters (Tables S4–S6) that already encode the empirical expert/novice contrasts the paper claims to reproduce. Domain assumptions about active inference, GNW, NPH, and the four-state cycle are invoked without independent evidence. Thoughtseeds are latent constructs with no external falsifiable handle provided here.

free parameters (6)
  • Network attractor profiles μx(s) by phenotype (Table S4) = e.g., BF DMN 0.50 novice vs 0.35 expert; MW DMN 0.82 vs 0.70
    Hand-set per state and phenotype to encode expected expert/novice network profiles; L1 is then attracted to these values (Eq. S2.1), so reported activation differences are near-preordained.
  • Dwell time ranges per state and phenotype (Table S5) = BF novice 10–18, expert 15–25; MW novice 15–25, expert 10–18; MA/RA 5–10 vs 3–6
    Dwell durations sampled from these ranges feed the dwell-aware prior (Eq. S4.5); the simulated expert dwell-time differences directly inherit them.
  • Exit transition priors P(s'|s) per phenotype (Table S6) = e.g., expert RA→BF 0.80, novice RA→BF 0.40; expert MW→MA 0.75 vs novice 0.60
    The transition matrix used to construct switch probabilities (Eq. S4.5, Table S6) already contains the canonical recovery loop; the simulation's transition matrices reflect it.
  • Thoughtseed priors μz(s) (Table S7) = e.g., BF attend_breath 0.85, MW pain_discomfort 0.65, MA aha_moment 0.85, RA equanimity 0.85
    State beliefs q(s) are softmax distances to these priors (Eq. S3.5); the latent trajectories are attracted to them via Eq. S3.1.
  • L1 coupling matrix Θ(s) (Table S3) plus expert modifications = diagonal base 0.50 in MW else 0.15; off-diagonals like DAN→DMN 0.60 in BF; expert BF δθ=0.4; global scale κθ 1.0/1.1
    Hand-chosen competitive/cooperative couplings; the expert BF self-stabilization directly encodes stiffness differences.
  • Core training hyperparameters = σ²=0.002, lr 0.01/0.02, BPTT T=25, VI steps=2, clip [0.05,0.9], λmin=0.05, PRECISION TAU, NOISE LEVEL
    Chosen by hand; not fitted to empirical data; affect convergence and smoothness rather than the qualitative expert/novice story.
axioms (6)
  • domain assumption Three-layer nested Markov-blanket hierarchy (networks → thoughtseeds → metacognition) is the correct decomposability of the meditating agent.
    Section 2.2 and Fig. 2 assume this architecture without deriving it from lower-level biology.
  • domain assumption Global Neuronal Workspace (GNW) capacity bottleneck and ignition dynamics describe conscious access.
    L3 implements GNW as the metacognitive gate (Section 2.2, Eq. S4.4); GNW is one contested consciousness theory but treated here as given.
  • domain assumption The four canonical meditation states (BF, MW, MA, RA) are exhaustive, discrete, and mutually exclusive.
    Explicitly stated in Section 2.2: 'one cannot simultaneously be in Breath Focus and Mind-Wandering, serving as an important model assumption.'
  • domain assumption Free Energy Principle and active inference are valid normative descriptions of cognitive agents.
    Section 1 treats the FEP as the overarching framework; the validity of the framework itself is not argued in this paper.
  • domain assumption Neuronal Packet Hypothesis (NPH) with transient Markov blankets underlies functional brain organization.
    Section 2.1 grounds L1 in NPH (refs [24,25]); this is a specific neuro-computational hypothesis, not standard settled neuroscience.
  • ad hoc to paper Dwell-aware quadratic hazard function enforces biological refractory pacing.
    Eq. S4.5 defines pdwell(stay)=1−d_t²; introduced to reproduce dwell behavior but not derived from data or from a more basic principle.
invented entities (1)
  • Thoughtseeds as latent causes no independent evidence
    purpose: Low-dimensional latent variables that generate/encode mental content and bias policies in L2 (System 1).
    No direct neural correlate or external behavioral measurement is given in this paper; they are inherited from the authors' prior work (refs [33,34]) and operationalized via hand-specified priors (Table S7).

pith-pipeline@v1.3.0-alltime-deepseek · 18977 in / 15960 out tokens · 133133 ms · 2026-08-02T00:55:17.606778+00:00 · methodology

0 comments
read the original abstract

Meditative expertise involves sustained attention, rapid recovery from distraction, and coordinated dynamics of large-scale brain networks. We present a computational phenomenology of focused-attention meditation traversing four attractor states: breath focus, mind-wandering, meta-awareness, and redirect attention. Within a dual-process active inference formulation, the model implements a three-layer nested Markov-blanket architecture: (L1) a high-dimensional physiological neuronal substrate modeled as a stochastic multivariate Ornstein--Uhlenbeck process over attentional Yeo networks; (L2) a low-dimensional generative model (System 1) that encodes latent mental content as thoughtseeds and evaluates autonomic action tendencies; and (L3) an agentic metacognitive monitor (System 2) that implements a Global Neuronal Workspace (GNW) capacity bottleneck to selectively gate these tendencies. In L3, meta-awareness functions as the GNW ignition signal, derived from policy-prior divergence and dynamically gated by direct competition between orchestrator and distractor thoughtseeds. Policy selection actively minimizes expected free energy, and L2 actions furnish descending predictions over network activity to close the enactive perception--action cycle. Training uses variational Expectation-Maximization (EM) across expert and novice phenotypes. Simulations reproduce behavior consistent with empirical observations and findings in contemplative neuroscience, providing a tractable link between first-person phenomenology and objective neurophysiological measures.

Figures

Figures reproduced from arXiv: 2607.14833 by Daniel Ari Friedman, Gustavo Patow, Prakash Chandra Kavi.

Figure 1
Figure 1. Figure 1: The Canonical Cycle of Focused-Attention Meditation. The cycle is conceptualized as four distinct phases [9]: (1) Breath Focus (BF), characterized by sustained attention on the breath (recruiting the dorsal attention network, DAN, and frontoparietal network, FPN) [6, 12]; (2) Mind-wandering (MW), where attention involuntarily drifts to self-referential thoughts (dominated by the default mode network, DMN) … view at source ↗
Figure 2
Figure 2. Figure 2: Conceptual Architecture. Layer 1: Physiological Neural Substrate (Internal Generative Process) This foundational layer represents the self-organization of neuronal packets into functional superordinate ensembles [24, 25, 27] and is described using Yeo attentional networks for neurobiological plausibility and model tractability [42]. Layer 2: Low-Dimensional Thought￾seeds Network (System 1: Generative Model… view at source ↗
Figure 3
Figure 3. Figure 3: Diagnostic Views of Stabilized Profiles. (A) Network activation profiles. Average activations for expert and novice phenotypes across the four canonical stages. (B) Dwell times. Average step durations for each state, demonstrating prolonged focus and shorter regulatory lapses in experts. (C) Transition dynamics. Transition matrices (self-transitions excluded) highlighting the structured recovery loop in ex… view at source ↗
Figure 4
Figure 4. Figure 4: Hierarchical dynamics with continuous thoughtseed trajectories. Across panels, L3 shows meta-awareness, L2 shows continuous thoughtseed trajectories, and L1 shows network activations (DMN/VAN/DAN/FPN). (A) Novice. Hierarchical profiles illustrating high volatility and prolonged Mind-Wandering. (B) Expert. Hierarchical profiles illustrating structural coherence and stable Breath-Focus. 4.1.2 Hierarchical Dy… view at source ↗
Figure 5
Figure 5. Figure 5: PCA Trajectories Across the Hierarchy. Plot-window trajectories (final 2,000 steps) projected using pooled PCA. Top row: L2 thoughtseed trajectories (novice left, expert right). Bottom row: L1 network trajectories (novice left, expert right). The separation between L2 and L1 trajectories underscores the hierarchical architecture, illustrating that low-dimensional bottleneck dynamics remain more orderly and… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 22 canonical work pages

  1. [1]

    2021 Meditation and the wandering mind: a theoretical framework of underlying neurocognitive mechanisms.Perspect

    Brandmeyer T, Delorme A. 2021 Meditation and the wandering mind: a theoretical framework of underlying neurocognitive mechanisms.Perspect. Psychol. Sci.16, 39-66. (doi:10.1177/1745691620917340)

    work page doi:10.1177/1745691620917340 2021

  2. [2]

    2015 The neuroscience of mindfulness meditation.Nat

    Tang YY, H¨ olzel BK, Posner MI. 2015 The neuroscience of mindfulness meditation.Nat. Rev. Neurosci.16, 213-225. (doi:10.1038/nrn3916) 13

    work page doi:10.1038/nrn3916 2015

  3. [3]

    2015 Reconstructing and deconstructing the self: cognitive mech- anisms in meditation practice.Trends Cogn

    Dahl CJ, Lutz A, Davidson RJ. 2015 Reconstructing and deconstructing the self: cognitive mech- anisms in meditation practice.Trends Cogn. Sci.19, 515-523. (doi:10.1016/j.tics.2015.07.001)

    doi:10.1016/j.tics.2015.07.001 2015

  4. [4]

    Biobehav

    Fox KCR, Nijeboer S, Dixon ML, Floman JL, Ellamil M, Rumak SPet al.2014 Is medi- tation associated with altered brain structure? A systematic review and meta-analysis of morphometric neuroimaging in meditation practitioners.Neurosci. Biobehav. Rev.43, 48-73. (doi:10.1016/j.neubiorev.2014.03.016)

    work page doi:10.1016/j.neubiorev.2014.03.016 2014

  5. [5]

    Lutz A, Abdoun O, Dor-Ziderman Y, Trautwein FM, Berkovich-Ohana A. 2024. An overview of neurophenomenological approaches to meditation and their relevance to clinical research. Biol. Psychiatry Cogn. Neurosci. Neuroimaging 10, 411–424. (doi:10.1016/j.bpsc.2024.11.008)

    doi:10.1016/j.bpsc.2024.11.008 2024

  6. [6]

    Lutz A, Slagter HA, Dunne JD, Davidson RJ. 2008. Attention regulation and monitoring in meditation.Trends Cogn Sci12, 163–169. (doi:10.1016/j.tics.2008.01.005)

    work page doi:10.1016/j.tics.2008.01.005 2008

  7. [7]

    Laukkonen RE, Slagter HA. 2021. From many to (n)one: Meditation and the plas- ticity of the predictive mind.Neuroscience & Biobehavioral Reviews128, 199–217. (doi:10.1016/j.neubiorev.2021.06.021)

    work page doi:10.1016/j.neubiorev.2021.06.021 2021

  8. [8]

    2019Satipat.t.h¯ ana meditation: a practice guide

    An¯ alayo B. 2019Satipat.t.h¯ ana meditation: a practice guide. Cambridge, UK: Windhorse Publications

  9. [9]

    2012 Mind wandering and attention during focused meditation: a fine-grained temporal analysis of fluctuating cognitive states.Neuroimage59, 750–760

    Hasenkamp W, Wilson-Mendenhall CD, Duncan E, Barsalou L W. 2012 Mind wandering and attention during focused meditation: a fine-grained temporal analysis of fluctuating cognitive states.Neuroimage59, 750–760. (doi:10.1016/j.neuroimage.2011.07.008)

    work page doi:10.1016/j.neuroimage.2011.07.008 2012

  10. [10]

    Zamora-Lopez G, Chen Y, Deco G, Kringelbach ML, Zhou C. 2016. Functional complexity emerging from anatomical constraints in the brain: The significance of network modularity and rich-clubs.Sci Rep6, 38424. (doi:10.1038/srep38424)

    doi:10.1038/srep38424 2016

  11. [11]

    Barrett LF, Simmons WK. 2015. Interoceptive predictions in the brain.Nat Rev Neurosci16, 419-429. (doi:10.1038/nrn3950)

    work page doi:10.1038/nrn3950 2015

  12. [12]

    MacLean KA, Ferrer E, Aichele SR, Bridwell DA, Zanesco AP, Jacobs TL et al. 2010. Intensive meditation training improves perceptual discrimination and sustained attention. Psychol. Sci. 21, 829–839. (doi:10.1177/0956797610371339)

    doi:10.1177/0956797610371339 2010

  13. [13]

    2015 The wandering brain: meta-analysis of functional neuroimaging studies of mind-wandering and related spontaneous thought processes.Neuroimage111, 611-621

    Fox KCR, Spreng RN, Ellamil M, Andrews-Hanna JR, Christoff K. 2015 The wandering brain: meta-analysis of functional neuroimaging studies of mind-wandering and related spontaneous thought processes.Neuroimage111, 611-621. (doi:10.1016/j.neuroimage.2015.02.039)

    work page doi:10.1016/j.neuroimage.2015.02.039 2015

  14. [14]

    Christoff K, Irving ZC, Fox KCR, Spreng RN, Andrews-Hanna JR. 2016. Mind- wandering as spontaneous thought: a dynamic framework.Nat Rev Neurosci17, 718–731. (doi:10.1038/nrn.2016.113)

    work page doi:10.1038/nrn.2016.113 2016

  15. [15]

    Seeley WW, Menon V, Schatzberg AF, Keller J, Glover GH, Kenna H et al. 2007. Dissociable intrinsic connectivity networks for salience processing and executive control. J. Neurosci. 27, 2349–2356. (doi:10.1523/JNEUROSCI.5587-06.2007)

    work page doi:10.1523/jneurosci.5587-06.2007 2007

  16. [16]

    Schooler JW, Smallwood J, Christoff K, Handy TC, Reichle ED, Sayette MA. 2011. Meta- awareness, perceptual decoupling and the wandering mind. Trends Cogn. Sci. 15, 319–326. (doi:10.1016/j.tics.2011.05.006) 14

    work page doi:10.1016/j.tics.2011.05.006 2011

  17. [17]

    An¯ alayo B. 2021. Dependent arising and interdependence.Mindfulness12, 1094–1102. (doi:10.1007/s12671-020-01544-x)

    work page doi:10.1007/s12671-020-01544-x 2021

  18. [18]

    Escrichs A, Sanju´ an A, Atasoy S, L´ opez-Gonz´ alez A, Garrido C, C` amara E et al. 2019. Characterizing the dynamical complexity underlying meditation. Front. Syst. Neurosci. 13, 27. (doi:10.3389/fnsys.2019.00027)

    arXiv 2019

  19. [19]

    2010 The free-energy principle: A unified brain theory?Nat

    Friston K. 2010 The free-energy principle: A unified brain theory?Nat. Rev. Neurosci.11, 127–138. (doi:10.1038/nrn2787)

    doi:10.1038/nrn2787 2010

  20. [20]

    Parr T, Friston KJ. 2018. The graphical brain: belief propagation and active inference.Netw Neurosci2, 381–414. (doi:10.1162/netn a 00018)

    doi:10.1162/netn 2018

  21. [21]

    2018 The Markov blankets of life: Autonomy, active inference and the free energy principle.J

    Kirchhoff M, Parr T, Palacios E, Friston K, Kiverstein J. 2018 The Markov blankets of life: Autonomy, active inference and the free energy principle.J. R. Soc. Interface15, 20170792. (doi:10.1098/rsif.2017.0792)

    arXiv 2018

  22. [22]

    Friston KJ, FitzGerald T, Rigoli F, Schwartenbeck P, Pezzulo G. 2017. Active inference: A process theory.Neural Computation29, 1–49. (doi:10.1162/NECO a 00912)

    doi:10.1162/neco 2017

  23. [23]

    Ramstead MJD, Kirchhoff MD, Friston KJ. 2020. A tale of two densities: Active inference is enactive inference.Adaptive Behavior28, 225–239. (doi:10.1177/1059712319862774)

    work page doi:10.1177/1059712319862774 2020

  24. [24]

    Yufik YM. 2019. The understanding capacity and information dynamics in the human brain. Entropy21, 308. (doi:10.3390/e21030308)

    work page doi:10.3390/e21030308 2019

  25. [25]

    Ramstead MJD, Hesp C, Tschantz A, Smith R, Constant A, Friston KJ. 2021. Neural and phenotypic representation under the free-energy principle.Neuroscience and Biobehavioral Reviews120, 109–122. (doi:10.1016/j.neubiorev.2020.11.024)

    work page doi:10.1016/j.neubiorev.2020.11.024 2021

  26. [26]

    Sporns O, Betzel RF. 2016. Modular brain networks.Annual Review of Psychology67, 613–640. (doi:10.1146/annurev-psych-122414-033634)

    doi:10.1146/annurev-psych-122414-033634 2016

  27. [27]

    Hip´ olito I, Ramstead MJD, Convertino L, Bhat A, Friston K, Parr T. 2021. Markov blankets in the brain.Neurosci Biobehav Rev125, 88–97. (doi:10.1016/j.neubiorev.2021.02.003)

    work page doi:10.1016/j.neubiorev.2021.02.003 2021

  28. [28]

    Ramstead MJD, Seth AK, Hesp C, Sandved-Smith L, Mago J, Lifshitz M, et al. 2022. From generative models to generative passages: A computational approach to (neuro)phenomenology. Rev Philos Psychol13, 829–857. (doi:10.1007/s13164-021-00604-y)

    doi:10.1007/s13164-021-00604-y 2022

  29. [29]

    Palacios ER, Razi A, Parr T, Kirchhoff M, Friston K. 2020. On Markov blankets and hierarchical self-organisation.J Theor Biol486, 110089. (doi:10.1016/j.jtbi.2019.110089)

    arXiv 2020

  30. [30]

    Czajko S, Zorn J, Daumail L, Chetelat G, Margulies DS, Lutz A. 2024. Exploring the embodied mind: Functional connectome fingerprinting of meditation expertise.Biol Psychiatry Glob Open Sci4, 100372. (doi:10.1016/j.bpsgos.2024.100372)

    arXiv 2024

  31. [31]

    2013.The Predictive Mind

    Hohwy J. 2013.The Predictive Mind. Oxford, UK: Oxford University Press. (doi:10.1093/acprof:oso/9780199682737.001.0001)

    arXiv 2013

  32. [32]

    Varela FJ, Thompson E, Rosch E. 2017. The embodied mind: cognitive science and human experience. Cambridge, MA: MIT Press. (doi:10.7551/mitpress/9780262529365.001.0001) 15

    arXiv 2017

  33. [33]

    Kavi PC, Zamora-L´ opez G, Friedman DA, Patow G. 2025. Thoughtseeds: a hierarchical and agentic framework for investigating thought dynamics in meditative states. Entropy27, 459. (doi:10.3390/e27050459)

    work page doi:10.3390/e27050459 2025

  34. [34]

    2026 Dynamic Attentional Agents in Focused Attention Meditation: Hierarchical Computational Modeling of Expert-Novice Differences

    Kavi PC, Friedman DA, Patow G. 2026 Dynamic Attentional Agents in Focused Attention Meditation: Hierarchical Computational Modeling of Expert-Novice Differences. InActive Inference: 6th International Workshop, IWAI 2025, Montreal, QC, Canada, October 15–17, 2025, Revised Selected Papers, p. 182. Springer Nature

    2026

  35. [35]

    Bengio Y. 2017. The consciousness prior. arXiv. (doi:10.48550/arXiv.1709.08568)

    Pith/arXiv arXiv doi:10.48550/arxiv.1709.08568 2017

  36. [36]

    Miller GA. 1956. The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychol. Rev. 63, 81–97. (doi:10.1037/h0043158)

    doi:10.1037/h0043158 1956

  37. [37]

    Mashour GA, Roelfsema PR, Changeux J-P, Dehaene S. 2020. Conscious processing and the global neuronal workspace hypothesis. Neuron105, 776–798. (doi:10.1016/j.neuron.2020.01.026)

    doi:10.1016/j.neuron.2020.01.026 2020

  38. [38]

    2011 Experimental and theoretical approaches to conscious processing

    Dehaene S, Changeux J-P. 2011 Experimental and theoretical approaches to conscious processing. Neuron70, 200–227. (doi:10.1016/j.neuron.2011.03.018)

    work page doi:10.1016/j.neuron.2011.03.018 2011

  39. [39]

    Safron A. 2020 An integrated world modeling theory (IWMT) of consciousness: combining integrated information and global neuronal workspace theories with the free energy principle and active inference.Entropy22, 646. (doi:10.3390/e22060646)

    work page doi:10.3390/e22060646 2020

  40. [40]

    2011.Thinking, fast and slow

    Kahneman D. 2011.Thinking, fast and slow. New York, NY: Farrar, Straus and Giroux

    2011

  41. [41]

    2022.Active Inference: The Free Energy Principle in Mind, Brain, and Behavior

    Parr T, Pezzulo G, Friston KJ. 2022.Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. Cambridge, MA: MIT Press. (doi:10.7551/mitpress/12441.001.0001)

    doi:10.7551/mitpress/12441.001.0001 2022

  42. [42]

    Neurophysiol.106, 1125–1165

    Yeo BTT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead Met al.2011 The organization of the human cerebral cortex estimated by intrinsic functional connectivity.J. Neurophysiol.106, 1125–1165. (doi:10.1152/jn.00338.2011)

    arXiv 2011

  43. [43]

    2025 Deep computational neurophenomenology: a methodological framework for investigating the how of experience

    Sandved-Smith L, Bogot´ a JD, Hohwy J, Kiverstein J, Lutz A. 2025 Deep computational neurophenomenology: a methodological framework for investigating the how of experience. Neurosci. Conscious.2025, niaf016. (doi:10.1093/nc/niaf016)

    doi:10.1093/nc/niaf016 2025

  44. [44]

    A., Worhunsky, P

    Brewer, J. A., Worhunsky, P. D., Gray, J. R., Tang, Y.-Y., Weber, J. and Kober, H. 2011 Meditation experience is associated with differences in default mode network activity and connectivity.Proc. Natl Acad. Sci. USA108, 20 254–20 259. (doi:10.1073/pnas.1112029108)

    work page doi:10.1073/pnas.1112029108 2011

  45. [45]

    and Auletta, G

    Friston, K., Sengupta, B. and Auletta, G. 2014 Cognitive dynamics: from attractors to active inference.Proc. IEEE102, 427–445. (doi:10.1109/JPROC.2014.2306251)

    arXiv 2014

  46. [46]

    and Deco, G

    Gilson, M., Moreno-Bote, R., Ponce-Alvarez, A., Ritter, P. and Deco, G. 2016 Estimation of directed effective connectivity from fMRI functional connectivity hints at asymmetries of cortical connectome.PLoS Comput. Biol.12, e1004762. (doi:10.1371/journal.pcbi.1004762)

    work page doi:10.1371/journal.pcbi.1004762 2016

  47. [47]

    Kingma DP, Welling M. 2014. Auto-encoding variational Bayes.Proc. 2nd Int. Conf. Learn. Represent. (ICLR). (arXiv:1312.6114)

    Pith/arXiv arXiv 2014

  48. [48]

    2021 Dynamic expectation maximization algorithm for estimation of linear systems with colored noise.Entropy23, 1306

    Anil Meera A, Wisse M. 2021 Dynamic expectation maximization algorithm for estimation of linear systems with colored noise.Entropy23, 1306. (doi:10.3390/e23101306) 16

    doi:10.3390/e23101306 2021

  49. [49]

    Raffone A, Srinivasan N. 2017. Mindfulness and cognitive functions: Toward a unifying neurocognitive framework.Mindfulness8, 1–9. (doi:10.1007/s12671-016-0654-1)

    work page doi:10.1007/s12671-016-0654-1 2017

  50. [50]

    Vidaurre D, Masaracchia L, Larsen NY, Ruijters LRPT, Alonso S, Ahrends C, Woolrich MW

  51. [51]

    Beck J, Ramstead MJD. 2025. Dynamic Markov Blanket Detection for Macroscopic Physics Discovery.arXiv[Preprint] arXiv:2502.21217. (doi:10.48550/arXiv.2502.21217) 17 Supplementary Material This supplementary section provides detailed mathematical specification, parameter tables, and convergence diagnostics that align with the main manuscript. S1 Supplementa...

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.21217 2025

  52. [2025]

    (doi:10.1162/imag a 00460)

    The Gaussian-linear hidden Markov model: A Python package.Imaging Neuroscience3, imag a 00460. (doi:10.1162/imag a 00460)

    doi:10.1162/imag