Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation

summary

The gist

Meditative expertise involves sustained attention, rapid recovery from distraction, and coordinated dynamics of large-scale brain networks The model implements a three-layer nested Markov-blanket…

In short

The model uses a three-layer nested Markov-blanket architecture to simulate focused-attention meditation across four states: breath focus, mind-wandering, meta-awareness, and redirect attention. It applies the Free Energy Principle to show how biological systems minimize surprise through perception and action. Training distinguishes experts by modeling differences in brain network dynamics and attractor state transitions.

Key concepts

Free Energy Principle (FEP)
This theory suggests living organisms try to minimize 'surprise' or variational free energy. It frames the brain as a system trying to accurately predict sensory input and minimize discrepancies between what is expected and what is actually perceived, guiding learning through perception and action.
Markov Blanket Architecture
The model uses nested Markov blankets—transient computational structures—to describe cognitive functions. These blankets provide conditional independence across different scales, allowing the system to manage complex attentional dynamics by interacting with different levels of neural processing.
Dual-Process Architecture
The system has two main parts: System 1 (generative model) handles fast, automatic inference about potential thoughts, while System 2 (metacognitive monitor) acts as a slower executive gate. This allows the model to separate rapid thought generation from deliberate policy selection.
Attractor States
The four states—breath focus, mind-wandering, meta-awareness, and redirect attention—are stable points in the system's dynamics. The model shows how expert practitioners move efficiently between these states using learned transition rules to achieve focused attention.

Terminology used across episodes

This episode discusses

The paper

Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation · Read on arXiv

Universitat de Girona · Active Inference Institute · Universitat Pompeu Fabra

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "Recognizing Mind Wandering".

Marcus: Here's the gist of the paper. Meditative expertise involves sustained attention, rapid recovery from distraction, and coordinated dynamics of large-scale brain networks The model implements a three-layer nested Markov-blanket architecture to describe focused-attention meditation traversing four attractor states: breath focus, mind-wandering, meta-awareness, and redirect attention This computational phenomenology traverses four attractor states: breath focus, mind-wandering, meta-awareness, and redirect attention Theoretical Framework The overarching theoretical framework adopted is the Free Energy Principle (FEP), which posits that biological systems can be modeled as if they sustain their existence by minimizing variational free energy, an information-theoretic upper bound on sensory surprise Active Inference offers a process theory specifying how biological agents realize free energy minimization through recurrent cycles of perception, learning, and action The dynamics are situated within the Neuronal Packet Hypothesis (NPH), which posits that cognitive functions emerge from self-organizing ensembles of neurons termed “neuronal packets” These packets can be formally characterized as transient Markov blankets that confer a degree of computational autonomy and conditional independence across multiple spatiotemporal scales Dual-Process Architecture The model is structured into three nested layers: (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 Layer 1 evolves as the generative process and is not optimized by EM.

Ines: Training Signal Construction The training uses variational Expectation-Maximization (EM) across expert and novice phenotypes The M-step explicitly updates the parameters of the L2 generative model: the encoder ϕ, decoder θ, and forward model ψ.

Marcus: Concurrently, the M-step updates the learned habit prior pseudo-counts in L3 from buffered E-step state beliefs and policy posteriors The training minimizes a composite loss: Lt = Ft + Sfwd,t+1 + αrec Lrec,t.

Ines: Phenotype Signatures Expert phenotypes show stronger DMN suppression (0.332 vs 0.539) and higher DAN (0.673 vs 0.553) and FPN (0.701 vs 0.653) in Breath Focus compared to novices Experts sustain Breath Focus longer (87.8 vs 75.1 steps) and remain in Mind Wandering less (69.5 vs 96.0), matching findings that intensive training increases on-task time while reducing mind-wandering Transition dynamics reveal that, conditioned on switching, experts show a clean recovery loop (MW→MA, MA→RA, and RA→BF all at 1.00), mirroring the canonical sequence Attractor Dynamics In the low-dimensional thoughtseed space (L2), PC1 and PC2 explain 94.1% and 88.8% of the variance for novices, and 82.1% and 9.4% for experts Expert trajectories are more compact in network space (PC1/PC2 std ≈ 0.239/0.089) than those of novices (≈ 0.286/0.090), indicating tighter attractor dynamics at L1 for experienced practitioners Limitations and Future Directions The present model seeks to operationalize the boundary between automatic and deliberate processing in a manner consistent with Helmholtzian predictive-coding frameworks The current state-transition mechanism performs single-step policy selection via a softmax over Expected Free Energy together with a learned habit prior and an explicit dwell-aware temporal heuristic Future iterations of the model could introduce phenotype-specific baseline parameters for L3 meta-awareness to further distinguish expert from novice metacognitive capacities The framework may inform models of attention, mind-wandering, and meta-awareness across cognitive domains Acknowledgements We thank the organizers and participants of IWAI 2025 for valuable feedback that helped refine this work from an initial rules-based approach to a tractable enactive inference implementation This research was partially funded by grant PID2021-122136OB-C22 from MICIU/AEI/10.13039/501100011033 and by ERDF “A way of making Europe”, and by the AGAUR research support grant 2021 SGR 01035 from the Department of Research and Universities of the Generalitat de Catalunya (G.P.) The authors declare no conflicts of interest.

Title and authors: Ines: So we're looking at this paper now titled "Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation". It really tries to model meditation as a process with four distinct states.

Marcus: Right. It’s looking at how sustained attention, recovery from distraction, and the coordination of large brain networks actually work when someone is meditating. The authors are using this dual-process active inference framework to do it.

Yuki: That sounds like they're trying to map the cognitive experience onto a system that minimizes surprise in real time. I wonder what they think about how much of that process is automatic versus deliberate?

Ines: They set up this three-layer nested Markov blanket architecture. Layer one is a high-dimensional physiological substrate modeled as a stochastic multivariate Ornstein–Uhlenbeck process over attentional Yeo networks.

Marcus: And then layer two, which they call System one is the low-dimensional generative model that encodes latent mental content as thoughtseeds and evaluates autonomic action tendencies <ref:2607.14833#pg1,that encodes latent mental content as thoughtseeds and evaluates autonomic action tendencies>.

Yuki: So System one is essentially the fast, automatic part of the brain making predictions about what’s going on in our heads before we even consciously notice it. What do you think about how that fits into our understanding of species-wide cognitive patterns?

Ines: That's where population genetics comes in. They are using this framework to describe the "discursive mind," which is basically the stream of thoughts and sensory impressions unfolding over time eight <ref:2607.14833#pg2>. They map this onto a canonical cycle with four stages.

Marcus: So they break down the experience into breath focus, mind-wandering, meta-awareness, and then redirecting attention. It’s like a set of observable states that we can track computationally.

Yuki: And how does this relate to what we know about meditation across different cultures or species? Are these four states universal in some sense?

Ines: The paper suggests the cycle reflects competitive and cooperative dynamics among large-scale brain networks ten <ref:2607.14833#pg2,cycle reflects competitive and cooperative dynamics among large-scale brain networks 10>. It aligns with predictive processing theories where cognition is framed as continuously minimizing prediction errors across different scales eleven seven <ref:2607.14833#pg2>.

Marcus: The key part for us is how they connect this to the underlying biology. They situate these dynamics within the Neuronal Packet Hypothesis, which says cognitive functions emerge from self-organizing ensembles of neurons called neuronal packets twenty-four twenty-five <ref:2607.14833#pg3,cognitive functions emerge from self-organizing ensembles of neurons>.

Yuki: Wait, so these packets are transient Markov blankets that give them a degree of computational autonomy across different scales? That’s a big idea for how we view the brain.

Title and authors: Ines: Exactly. And they say these packets can self-organize into higher-order, superordinate ensembles to embody complex cognitive processes twenty-five <ref:2607.14833#pg3,can self-organize into higher-order, superordinate ensembles to embody complex cognitive>. This is supported by the sparse connectivity architecture of the brain, which allows for localized functional units twenty-six <ref:2607.14833#pg3>.

Marcus: And this leads to that hierarchical, nested Markov blanket structure they use in their model. It’s a way of modeling how those smaller packets organize into bigger structures.

Yuki: So if we think about the wider history of species, could these nesting dynamics help us understand why certain attentional patterns become ingrained?

Ines: It helps show how that organization happens dynamically during practice, not just as a fixed structure. The paper focuses on how this system evolves through focused-attention meditation.

Marcus: And they talk about the training signal construction using variational expectation-maximization across expert and novice phenotypes. They are explicitly trying to train the AI model on what makes someone an expert versus a beginner.

Yuki: That’s interesting because it implies that expertise isn't just about practicing more; it’s about how that practice reshapes the underlying computational structure of attention itself.

Ines: They show that expert phenotypes show stronger suppression of the default mode network, DMN, and higher activity in the dorsal attention network, DAN, and frontal parietal network, FPN in breath focus compared to novices

Expert phenotypes show stronger DMN suppression (zero point three three two vs zero point five three nine) and higher DAN (zero point six seven three vs zero point five five three) and FPN (zero point seven zero one vs zero point six five three) in Breath Focus: .

Marcus: And they quantify the duration of these states too, showing experts sustain breath focus longer, like eighty-seven point eight steps compared to seventy-five point one steps for novices

Experts sustain Breath Focus longer (eighty-seven point eight vs seventy-five point one steps): . That’s a measurable difference in performance.

Yuki: So if we look at this from a population genetic viewpoint, does this suggest that the cognitive machinery itself is being selected for through meditation practice?

Ines: It suggests that the dynamics of these states are being tuned by the training process. And they show transition dynamics where conditioned on switching, experts have a clean recovery loop: mind-wandering to meta-awareness, then meta-awareness to redirect attention, and finally redirecting back to breath focus all at one point

Transition dynamics reveal that, conditioned on switching, experts show a clean recovery loop (MW→MA, MA→RA, and RA→BF all at one point zero zero): <ref:2607.14833#pg1>.

Marcus: That clean loop is telling. It means the system recovers from distraction much more efficiently than someone who just wanders around randomly. It mirrors the canonical sequence of how we think attention should flow when we try to focus.

Title and authors: Yuki: For someone who only listens to this, what does that clean recovery loop mean for their daily life? Is it just about being able to get back on task faster?

Ines: It means the system is learning a more efficient way to manage its internal resources during attention shifts. They are learning the constraints and preferred outcomes twenty-two <ref:2607.14833#pg3,constraints and preferred outcomes 22>.

Marcus: And they do show that in the low-dimensional thoughtseed space, PC one and PC two explain ninety-four point one percent and eighty-eight point eight percent of the variance for novices, but only eighty-two point one percent and nine point four percent for experts

In the low-dimensional thoughtseed space (L2), PC1 and PC2 explain ninety-four point one percent and eighty-eight point eight percent of the variance for novices, and eighty-two point one percent and nine point four percent for experts: .

Yuki: So, even when we simplify the mental content down to just two principal dimensions, we see a clear separation between what novices experience versus what experts experience in terms of thought patterns.

Ines: That compactness in network space is also telling. Expert trajectories are more compact in network space with standard deviations around zero point two three nine for PC one and zero point zero eight nine for PC two compared to novice trajectories, which are around zero point two eight six for PC one and zero point zero nine for PC two

Expert trajectories are more compact in network space (PC1/PC2 std ≈ zero point two three nine/zero point zero eight nine) than those of novices (≈ zero point two eight six/zero point zero nine zero): .

Marcus: That compactness means the expert system is operating in a much tighter, more constrained space within its own network dynamics at layer one indicating tighter attractor dynamics at L1 for experienced practitioners. It’s less noisy.

Yuki: So, if we think about what this does for the wider field of attention research, does it offer a way to operationalize that boundary between automatic and deliberate processing?

Ines: The authors are trying to do just that. They want to operationalize the boundary between automatic and deliberate processing in a way that matches Helmholtzian predictive-coding frameworks The present model seeks to operationalize the boundary between automatic and deliberate processing in a manner consistent with Helmholtzian predictive-coding frameworks.

Marcus: And their current state transition mechanism performs single-step policy selection via a softmax over expected free energy together with a learned habit prior and an explicit dwell aware temporal heuristic.

Yuki: That means the AI system is choosing its next move based on what it expects to minimize surprise right now, informed by past habits and how long it’s been in the current state.

Title and authors: Ines: Future work could introduce phenotype specific baseline parameters for layer three meta-awareness to further distinguish expert from novice metacognitive capacities Future iterations of the model could introduce phenotype specific baseline parameters for L3 meta-awareness to further distinguish expert from novice metacognitive capacities.

Marcus: And that framework might inform models of attention, mind wandering, and meta-awareness across different cognitive domains. It’s an attempt to build a unified computational phenomenology.

Yuki: So for the listener who just wants to know what this means outside of the model itself, it suggests that meditation isn't just about relaxation; it’s about training a specific kind of dynamic system in your brain.

Ines: Precisely. It’s about optimizing the way your attention networks organize themselves under pressure, and how that optimization becomes a stable state for an expert practitioner

Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation: .

Marcus: So we've looked at how they build this model, how they train it on expert versus novice data, and what the resulting differences are in terms of network compactness and recovery loops.

Yuki: It really suggests that what we see as a skill is actually a change in the system’s architecture itself rather than just a set of learned habits.

Ines: Exactly. It’s about the underlying dynamical constraints being adjusted by focused attention practice over time

Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation: .

Marcus: We've covered the title and authors, then their summary of the paper, and looked at what they suggest about improving the model's architecture.

Yuki: And we’ve gone through some of their conclusions about expert phenotypes showing stronger DMN suppression and tighter attractor dynamics in those low-dimensional spaces.

Ines: We’re wrapping up here with a look at what this computational phenomenology of focused-attention meditation means for how we understand the mind itself.

Marcus: It suggests that expertise is really about tuning the system’s capacity to minimize prediction errors across different scales

Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation: .

Yuki: We've explored how this model helps us see the four attractor states and what they look like computationally.

Ines: It’s about understanding how those states interact in a coordinated dynamic, which is pretty important for any cognitive science trying to map the system.

Marcus: So we've looked at how they train this dual-process active inference formulation and the resulting phenotype signatures.

Yuki: That’s where we stop for this paper for now, leaving us with the idea that meditation might be a way to actively shape our own cognitive architecture.

The paper's summary: Ines: So, this paper is basically trying to take meditation—specifically focused-attention meditation—and turn it into a computer model. It’s looking at how that sustained attention works by breaking it down into four different mental states.

Marcus: Right. They’re using this three-layer structure called a nested Markov blanket architecture to describe the whole process, from breath focus all the way to mind wandering and back again.

Yuki: So they are trying to build a computational map of the experience itself, rather than just looking at brain scans in isolation?

Ines: Exactly. They’re using this framework rooted in the Free Energy Principle, which is basically an idea that biological systems exist by trying to minimize surprise or variational free energy.

Marcus: And they frame meditation as an active inference process. So instead of just passively observing a state, the brain is actively learning and adjusting its internal models through cycles of perception and action.

Yuki: That connects it to how we think about learning across species, right? Like how those predictive models get refined over generations.

Ines: Right. And they’re looking at this through the lens of the Neuronal Packet Hypothesis, which suggests that cognitive functions aren't just in single neurons but emerge from these self-organizing groups called neuronal packets.

Marcus: These packets are transient Markov blankets, meaning they have a kind of independent computational power across different scales and times.

Yuki: That’s interesting because it gives us a way to think about how complex behaviors, like sustained attention, aren't just fixed wiring but something that organizes itself in real time.

Ines: And they use this structure to see how those packets transition between the four states they identified: breath focus, mind wandering, meta-awareness, and redirecting attention.

Marcus: The training method is pretty interesting too. They use a variational expectation maximization process across expert and novice people to tune the model parameters.

Yuki: So they are essentially training an AI to recognize the pattern of an expert practitioner by seeing how their brain dynamics shift during meditation practice?

Ines: Precisely. They compare how experts handle things like default mode network activity versus dorsal attention network engagement when they are focusing on their breath.

Marcus: The results show that experts suppress the default mode network way more than novices do when they are in breath focus, and they stay in that focused state much longer—we’re talking eight point seven eight steps versus seventy-five point one for novices.

Yuki: That difference in "dwell time" is significant. It suggests that the training actually changes the underlying attentional machinery over time.

Ines: And they show a clean recovery loop for experts, meaning when they switch states, like from mind wandering to meta-awareness, then back to breath focus, it happens all at one point with perfect transition dynamics.

Marcus: That clean loop is what tells us that the expert system has learned a much more efficient way to manage its internal resources during those transitions.

Yuki: It moves this beyond just saying "meditation helps you relax." It’s suggesting meditation might be actively tuning the fundamental structure of your attention system itself.

Ines: That’s the core idea. They are operationalizing that line between automatic thinking and deliberate control using this model to see how it scales from individual practice to whole-system organization.

The paper's improvements: Tom: So, we’re looking at how they suggest making this AI model better for real use. They aren't just stopping at their initial version; they’re thinking about how to make the expert distinction even clearer in the future.

Ines: They point out that right now, Layer three, which is the agentic metacognitive monitor, doesn't have specific baseline parameters for meta-awareness yet.

Marcus: That means if we want to truly distinguish an expert from a novice in terms of their thinking process beyond just the state durations they see now, we need to add those specific tuning knobs.

Yuki: So they are suggesting phenotype-specific parameters for that Layer three monitor, something that would really allow the AI to see what kind of metacognitive capacity you have.

Ines: Right. They want to see if we can tune those parameters so the system can differentiate between expert and novice metacognitive capacities more clearly than it does currently.

Marcus: From a data perspective, that means we need more cohort data that specifically captures the dynamics of that meta-awareness layer, not just the surface-level states.

Yuki: And from a population genetics view, if we can tune those parameters, it might tell us if certain attentional patterns are being selected for across different human populations or developmental stages.

Ines: Exactly. The implication is that the current model gives us a good start, but to really map out the full spectrum of expert behavior, we need to refine how we model that higher-order decision making.

Marcus: So they’re moving from describing what happens now to building a more precise tool for identifying *how* expertise is computationally encoded.

Yuki: It opens up the possibility of using this framework not just to study meditation, but as a way to model how any complex system learns its own optimal operating rules through practice.

Ines: That’s the big picture. They’re aiming for a framework that can inform models of attention and mind wandering across many different cognitive domains, not just sitting still in silence.

Marcus: It moves it from being a description of meditation to being a general methodology for understanding how active inference works in any biological system.

Yuki: So the next step is taking those refined parameters and seeing if we can find similar organizational principles in other areas of human cognition or even other species.

Conclusion: Tom: So we're wrapping up our look at "Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation." Basically, this paper uses an active inference approach to map out the four states of meditation—breath focus, mind wandering, meta-awareness, and redirecting attention.

Ines: It’s about showing how these mental states are not just random moments but parts of a coordinated system that minimizes surprise over time.

Marcus: And they use this three-layer model to break down the brain into physical substrate, a low-dimensional generative model for thoughtseeds, and an agentic monitor that acts as the executive gate.

Yuki: So it suggests that meditation is training the system's ability to manage its internal predictive models more effectively.

Ines: Right. The key finding is how expert trajectories are much tighter in network space when you look at those low-dimensional thoughtseed spaces, with standard deviations around zero point two three nine for the first two principal components.

Marcus: That compactness means the expert system is operating in a much more constrained space within its own network dynamics at layer one, which is what they call tighter attractor dynamics.

Yuki: From a population genetic viewpoint, if we can see this tightness in the data, it might hint at how cognitive structures are being selected for across different human groups.

Ines: It suggests that the training process actively sculpts the underlying way attention networks organize themselves, which is something we need to test with more varied cohorts.

Marcus: And they do flag that their current method only performs single-step policy selection using a softmax over expected free energy combined with a learned habit prior and an explicit dwell aware temporal heuristic.

Yuki: So the caveat is that it doesn't fully capture the slow, deep organizational changes in the brain; it’s more of a fast transition model for now.

Ines: Exactly. But even with those limitations, this framework gives us a way to operationalize that boundary between automatic and deliberate processing in a way that aligns with predictive coding theories.

Marcus: It changes things for us by giving us a mathematical structure to test hypotheses about how we gain skills through repeated experience, like meditation.

Yuki: It opens the door to looking at how this kind of self-organization might apply to other systems, not just the human brain.

Ines: We’ve talked about how they build this model and what it recovers in terms of expert signatures for breath focus and recovery loops.

Marcus: And we discussed how those tight attractor dynamics imply a more constrained, efficient system for the expert practitioner.

Yuki: It really shows that skill isn't just adding more practice to the equation; it’s about reorganizing the underlying computational structure itself.

Ines: So "Recognizing Mind Wandering: A Computational Phenomenology of Focused-Attention Meditation" gives us a clear, if complex, picture of how focused attention works dynamically.

Marcus: It’s a solid start for using AI to model cognitive skills that are currently hard to measure directly.

More episodes

← Home