Thoughtseeds: A Hierarchical and Agentic Framework for Investigating Thought Dynamics in Meditative States

Abstract

The Thoughtseeds Framework introduces a novel computational approach to modeling thought dynamics in meditative states, conceptualizing thoughtseeds as dynamic attentional agents that integrate information. This hierarchical model, structured as nested Markov blankets, comprises three interconnected levels: (i) knowledge domains as information repositories, (ii) the Thoughtseed Network where thoughtseeds compete, and (iii) meta-cognition regulating awareness. It simulates focused-attention Vipassana meditation via rule-based training informed by empirical neuroscience research on attentional stability and neural dynamics. Four states-breath_control, mind_wandering, meta_awareness, and redirect_breath-emerge organically from thoughtseed interactions, demonstrating self-organizing dynamics. Results indicate that experts sustain control dominance to reinforce focused attention, while novices exhibit frequent, prolonged mind_wandering episodes, reflecting beginner instability. Integrating Global Workspace Theory and the Intrinsic Ignition Framework, the model elucidates how thoughtseeds shape a unitary meditative experience through meta-awareness, balancing epistemic and pragmatic affordances via active inference. Synthesizing computational modeling with phenomenological insights, it provides an embodied perspective on cognitive state emergence and transitions, offering testable predictions about meditation skill development. The framework yields insights into attention regulation, meta-cognitive awareness, and meditation state emergence, establishing a versatile foundation for future research into diverse meditation practices (e.g., Open Monitoring, Non-Dual Awareness), cognitive development across the lifespan, and clinical applications in mindfulness-based interventions for attention disorders, advancing our understanding of the nature of mind and thought.

Keywords: Markov blanket; Vipassana; active inference; content of consciousness; embodied cognition; global workspace; meditation; meta-cognition; thoughtseed.

Figure 1

Hierarchical Thoughtseed Framework within a sentient being. (A) Partitioning of an agent’s internal states and external states through the Markov blanket (Ref [34]). The Markov blanket, comprising sensory (s) and active states (a), mediates the interaction between internal states internal states (μ), and external states (η). Internal states can only influence external states through active states, while external states can only influence internal states through sensory states. This separation allows the internal states, which house the Thoughtseed Network (TN), to operate with a degree of autonomy, generating predictions and selecting actions based on its internal model of the world. (B) The nested hierarchical organization of attentional processes within the brain’s internal states, as detailed in Figure 1A. It illustrates three levels—knowledge domains (KDs), the Thoughtseed Network (TN), and meta-cognition—each enclosed in its own Markov blanket, forming a hierarchical structure rooted in active inference and Global Workspace Theory (GWT). Knowledge domains (KDs): The base level comprises self-organizing units of embodied knowledge. Each KD has a Markov blanket that interfaces with sensory inputs and actions, providing a neuronal foundation for both conscious and unconscious processing. Thoughtseed Network (TN): The TN represents interactions among thoughtseeds—attentional agents—competing for dominance in the Global Workspace via winner-takes-all dynamics, shaping the content of conscious experience. Meta-cognition: The meta-cognitive agent oversees the Thoughtseed Network via attentional precision, and meta-awareness monitors a detection system so that the behavior is aligned with intentionality, guided by global goals and policies. Bidirectional information flow: Blue arrows indicate bottom-up processing (e.g., prediction errors), whereas red arrows indicate top-down processing (e.g., predictions). This bidirectional flow reflects the dynamic interaction across cognitive levels, where each layer functions as a Markov blanket, facilitating selective information exchange. Markov Blanket interactions: At the system’s boundary, the Markov blanket interfaces with the external states, labeled “Umwelt + Environment,” through sensory states and active states. This interaction aligns with embodied cognition’s emphasis on the sentient being’s engagement with its surroundings, shaping cognitive processes through sensory inputs and active outputs.

Figure 2

Neuronal packets and knowledge domains. (A) States of a neuronal packet (NP), showing how it transitions from an unmanifested to a manifested state, forming a Markov blanket that encapsulates its knowledge structure. In its activated state, the NP processes sensory inputs and generates actions. (B) Multiple NPs organize into knowledge domains (KDs), integrating sensory data and actions into a hierarchical and heterarchical framework (Level 1 in Thoughtseed Framework). The responses from NPs’ activated states feed into KDs, enabling thoughtseed dynamics and competition in the Global Workspace, ultimately shaping conscious experience. (A) This figure illustrates the states of a neuronal packet (NP) using a free energy landscape, with the x-axis representing internal states and the y-axis indicating free energy. The NP exists in three states: Unmanifested State: A potential neural configuration shaped by evolutionary priors, shown as a shallow local minimum (blue curve). Manifested State: Forms after repeated stimulus exposure, leading to a phase transition and the formation of a stable Markov Blanket—with its encapsulated knowledge structure. It includes a core attractor (red dot) as the primary neural pattern and subordinate attractors (red ‘x’) as secondary patterns. The vertical dashed line marks the energy barrier, and binding energy is the distance from the core attractor to the zero-free energy level (horizontal dashed line). Activated (or Spiking) State: A transient state characterized by heightened neural activity within the manifested NP ensemble, triggered by the dominant thoughtseed and generating a response influencing behavior or cognition (green curve). (B) KDs are shown as colored squares, each representing specialized knowledge areas: Unfilled squares indicate localized KDs (superordinate ensembles of NPs, or neuronal packet domains, NPDs). A filled square represents a higher-order, heterarchical KD integrating knowledge across domains. Organization: The arrangement of NPs within or connected to KDs visually represents the hierarchical nature of knowledge representation. The higher-order KD suggests a heterarchical organization that allows for a more complex integration of knowledge across domains. Dynamic Interplay: Connections between NPs and KDs represent the flow of information and influence. NPs provide raw data, while KDs interpret and contextualize this information, contributing to thoughtseed emergence from the dynamic interplay of information processing within and between KDs.

Figure 2

Neuronal packets and knowledge domains. (A) States of a neuronal packet (NP), showing how it transitions from an unmanifested to a manifested state, forming a Markov blanket that encapsulates its knowledge structure. In its activated state, the NP processes sensory inputs and generates actions. (B) Multiple NPs organize into knowledge domains (KDs), integrating sensory data and actions into a hierarchical and heterarchical framework (Level 1 in Thoughtseed Framework). The responses from NPs’ activated states feed into KDs, enabling thoughtseed dynamics and competition in the Global Workspace, ultimately shaping conscious experience. (A) This figure illustrates the states of a neuronal packet (NP) using a free energy landscape, with the x-axis representing internal states and the y-axis indicating free energy. The NP exists in three states: Unmanifested State: A potential neural configuration shaped by evolutionary priors, shown as a shallow local minimum (blue curve). Manifested State: Forms after repeated stimulus exposure, leading to a phase transition and the formation of a stable Markov Blanket—with its encapsulated knowledge structure. It includes a core attractor (red dot) as the primary neural pattern and subordinate attractors (red ‘x’) as secondary patterns. The vertical dashed line marks the energy barrier, and binding energy is the distance from the core attractor to the zero-free energy level (horizontal dashed line). Activated (or Spiking) State: A transient state characterized by heightened neural activity within the manifested NP ensemble, triggered by the dominant thoughtseed and generating a response influencing behavior or cognition (green curve). (B) KDs are shown as colored squares, each representing specialized knowledge areas: Unfilled squares indicate localized KDs (superordinate ensembles of NPs, or neuronal packet domains, NPDs). A filled square represents a higher-order, heterarchical KD integrating knowledge across domains. Organization: The arrangement of NPs within or connected to KDs visually represents the hierarchical nature of knowledge representation. The higher-order KD suggests a heterarchical organization that allows for a more complex integration of knowledge across domains. Dynamic Interplay: Connections between NPs and KDs represent the flow of information and influence. NPs provide raw data, while KDs interpret and contextualize this information, contributing to thoughtseed emergence from the dynamic interplay of information processing within and between KDs.

Figure 3

State transition matrices derived from empirical research [37,38]. For experts (left panel), in the ‘breath_control’ state, practitioners maintain focus 55% of the time and engage in meta-awareness 25% of the time, reflecting adaptive attentional dynamics [36,78]. In the ‘mind_wandering’ state, experts exhibit self-regulation, spending 50% of the time in this state with a 25% probability of transitioning to ‘meta_awareness’, indicating timely recognition of competing thoughts. The ‘meta_awareness’ state transitions to ‘redirect_breath’ with an 80% probability, which then returns to ‘breath_control’ 90% of the time [38]. For novices (right panel), ‘breath_control’ shows reduced stability with focus maintained 50% of the time and a 35% transition to ‘mind_wandering’. In the ‘mind_wandering’ state, novices remain 70% of the time, while ‘meta_awareness’ transitions to ‘redirect_breath’ 65% of the time. The ‘redirect_breath’ state returns to ‘breath_control’ 70% of the time but lapses to ‘mind_wandering’ 20%, indicating weaker attentional control [79,80]. These matrices, derived from focused attention meditation research [78,79,80], serve as a reference for understanding the 4 state cyclical patterns [37,38] in Vipassana meditation practice.

Figure 4

Learned weight matrices after learning. The matrices (left: expert, right: novice) reveal experience-dependent differences: Attentional Focus Enhancement: Experts display stronger breath_focus activation during breath_control (0.98 vs. 0.78 in novices), reflecting enhanced executive control [79]. Distraction Reduction: Novices show higher distraction thoughtseed activation in mind_wandering (e.g., pain_discomfort: 0.55, pending_tasks: 0.58) compared to experts (0.30, 0.14), consistent with reduced default mode network activity in experienced meditators [78]. Equanimity Development: Experts exhibit stronger equanimity weights, especially during redirect_breath (0.75 vs. 0.28 in novices), supporting improved emotional regulation [81]. Meta-cognitive Awareness and Redirect Breath: Both groups show similar self_reflection activation during meta_awareness (experts: 0.45, novices: 0.46), but experts show greater redirecting attention to breath capabilities (experts: 0.75, novices: 0.28). Neural Pattern Consistency: Experts’ tighter weight clustering in adaptive states (breath_control, redirect_breath) indicates consistent neural recruitment, reflecting neuroplasticity from long-term practice [82,83].

Figure 5

Thoughtseed activations and meta-awareness during learning. This composite plot shows the activation trajectories of five thoughtseeds (‘self-reflection’, ‘breath_focus’, ‘equanimity’, ‘pain_discomfort’, ‘pending_tasks’) along with meta-awareness for novice meditators (top panel) and expert meditators (bottom panel) during the learning process.

Figure 6

Mediation states’ evolution during learning. These plots illustrate the temporal progression of meditation states—breath_control, mind_wandering, meta-awareness, and redirect_breath—over 200 timesteps for novice (top panel) and expert (bottom panel) meditators. State transitions, constrained by dwell time limits (mean ± 2 SD) and modulated by meta-awareness (0.55–0.9), demonstrate a 4-state cyclical model of focused attention in Vipassana meditation [36,38].

Figure 7

Thoughtseed interaction network. These thoughtseed interactions differ between meditation experience levels (novice in right panel and expert in left panel), where red represents inhibitory connections and green represents facilitatory connections. The data-driven analysis reveals several key differences: Enhanced Meta-Cognitive Processing: Experts show stronger facilitatory connections from breath focus to self-reflection (+0.70 vs. novices’ +0.00), indicating that sustained attention to breath becomes integrated with meta-cognitive awareness through practice [36]. Refined Distraction Management: Novices show mutual reinforcement between distraction types (pain_discomfort → pending_tasks: +0.70), while experts demonstrate no such reinforcement (+0.00), indicating better separation between different distraction categories. Improved Breath-related Regulation: Experts develop more targeted inhibitory control (breath_focus → pending_tasks: −0.69) while simultaneously reducing inhibition toward sensations (breath_focus → pain_discomfort: +0.00 vs. novices’ −0.70). Equanimity Cultivation: Experts show a positive connection from breath_focus to equanimity (+0.25) that is absent in novices, supporting theoretical accounts of breath awareness fostering equanimity with sustained practice [80]. Reduced Negative Interference: Both self_reflection and equanimity show less inhibitory relationships with pending_tasks in experts, suggesting more balanced integration of attention networks rather than oppositional relationships.

Figure 8

Hierarchical organization of meditation dynamics. This figure illustrates a three-level organization of meditation dynamics for novice (top panel) and expert (bottom panel) meditators: It demonstrates how meditation emerges from three interconnected levels: Level 1—Thoughtseed Activations: Competitive dynamics between five color-coded thoughtseeds (e.g., breath_focus, self_reflection), with continuous activation trajectories showing the evolution of attention, distraction, and meta-cognitive processes over time. Level 2—Dominant Thoughtseed (middle): Discrete colored dots visualizing the winner-takes-all competition, indicating which thoughtseed has the highest activation at each moment. Level 3—Meta-Awareness (top): A purple line depicting meta-awareness fluctuations in response to changes in dominant thoughtseeds and meditation states. Novice vs. Expert Differences: Activation Stability: Experts exhibit more consistent thoughtseed activations, particularly for breath_focus and equanimity, with less noise than novices. Dominance Patterns: Novices display rapid switches in dominant thoughtseeds, reflecting frequent distractions, while experts sustain longer periods of breath_focus dominance with fewer interruptions. Meta-Awareness Fluctuations: Novices show more pronounced meta-awareness drops during mind-wandering, whereas experts maintain a higher baseline, even amidst distractions [39].