Qurated: Thoughtseeds as Latent Causes: A Dual-Process Computational Phenomenology of Focused-Attention Meditation
The Mind Doesn't Wander Randomly — It Falls Into Waiting Attractors
Every meditator knows the moment: breath awareness dissolves, and suddenly you're planning dinner. This isn't chaos. It's physics. A new computational model shows that mind-wandering, meta-awareness, and redirected attention aren't separate skills you learn — they're attractor states your brain was always going to fall into, and expertise is just learning which doors to leave open.
The Core Insight: Thoughts Are Latent Causes, Not Events
The paper's central move is deceptively simple: it treats "thoughtseeds" — units of mental content like breath-focus or distraction — as latent causes competing to explain your ongoing experience, the same way your brain infers "there's a cup" from raw visual noise. Meditation isn't fighting distraction. It's a competition between generative hypotheses about what your mind should be doing right now.
This reframes the entire practice. You're not suppressing the wandering mind — you're changing which hypothesis wins.
The Architecture: Three Nested Layers, Two Speeds of Mind
The model stacks three Markov blankets, each a boundary between inference and the world it infers:
- L1 — The Substrate: Raw neuronal dynamics across attention networks, modeled as noisy oscillators (an Ornstein-Uhlenbeck process) — the wet physics beneath thought.
- L2 — System 1: A fast, low-dimensional model encoding thoughtseeds and their action tendencies. This is the autopilot — quick, cheap, usually right.
- L3 — System 2: A slow metacognitive monitor implementing a Global Neuronal Workspace — a capacity-limited bottleneck that ignites only when something demands conscious arbitration.
The key mechanic: meta-awareness is not constant vigilance — it's an ignition event, triggered when the divergence between competing thoughtseeds (orchestrator vs. distractor) crosses a threshold. You don't watch your mind non-stop. You get notified when arbitration is needed.
The Framework: Expected Free Energy as the Selection Rule
Policy selection — what your mind does next — minimizes expected free energy: a joint bet on prediction accuracy and information gain. This is the mental model worth stealing:
Attention is not effort. Attention is inference under a budget.
Every redirect of focus is your brain choosing the policy that best resolves uncertainty and fulfills your goals — not a moral victory over weakness.
Why Experts Differ From Novices
Training the model (via variational EM) across expert and novice phenotypes reveals the real signature of expertise: not fewer distractions, but faster ignition and cheaper redirection. Experts don't have quieter minds — they have lower-latency arbitration circuits. The GNW bottleneck fires sooner, the orchestrator thoughtseed wins faster, and descending predictions correct the network dynamics before the wandering entrenches.
Practical Takeaways
- Stop treating distraction as failure. It's an attractor state your architecture is built to visit — the goal is faster exit, not permanent avoidance.
- Meta-awareness is a signal, not a state. You can train the ignition threshold — noticing sooner — more than you can train constant vigilance.
- Redirecting attention is prediction, not punishment. Each return to breath is your system updating its model of what matters — treat it as data, not discipline.
- Expertise = latency reduction. The felt sense of "calm mind" in experienced meditators is really a faster feedback loop between noticing and correcting.