Thoughtseeds Model Uses Active Inference to Simulate Meditation Brain States

A new paper on arXiv introduces "Thoughtseeds as Latent Causes," a computational phenomenology of focused-attention meditation built on active inference and dual-process theory. The model, authored by Prakash Chandra Kavi, Daniel Ari Friedman, and Gustavo Patow, uses a three-layer nested Markov-blanket architecture. Layer 1 (L1) models the high-dimensional physiological neuronal substrate as a stochastic Ornstein-Uhlenbeck process over attentional Yeo networks. Layer 2 (L2) implements a low-dimensional generative model (System 1) that encodes latent mental content as "thoughtseeds" and evaluates autonomic action tendencies. Layer 3 (L3) acts as an agentic metacognitive monitor (System 2), implementing a Global Neuronal Workspace (GNW) capacity bottleneck to selectively gate these tendencies.

Simulations traverse four attractor states central to meditation: breath focus, mind-wandering, meta-awareness, and redirect attention. Meta-awareness functions as the GNW ignition signal, derived from policy-prior divergence and dynamically gated by competition between orchestrator and distractor thoughtseeds. Training uses variational Expectation-Maximization across expert and novice phenotypes, successfully reproducing empirical findings from contemplative neuroscience. The work provides a tractable link between subjective meditation experience and objective neurophysiological measures, with potential applications in brain-computer interfaces, digital therapeutics for attention disorders, and AI systems that emulate metacognitive control.