Meta-awareness, the ability to notice the current content of consciousness, is crucial for controlling cognitive states like directing attention. This paper models meta-awareness and attentional control using hierarchical active inference, treating mental actions as policy choices over higher-level cognitive states. A further hierarchical level represents meta-awareness states that modulate the expected confidence in the mapping between observations and hidden cognitive states. Simulations of mind-wandering during a sustained selective attention task illustrate how this inferential architecture enables accessing and controlling cognitive states, offering a computational foundation for a phenomenology of mental action and self-monitoring.
This theoretical paper proposes a computational model of mental action—the deliberate control of one's own thoughts and attention—by combining insights from phenomenology and active inference. The authors develop a deep-parametric active inference framework to simulate meta-awareness and attentional control, showing how agents can learn to monitor and regulate their own cognitive processes. The model suggests that meta-awareness emerges from hierarchical inference about attentional states, enabling flexible control of attention. This work bridges phenomenological philosophy and computational neuroscience, offering a formal account of how conscious agents can intentionally shape their own mental activity.
A predictive processing framework, grounded in active inference and free-energy minimization, can explain how focused attention meditation works. Paying voluntary attention to the body during meditation downweights habitual automatic reactions and distracting spontaneous thoughts, thereby settling the mind. The framework also links phenomenological concepts like opacity and de-reification to the voluntary allocation of attention and precision-weighting. The authors propose this theoretical approach as a promising strategy for contemplative research, though explicit computational simulations and comparisons with experimental data are still needed.