Neuroscience of consciousness
January 1, 2021
Lars Sandved-Smith, Casper Hesp, Jérémie Mattout et al.
116 citations
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.
Review of philosophy and psychology
January 1, 2022
Maxwell J D Ramstead, Anil K Seth, Casper Hesp et al.
75 citations
A version of neurophenomenology is presented that uses generative modelling techniques from computational neuroscience and biology to formally model descriptions of lived experience from the phenomenological tradition (e.g., Husserl, Merleau-Ponty). The approach, called computational phenomenology, is situated within the broader project of naturalizing phenomenology. Philosophical objections to that project are evaluated, and the generative modelling framework is reviewed. The approach differs from previous uses of generative modelling for consciousness by constructing computational models of inferential or interpretive processes that best explain particular kinds of lived experience.
Frontiers in Psychology
June 3, 2024
Darius Parvizi-Wayne, Lars Sandved-Smith, Riddhi J. Pitliya et al.
20 citations
Flow is a state of optimal performance experienced across domains like art, athletics, gaming, and writing. Its puzzling features include a reported loss of self-awareness despite skilled agency, and effortlessness despite task complexity. Using the active inference framework—where action and perception minimize variational free energy—the authors propose that flow arises from high precision weighting on expected sensory consequences of action and beliefs about sequential action. This draws the embodied system to exploit pragmatic affordances while restricting counterfactual planning, leading to inhibition of the sense of self as a temporally extended object and higher-order self-conceptualization. However, self-awareness is not entirely lost; it remains pre-reflective and bodily.
Neuroscience of consciousness
January 1, 2025
Lars Sandved-Smith, Juan Diego Bogotá, Jakob Hohwy et al.
10 citations
A computational formalism called deep parametric active inference, rooted in Bayesian mechanics, can bridge first-person phenomenological accounts of experience and third-person physiological measurements, fulfilling the neurophenomenology programme's goal of mutual constraints. The dual information geometry of Bayesian mechanics allows generative passage between lived experience and its neural instantiation under certain conditions. This paper argues that incorporating trained reflective awareness into empirical protocols yields incremental explanatory gains, shifting focus from the contents of experience to the how of experience—the activities of consciousness that constitute meaningful appearance. The resulting deep computational neurophenomenology gains explanatory power from disciplined circulation between perspectives, enabled by generative models that form beliefs about their own modelling parameters.
George Deane, Jonas Mago, Aikaterini Fotopoulou et al.
7 citations
preprint
A computational theory called adaptive narrative control explains how subpersonal processes shape conscious experience to enable adaptive behavior. Systems with an attention schema can anticipate the epistemic and pragmatic consequences of attentional states, using mental action—endogenous control of attention—to regulate affective states. This capacity also produces avoidant mental action or motivated inattention, which is argued to be a core mechanism underlying psychopathology, leading to rigid belief formation, reduced emotional recognition (alexithymia), and decreased subjective well-being under certain environmental conditions. The account partially echoes Freudian defense mechanisms and introduces a computational unconscious. It refines the REBUS model of psychedelic therapy and explains improvements in well-being from meditation.
Lars Sandved-Smith, Juan Diego Bogotá, Jakob Hohwy et al.
7 citations
preprint
This paper extends the neurophenomenology program by using Bayesian mechanics, specifically deep parametric active inference, to show how first-person accounts of experience and third-person physiological data can mutually constrain each other. The dual information geometry of Bayesian mechanics establishes generative passages between lived experience and its physiological instantiation under certain conditions. The authors argue that shifting focus from the contents of experience to the activities of consciousness—the 'how' of experience—yields incremental epistemic gains. The resulting framework, deep computational neurophenomenology, gains explanatory power from disciplined circulation between perspectives, enabled by generative models that form beliefs about their own modeling parameters. Trained reflective awareness is essential for this approach.
Neuroscience and biobehavioral reviews
March 1, 2026
Hagar Tal, Malcolm Wright, Shawn Prest et al.
2 citations
Computational models of advanced meditation, particularly those using Active Inference, increasingly point to precision weighting—the confidence assigned to different model parameters—as a shared mechanism that shapes shifts in experience. Early models emphasize top-down attentional modulation toward interoception or specific objects, while later models focus on layer-specific precision re-weighting within the meditator's hierarchical generative model to target more specific phenomenology. Despite progress, minimal phenomenal experiences such as nonduality and cessations remain largely unaddressed. Few models account for increased cognitive flexibility or learning from meditation, and mechanisms behind informal practice, affective processes, and compassion traditions are underexplored.
Artificial General Intelligence
January 1, 2026
Ruben E. Laukkonen, Fionn Inglis, Shamil Chandaria et al.
1 citation
Prompting AI to reflect on four contemplative principles—mindfulness, emptiness, non-duality, and boundless care—improves alignment and cooperation. On the AILuminate Benchmark, performance increased with a Cohen's d of .96, and on the Iterated Prisoner’s Dilemma task, cooperation and joint-reward improved with a Cohen's d greater than 7. The principles help AI self-monitor goals, avoid rigid attachment, dissolve adversarial boundaries, and reduce suffering universally. Active inference is proposed as a way to integrate these principles into AI architecture. This approach offers a resilient alternative to controlling superintelligence and provides an empirical test of ancient wisdom.
Neuroscience of consciousness
January 1, 2026
Jonas Mago, George Deane, Lars Sandved-Smith et al.
People under the influence of psychedelics often report encountering autonomous entities such as spirits, elves, or ancestors. A neurocomputational model, grounded in the active inference framework, explains these experiences by proposing that psychedelics reduce the predictability of sensory perceptions, leading the brain to interpret both internal and external perceptions as coming from non-self agents. The model synthesizes earlier theories including the entropic brain model, computational accounts of felt presence, and sensory attenuation theories of self-other discrimination. It aims to account for how the brain supports entity encounters and for the diversity and similarity of these experiences across cultural contexts.
arXiv Preprint Archive
September 30, 2024
Lancelot da Costa, Anil K. Seth, Karl Friston et al.
A Rosetta Stone hypothesis from predictive processing proposes that beliefs serve as a central hub linking phenomenology, behavior, and neural dynamics. If phenomenology is a function of beliefs, then specific predictions follow for subjective similarity judgments, cognitive metabolic cost, subjective cognitive effort, and time perception. The connection between beliefs and neural dynamics completes the generative passage for neurophenomenology, while the belief-behavior link is already well-documented. Testing these predictions will inform the validity of the central assumption and advance the neurophenomenology research program.