Subproject 1:

Psychometrics of pure awareness

Subproject 1 aims to develop a finer-grained phenomenological description of MPE. We use multi-language surveys and statistical methods to filter out invariances across different populations.

Example: Results of exploratory factor analysis of the MPE-92M questionnaire (N = 1393).

Questionnaire items are listed on the y-axis. Each column corresponds to a full factor solution with the number of extracted factors given on the x-axis. The size (area) of the circles is proportional to the factor loadings. Open circles indicate negative loadings, and loadings below 0.3 are not shown. Color represents factor membership. An attempt was made to identify the “same” factors across all factor solutions and color them identically. “Sameness” here is not objectively definable, but the degree of similarity was operationalized as membership in the same cluster following hierarchical cluster analysis. Items are sorted by factor membership and size of factor loading in the 12-factor solution, as it was this solution that was ultimately chosen as optimal.

Subproject 2:

Pure awareness during dreamless deep sleep

We are beginning a new research initiative investigating the phenomenon of “clear light sleep”. This is a state in which nonconceptual awareness continues during sleep with minimal phenomenal content. The minimal phenomenal content that remains is sometimes experientially described as including qualities such as expansiveness, luminosity and bliss, but is otherwise devoid of typical sensory or conceptual content. The state may be accessed from a lucid dream or by maintaining a witnessing mode of consciousness into sleep from the waking state.

The goal of this research is to perform the first scientific study and sleep laboratory validation of the experience of clear light sleep. This project could have multiple outcomes that are scientifically important. First, if an objective signal can be sent during this state, it could provide the first evidence of consciousness being maintained during deep NREM sleep. Second, it would enable for the first time the physiological and brain activity correlates of the state of clear light sleep to be measured. Both aspects could have far-reaching consequences for our understanding of the link between consciousness and brain activity.

Participants who have training in practices to enter the state, such as Tibetan “dream yoga,” are invited to work with us in this ongoing work. The main criterion for inclusion in the research is a capacity to reliably enter into a state of clear light sleep either from lucid dreams or the waking state. Participants spend between 1 and 4 nights in the sleep laboratory while brain activity and physiology are monitored using polysomnography (PSG). If you are interested in potentially being a participant in this study, please contact us to receive additional information.

Subproject 3:

Computational phenomenology of pure awareness

MPE no absorption episode graphic description

Subproject 3 aims to propose and refine computational models of MPE phenomenology. The current proposal leverages the active inference modelling framework, which casts experience as the output of an inferential process. The purpose behind this project is to arrive at novel and fruitful insights into the processes which might underpin MPE as well as to provide a conceptually more precise and fine-grained description than is possible in natural language.

Two models are proposed below, the first is a computational model of MPE without absorption (ie. containing other phenomenal content) and the second is a model of an MPE absorption episode. If you have any feedback or would like to propose your own model of MPE, please contact

This graphic illustrates an inferential architecture that would give rise to the experience of pure awareness as co-occurring with other forms experiential content. The phenomenal character of MPE is present, but thoughts, feelings, and sensations occur at the same time. Phenomenological examples would be episodic experiential states during dual mindfulness practice (e.g.Metzinger 2024,  Chapters 2-6) or the global MPE mode of witness consciousness (Chapter 19). The blue box at the first level describes perceptual experiences like seeing, hearing, or feeling the physical sensations caused by the breath or bodily movement. Inference at this level asks the question, “What are the likely causes of what I am experiencing through my senses?”.

Proposed deep generative model
of Minimal Phenomenal Experience with content

The red box containing the blue box illustrates a higher-level model of the “likelihood precision” parameter in the level below which modulates sensory gain, i.e. attention. Modelling this process opens an attentional mental action policy-space available to the system. Phenomenologically, this level describes the experience of conscious attention since inference here amounts to asking the question, “How is my attention allocated to perceptual objects?”. Therefore, changes at this level may describe the introspective experience of, for example, deliberately focussing on the breath by actively making perceptual observations (at the level below) more precise. Taken together, and in terms of our psychometric study, the first two boxes mostly refer to “Time, Effort, and Desire” (Factor 1), “Sensory Perception in Body and Space”(Factor 9), plus the experience of “Mental Agency” (Factor 11).

Level 3 in our figure (the green box) models the likelihood precision at the second level and describes all situations in which we gain or lose the non-conceptual, experientially direct awareness of our mind, specifically here the awareness of our current mental state of attention. Inference at this level asks the question “How aware am I of my own mental state(s)?” By performing inference at the first, second and third levels we can be consciously paying attention to, for instance, a flow of physical sensations and simultaneously be aware of sudden shifts in attention. This is a typical example of experiences occurring during mindfulness practice like Shamata or Vipassana meditation, in which we deliberately, but non-judgmentally, observe our breath and monitor the wandering of attention.

The fourth and largest box finally describes the phenomenology of “epistemic openness”, or pure awareness. Computationally, the precision parameter being modelled at each higher-level is the mathematical element which endows the system with a receptivity to incoming information. Fourth level state inference therefore amounts to inferring the precision mechanism which enables perception across levels of inference, i.e. the global ‘epistemic openness’ to incoming information. Phenomenologically, one might argue that to infer this level is to become aware of awareness itself by asking the question “How aware is awareness of itself?”.

Why is awareness consciously aware of itself in this situation? Because conscious awareness appears when a system is epistemically open to (i.e. aware) the world and additionally knows about this fact. It is Level 4 which really endows the content that is generated on all lower-order levels with phenomenality–with the quality of “being consciously experienced”–because it adds the expectation that something can (and likely will) now be known that generates this lower-order content. Consciousness is a global prediction of epistemic gain that functionally integrates all lower-order content by meta-representing it as something that is now available to be explored in a deeper way. Recall from chapter 4, MPE is a continuing, ongoing process of nonconceptually experiencing the organism’s current state of epistemic openness, i.e. of expecting new knowledge without yet having it. This is what modelling its states at the fourth level amounts to; a single, integrated model of its own epistemic openness in the form of a highest-level, but non-egoic self-model. Without this fourth level we could say that awareness is transparent: We see “with” it or “through” it, but are only aware of the lower-order content. But now, with the fourth level, awareness itself is “opacified”, as when we suddenly become aware of the window pane through which we look out into the garden.

In summary, this model describes the property of epistemic openness as the all-pervading (non-zero) likelihood precision which makes the system open and responsive to new information. Without this feature, experience can only arise from inferences based on already existing beliefs. So, by having a global model of the likelihood precision across all levels of the hierarchy, the system has an explicit representation of its capacity for epistemic gain–that is, its current degree of epistemic openness. The idea is that this provides a computational definition of MPE, which can help us to formulate a “minimal model explanation” for consciousness (cf. note 1 in Chapter 1).

This and the next figure are only meant as discussion starters to illustrate that, in principle, recent advances in computational phenomenology leveraging the active inference framework may begin to shed light on how to model or simulate MPE. It is also worth acknowledging and remembering that “the math is not the territory” (Andrews, 2021), in the sense that there will always be a categorical schism between the abstract and simplified computational model of a certain target phenomenology and the much more holistic, subtle, and fluid experience that hypothetically corresponds to it.

Technical remarks:

This figure shows a probabilistic Bayes graph of a deep generative model with four hierarchical levels of state inference. The model represents the active inference (Friston et al., 2016) process underpinning perception and action at multiple levels of ‘parametric depth’, and is structured to plausibly represent the inferential architecture required for an MPE experience with sensory and mental content. Parametric depth refers to a hierarchical model in which each higher-level state models a parameter of the inferential process unfolding at the level below (Hesp et al. 2019). In this case we consider higher-level states which parameterise the likelihood mapping precision, γA, at the level below (Sandved-Smith et al., 2021).

The likelihood mapping A is a matrix of probability densities which encode beliefs about how observations relate to causes, i.e. P(o | s). Inference at each level then inverts the likelihood mapping and combines this with prior beliefs about states to obtain (or approximate) the most probable state P(s | o). This approximate posterior belief is the computational equivalent of the system’s phenomenological perception under this framework. The initial state vector, D, specifies beliefs about the most likely state of the world independently of any observation, i.e. the prior beliefs P(s). The model is also equipped with transition beliefs B about how states evolve over time. These expectations are dependent on the actions, u, the system believes it is currently undertaking. Action selection is performed by selecting the sequence of state transitions (B matrices) associated with the least expected free energy G. Two additional parameters implicated in the action selection process are the prior preference mapping (or C matrix), which specifies prior beliefs about sensory outcomes, and the prior over policies (E). The E matrix encodes beliefs about what the agent would do, independent of the expected free energy in the current context. The expected free energy scores the posterior probability of different allowable action sequences in terms of outcomes. For a detailed introduction to the active inference framework readers are referred to (Smith et al., 2022).

In active inference, “attentional processes” have been formulated in terms of the precision of the likelihood mapping A (Feldman, Friston, 2010). Intuitively, we can see why precision-modulation over A corresponds with attentional processes: the precision on A represents the extent to which the agent believes their observations “accurately map onto” hidden states. Attending to some stimuli increases the relative weight or gain on inferences made on the basis of that particular data or observation. For example, by paying closer attention to an ambiguous sound, we have greater confidence in determining the location of its origin than when the sound was first heard without being heeded. The process of combining available data with estimations of that data’s reliability—in order to arbitrate its effect (relative to prior beliefs) on the overall inferential process— is known as “precision weighting” or “precision control”. Under active inference, this is the candidate mechanism for attentional modulation of perception. Crucially, this precision creates the computational conditions for the system’s ability to register incoming information.

Hence second level states, s(2), can be interpreted as attentional states, which modulate the confidence in interoceptive or external sensory observations. In the same way, meta-attentional states, s(3), modulate the confidence in higher-order observations of the attentional state and awareness-of-awareness states, s(4), modulate the confidence in the higher-order observations of the meta-awareness states.

This model is a proposal for an inferential structure that can be used to simulate the phenomenology of a form of minimal phenomenal experience at the fourth level. As we move up the hierarchy, each precision becomes the basis for the state inference above. Since precisions are summary statistics of a probability density, the information they encode becomes more abstract and ‘simplified’ as we climb to higher states. Hence by the fourth level all conceptual content is abstracted away and the state inference amounts to inferring soley the precision mechanism which enables perception across levels of inference, ie. “epistemic openness” to incoming information.

MPE absorption episode graphic description

This graphic illustrates the inferential architecture that would give rise to the experience of pure awareness during a full-absorption episode. The phenomenal character of MPE is present and is the only quality that can later be reported, as no thoughts, feelings, or sensations are experienced at the same time.

For example, states of conscious experience in which nothing but the experiential quality of pure awareness itself is present may occur during deeper states of formal practice, e.g. sitting meditation, but also in phases of NREM-sleep, sometimes called “clear light sleep” (Metzinger 2024, Chapter 20). Full-absorption episodes can also be described as a form of “non-egoic self-awareness”, because we find the phenomenal character of epistemic openness, and there is a reflexive element to it. Phenomenologically, there is an ownerless and non-agentive quality of “knowingness” or insight in which awareness (but not the meditator, who doesn’t exist any more) is effortlessly aware of itself. There is no subject, only awareness non-conceptually and timelessly knowing itself. The existence of full-absorption episodes clearly demonstrates that agentive and egoic forms of self-consciousness, time representation, and self-location in a spatial frame of reference are not necessary conditions for conscious experience to occur.

Proposed deep generative model an MPE absorption episode

Computationally, one may speculate that the inferential structure describing MPE as non-egoic self-awareness is implicitly present in all other, more complex forms of consciousness and that it may actually constitute the core causal factor. This would be in line with the view of MPE as related to tonic alertness and the functional process by which the brain activates itself, for example when we wake up in the morning. If this is correct, the structure presented above may constitute a first step towards a minimal model explanation for consciousness: Consciousness is an integrated meta-representation of epistemic capacity per se.

As described in the previous graphic, performing inference at each level of this proposed mental architecture relates to experiencing the world and self at varying degrees of information abstraction. At the highest, and most abstract, fourth level, the experience is simply a self-evidencing representation of the system’s fundamental ‘epistemic openness’, whereas at the nested, lower levels we experience the complex inferred causes of our sensory data and the movement of thought and attention. An MPE absorption episode can therefore be modelled as the situation in which experience is only a result of inference at this fourth level, without any perceptual inference at the levels below. This would occur when the system no longer attends to lower level observations and is unable to perform inference at those levels, whilst continuing to infer at the fourth level.

The non-egoic, unbounded, atemporal and aperspectival quality of this experience (Metzinger 2024, Chapter 2) arises from the absence of possible actions at this level. Since the capacity for awareness is always already present, this state is modelled as being unavailable to manipulation by some action policy. Because it is not something that can or must be controlled, it is also experienced as ownerless and not tied to a first-person perspective. This also makes it maximally simple, because it eliminates the scope for a temporally deep action model at this level, which entirely removes the agentic phenomenological quality related to temporal depth (Friston, 2018). The tentative model presented in the preceding two figures has been conceived by Thomas Metzinger and Lars Sandved-Smith; figures have been created by Lars Sandved-Smith.

Technical remarks:

This figure represents a particular case of the previous figure, demonstrating the active inference process underpinning perception and action at multiple, nested levels of “parametric depth” when multiple levels of perceptual experience are absent due to the mechanism of sensory attenuation.

When attenuated, incoming data is no longer relied on to perform perceptual inference and the related experience that data would give rise to is no longer perceived. This occurs when the system internally decreases its confidence in incoming data and amounts to reducing the precision (or confidence) in the learned mapping between how likely a particular cause is given some data (the likelihood mapping A). This can be a result of deliberate mental action (e.g. attending away from an annoying sound in an attempt to ignore it) or due to a shift in context (e.g. when the lights go out and visual sensations can no longer be considered informative). This decrease in the precision of the likelihood mapping is illustrated in the figure by the darkly shaded precision terms, γA. By attenuating the lower levels, whilst maintaining precision at the fourth, the system will experience only the most abstract form of content (MPE), i.e. an absorption episode of pure awareness.

Dark shaded circles represent a low level of likelihood precision on observations arising from levels 1-3 in the hierarchy. This in turn eliminates the confidence in the perceptual posterior calculation and has the phenomenological consequence of losing access to the related perception. Therefore, in the model shown, the only source of perceptual experience is arising from the inference calculation performed at the fourth level (where likelihood precision is maintained). Only states at that level are estimated (and experienced) successfully. This can plausibly represent an MPE episode without any phenomenological content.

Biologically, the likelihood precision has been related to the effects of the cholinergic neurotransmitter system and its modulatory impact on synaptic gain (i.e. the strength of neuron signals) (Parr & Friston 2017;  Dayan & Yu 2002). A plausible interpretation leading to an empirically testable interpretation therefore is that an MPE absorption episode is a high concentration of cholinergic neurotransmitters in regions encoding the high-level mappings, combined with a depletion of cholinergic neurotransmitters at the lower level mappings.

Scroll to Top