Laukkonen, Friston, and Chandaria’s Beautiful Loop Theory

Neuroscientists Ruben Laukkonen, Karl Friston, and Shamil Chandaria’s Beautiful Loop Theory proposes that phenomenal consciousness arises when a system generates a unified reality model, selects its contents through coherent inferential competition, and recursively makes that model available to its own inferential processes. The theory holds that an active inference system must satisfy three conditions: it generates a unified, phenomenal reality model that serves as an “epistemic field”; inferences compete to enter that model, and only those that coherently reduce long-term uncertainty are bound into it (“Bayesian binding”); and the global model is recursively shared back throughout the hierarchy, so that it “contains the knowledge that it exists” (“epistemic depth”) (Laukkonen et al., 2025).

This third condition, “epistemic depth,” is intended to explain experiential knowing through system-wide forecasting and control of precision—the confidence assigned to predictions and evidence. Formally, epistemic depth is a global hyper-model that forecasts and broadcasts the precision assigned to every level of inference, so that the system models how it is modeling the world. Awareness is thus distinguished from content: content is fixed by what wins the competition, the clarity, or luminosity, of knowing by the depth of recursive sharing. Consciousness need not include an ego or represented first-person perspective. Applied to sleep, lucid dreaming, meditation, psychedelics, and artificial intelligence, the theory identifies the most minimal consciousness as a nearly contentless world model reflexively knowing itself (Laukkonen et al., 2025).

From Active Inference to Phenomenal Consciousness

Friston’s free-energy principle describes how self-organizing systems maintain states compatible with their persistence. The theory builds on active inference, in which organisms persist by minimizing precision-weighted prediction error within a hierarchical generative model, thereby evidencing their own existence (Friston, 2010; Feldman and Friston, 2010; Hohwy, 2016; see Friston’s Free-Energy Principle and Active Inference). Active inference models perception and action through probabilistic inference: organisms update expectations and select actions that fulfill preferences and resolve uncertainty. Variational free energy bounds sensory surprisal; expected free energy concerns prospective policies, including their epistemic value. Neither is simply thermodynamic energy, and minimizing it need not mean avoiding novelty (Friston, 2010).

The authors call the organism’s whole lived world the “generative phenomenal, unified world model,” or reality model: generative because internally constructed, phenomenal because “there can be an experience of” it, and unified because it appears bound together as a whole. It is also an epistemic field, a space of what can be known and acted on, and is necessary for consciousness because it defines what can become conscious. On the authors’ own account, such a model explains what is experienced, not why or how there is experience.

Beautiful Loop Theory adds conditions intended to distinguish conscious systems from systems merely performing adaptive inference. Its target is both the contents of experience and their being experienced. Computational “beliefs” here are probabilistic estimates, not necessarily conscious judgments. The proposal is naturalistic, but its abstract architecture permits different implementations; it does not identify every living or free-energy-minimizing system as conscious (Laukkonen et al., 2025).

Inferential Competition and Bayesian Binding

The second condition of inferential competition sets the threshold. Binocular rivalry and inattentional blindness show that some interpretations of sensory data become conscious and others do not. The theory proposes that explanations of sensory causes compete through such precision weighting, and that coherence with the existing reality model, acting as empirical priors, is the decisive criterion, since incoherent inferences propagate irreducible error and degrade planning. The binding problem thus becomes self-solving as part of the solution: what is bound is what wins. The same mechanism binds features into a face and the whole scene, body included, into one field. Global workspace “ignition” is reinterpreted as coherence-constrained precision competition (Dehaene et al., 2017; see Baars’s and Dehaene’s Global Workspace Theory), consistent with the view that conscious contents are precise posteriors informing policy selection (Whyte and Smith, 2021). A unified global posterior, however, is held to be necessary but not sufficient for consciousness.

Epistemic Depth and the Global Hyper-Model

The third condition of epistemic depth addresses the sense of knowing. Whichever percept dominates in rivalry, the authors argue, the presence of consciousness is unchanged; they adopt the view that a subject is aware not merely when sensitive to a state of affairs but when she knows she is sensitive to it (Cleeremans et al., 2020). Epistemic depth is the recursive sharing of the reality model with the hierarchy that generated it, likened to hearing one’s own voice while speaking. Borrowing the contemplative term luminosity, the authors treat awareness as one knowing that varies in clarity, like a lamp illuminating itself, which, according to the authors, blocks regress.

Formally, local free energy is minimized at each layer while a global hyper free-energy is minimized over hyperparameters, Φ, that predict precision at every layer; descending precision predictions and ascending precision-weighted errors update one another recursively, and because every layer lies within the hyperparameters’ Markov blanket, the loop is non-local (Kanai et al., 2015; Parr and Friston, 2018). Precision control is also cast as regulating phenomenal transparency and opacity: layers become “opaque,” and thus known, when precision is predicted upon them (Limanowski and Friston, 2018; see Metzinger’s No-Self Representational Theory of Subjectivity). The authors distinguish this from higher-order theories: epistemic depth achieves “sensitivity to sensitivity” through recursion rather than metarepresentation, so metacognition itself may be conscious or unconscious (compare Lau’s Perceptual Reality Monitoring Theory and Cleeremans’s Learned Consciousness via Radical Plasticity). Even a two-layer system that predicts and rebroadcasts its own precisions would have rudimentary epistemic depth, so consciousness precedes introspection and self-modeling.

Levels of Awareness, Sleep, and Lucidity

The conditions yield graded states. Low-precision inferences remain unconscious, as in subliminal priming. Bound contents with little epistemic depth are barely known, as in absorbed driving. In deep NREM sleep, the reality model becomes minimal and reflective broadcasting collapses; in REM, a rich model is bound but epistemic depth is low, so the dreamer does not know she is dreaming; lucidity is a rise in epistemic depth within REM (see Dreaming and Phenomenal Consciousness). Lucid dreamless sleep, described in Tibetan and Upaniṣadic sources as luminous awareness without content, corresponds to minimal abstraction with high epistemic depth (Windt et al., 2016).

Minimal Phenomenal Experience, Cessation, and Psychedelics

The theory takes minimal phenomenal experience, contentless “pure awareness,” as a central explanatory target because it separates consciousness as such from selfhood, agency, and time (Metzinger, 2024; Josipovic, 2019). Formally, minimal phenomenal experience occurs when first-order precision is uniformly low, so that no posterior collapses onto a scene, while precision on precision is maximal: one knows very clearly that one knows nothing. Meditative cessations (nirodha) are modeled differently, as “Bayesian unbinding,” in which deconstructive practice accumulates evidence against the coherence of the reality model until inferential competition fails and awareness transiently ceases (Laukkonen et al., 2023). For psychedelics, the authors conjecture that changes in content reflect relaxation of high-level priors (Carhart-Harris and Friston, 2019; see Carhart-Harris’s Entropic Brain Hypothesis), whereas felt “expanded consciousness” and noetic certainty reflect increased epistemic depth, which need not track accuracy.

Mechanisms, Machines, and the Function of Consciousness

Declining to localize the loop, the authors propose two non-region-centric broadcast channels for Φ: neuromodulatory “sprays” setting slow baseline precision and endogenous electromagnetic fields providing millisecond phase alignment (Shine, 2019; Hunt and Schooler, 2019; see Hunt and Schooler’s General Resonance Theory). For artificial intelligence, they recast the question as whether a system meets the three conditions, identify epistemic depth as the main gap in current systems, and sketch how self-certainty signals in language models (Zhao et al., 2025) could become a primitive hyper-model. Granting that such a system could still be a philosophical zombie, they argue that consciousness should then be attributed to it as it is to other people, and urge ethical caution. Functionally, they hypothesize that epistemic depth enables cognitive self-optimization and underlies general intelligence.

Subsequent Development

A substantive extension is the Entangled Loop, coauthored by Laukkonen and Chandaria and published online in October 2026. It gives emotion an architecturally primary role and extends binding to interference-based, “quantum-like” inference coordinated through hierarchical brain dynamics (Kringelbach et al., 2026). The authors’ accompanying explanation treats classical Bayesian inference as a limiting case and distinguishes quantum-like probability from microscopic quantum mechanisms (Centre for Eudaimonia and Human Flourishing, 2026). This subsequent framework preserves the self-evidencing architecture while adding stronger biological and dynamical commitments (it should not be retroactively attributed to the 2025 theory).

Reception and Assessment

Beautiful Loop Theory’s strength is its attempt to explain experiential content, unity, and reflexive presence within one computational architecture, while accommodating minimal and nonordinary experience. Its central unresolved issue is whether global confidence forecasting explains why anything feels like anything, rather than explaining sophisticated regulation and the sense of knowing. Because nonlucid dreams already possess phenomenality, identifying epistemic depth with luminosity also requires separating experiential existence from explicit recognition and clarity.

The theory answers the charge that active inference offers theories for consciousness science rather than a theory of consciousness (Seth and Bayne, 2022). Whyte and colleagues, including Friston, extract a minimal theory implicit across active inference models, centered on precise posteriors guiding action selection (Whyte et al., 2024). Kringelbach and colleagues, including Laukkonen and Chandaria, extend it by making emotion architecturally primary and proposing oscillatory interference as a binding mechanism (Kringelbach et al., 2026).

A review of computational models of meditation adopts Bayesian unbinding but notes that cessations, as non-states, may resist modeling in a framework built on self-preserving dynamics (Tal et al., 2025). Laukkonen has acknowledged in an interview that the loop is the component hardest to test (Eaton, 2025). Establishing distinctive predictions beyond global-workspace, recurrent-processing, and higher-order models remains essential. The paper offers a conceptual framework and formal sketch, not a completed empirical demonstration of sufficient conditions for consciousness (Laukkonen et al., 2025). The theory explains the selection, structure, and graded clarity of experience; whether recursive precision-sharing constitutes phenomenal consciousness or only correlates with it, the question posed by the hard problem (Chalmers, 1995), remains open.

References

Carhart-Harris, R. L., and Friston, K. J. (2019). REBUS and the anarchic brain: Toward a unified model of the brain action of psychedelics. Pharmacological Reviews, 71(3), 316–344.

Centre for Eudaimonia and Human Flourishing. (2026). The Entangled Loop: Frequently asked questions. University of Oxford. Institutional explanation.

Chalmers, D. J. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200–219.

Cleeremans, A., Achoui, D., Beauny, A., Keuninckx, L., Martin, J.-R., Muñoz-Moldes, S., Vuillaume, L., and de Heering, A. (2020). Learning to be conscious. Trends in Cognitive Sciences, 24(2), 112–123.

Dehaene, S., Lau, H., and Kouider, S. (2017). What is consciousness, and could machines have it? Science, 358(6362), 486–492.

Eaton, J. (2025). Consciousness catching its own tail [Interview with R. Laukkonen]. Asterisk, 10. https://asteriskmag.com/issues/10/consciousness-catching

Feldman, H., and Friston, K. J. (2010). Attention, uncertainty, and free-energy. Frontiers in Human Neuroscience, 4, 215.

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.

Hohwy, J. (2016). The self-evidencing brain. Noûs, 50(2), 259–285.

Hunt, T., and Schooler, J. W. (2019). The easy part of the hard problem: A resonance theory of consciousness. Frontiers in Human Neuroscience, 13, 378.

Josipovic, Z. (2019). Nondual awareness: Consciousness-as-such as non-representational reflexivity. Progress in Brain Research, 244, 273–298.

Kanai, R., Komura, Y., Shipp, S., and Friston, K. (2015). Cerebral hierarchies: Predictive processing, precision and the pulvinar. Philosophical Transactions of the Royal Society B, 370(1668), 20140169.

Kringelbach, M. L., Rosas, F., Laukkonen, R., Chandaria, S., Sanz Perl, Y., and Deco, G. (2026). The Entangled Loop: Emotion, interference and hierarchy in the architecture of consciousness. Neuroscience & Biobehavioral Reviews. [Preprint: The Entangled Loop: Emotion, quantum-like interference and hierarchy in the architecture of consciousness. PsyArXiv. https://doi.org/10.31234/osf.io/besw9_v2]

Laukkonen, R., Friston, K., and Chandaria, S. (2025). A beautiful loop: An active inference theory of consciousness. Neuroscience & Biobehavioral Reviews, 106296. https://doi.org/10.1016/j.neubiorev.2025.106296

Laukkonen, R. E., Sacchet, M. D., Barendregt, H., Devaney, K. J., Chowdhury, A., and Slagter, H. A. (2023). Cessations of consciousness in meditation: Advancing a scientific understanding of nirodha samāpatti. Progress in Brain Research, 280, 61–87.

Limanowski, J., and Friston, K. (2018). “Seeing the dark”: Grounding phenomenal transparency and opacity in precision estimation for active inference. Frontiers in Psychology, 9, 643.

Metzinger, T. (2024). The Elephant and the Blind: The Experience of Pure Consciousness—Philosophy, Science, and 500+ Experiential Reports. Cambridge, MA: MIT Press.

Parr, T., and Friston, K. J. (2018). The anatomy of inference: Generative models and brain structure. Frontiers in Computational Neuroscience, 12, 90.

Seth, A. K., and Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience, 23(7), 439–452.

Shine, J. M. (2019). Neuromodulatory influences on integration and segregation in the brain. Trends in Cognitive Sciences, 23(7), 572–583.

Tal, H., Wright, M., Prest, S., Sandved-Smith, L., and Sacchet, M. D. (2025). Active inference, computational phenomenology, and advanced meditation: Toward the formalization of the experience of meditation. PsyArXiv. https://doi.org/10.31234/osf.io/pm5y2_v2

Whyte, C. J., Corcoran, A. W., Robinson, J., Smith, R., Moran, R. J., Parr, T., Friston, K. J., Seth, A. K., and Hohwy, J. (2024). On the minimal theory of consciousness implicit in active inference. arXiv:2410.06633.

Whyte, C. J., and Smith, R. (2021). The predictive global neuronal workspace: A formal active inference model of visual consciousness. Progress in Neurobiology, 199, 101918.

Windt, J. M., Nielsen, T., and Thompson, E. (2016). Does consciousness disappear in dreamless sleep? Trends in Cognitive Sciences, 20(12), 871–882.

Zhao, X., Kang, Z., Feng, A., Levine, S., and Song, D. (2025). Learning to reason without external rewards. arXiv:2505.19590.