From self-organizing systems to subjective temporal extension Authors: Jan Erik Bellingrath Published: 2026 ABSTRACT AbstractThe self-simulational theory of temporal extension (SST) describes an information-theoretically formalized mechanism by which the “width” of subjective temporality emerges from the architecture of self-modeling. In this paper, the perspective of the free energy principle will be assumed to cast the emergence of subjective temporal extension from first principles of the physics of self-organization and to formalize subjective temporal extension using information geometry. Using active inference, a deep parametric generative model of temporal inference is simulated, which realizes the described dynamics on a computational level. Two “biases” (i.e. variations) of time-perception emerge naturally from the simulated computational model. This concerns the intentional binding effect (i.e. the compression of the temporal interval between voluntarily initiated actions and subsequent sensory consequences) and empirically documented alterations of subjective time experience in deep states of meditative absorption (i.e. in minimal phenomenal experience). Generally, numerous systematic and domain-specific alterations of subjective temporal experience are computationally explained in a unified manner, as enabled by integration with current active inference accounts mapping onto the respective domains. This concerns, next to more general attentional and central tendency effects, the temporality-modulating role of valence, impulsivity, boredom, flow-states, near death-experiences, and various psychopathologies, among others. SST, from the perspective of the free energy principle, explains how the “width” of the subjective temporal moment emerges from first principles, accounting for why sometimes, subjective time seems to fly, and sometimes, moments feel like eternities; with the computational mechanism being readily deployable synthetically.Keywords: active inference, subjective temporal extension Introduction ”Time is the substance I am made of. Time is a river which sweeps me along, but I am the river.” (Borges, 1962) Has time ever seemed to pass by too quickly? Have moments ever seemed like an eternity? While the subjective experience of time seems as prevalent as the experience of change, the experience of the subjective present is not simply the experience of change, or change of change, etc. (Kent and Wittmann, 2021). Our experience of subjective temporality is the experience of an extended moment, a specious present (James, 1890), extending retrospectively to the moments that just happened, and prospectively to the moments that are predicted to subsequently arise. Historically, varying estimates for the “width” of a subjective temporal moment range back over the centuries. For example, the Buddhist Treasury of Abhidharma estimates there to be 65 temporal moments in the time it takes to snap one’s fingers, with the “width” of the subjective temporal “Now” thereby being taken to be about 1/65 of a second (i.e. about 15 ms) (Thompson, 2014). Essentially, nothing in physical time, considered as described by physics, straightforwardly corresponds to an extended phenomenal Now in the subjective sense. Accordingly, and given a characterization of conscious content in terms of hierarchical generative modeling, the “width” of subjective temporality cannot be perceptually inferred from external surroundings but must be the consequence of a counterfactual inference process (i.e. prospective or retrodictive inferences), entertained at a given moment of physical time. Thus, the cognizing system effectively entertains a hypothesis concerning how the world would look, if there were subjective temporal extension. Counterfactual self-simulations constitute one possibility for this inferential process: at least in the normal waking state, it looks as if there is always “someone” flowing in the temporal river. Mirroring this phenomenology, numerous empirical connections between temporal inference and self-modeling have been discovered. Here, and in the following, subjective temporal extension is assumed to have a systematic relationship to empirically measured duration inferences and passage-of-time judgments, with a longer temporal extension mapping onto increased duration estimates and a comparatively slower passage of time. As an example of an empirical connection between temporal inference and self-modeling, signals from the heart cycle and timing responses in a duration reproduction task are characterized by a certain synchrony (Pollatos et al. 2014b). Additionally, the amplitude of the heartbeat-evoked potential is associated with the length of the estimated durations (Richter and Ibáñez, 2021, Khoshnoud et al., 2024); and individuals who demonstrate a higher interoceptive accuracy (i.e. individuals that are more accurate at counting the number of heartbeats during a given interval) show a better performance in duration reproduction tasks (Meissner and Wittmann, 2011). Consistent with these studies suggesting the existence of a system-relative (i.e. embodied) temporal reference frame, neuroimaging meta-analyses reveal both self-modeling and temporal inference as centrally implicated in insula functioning (for a meta-analysis on neuroimaging studies of self-modeling, see e.g. Qin et al., 2020; for meta-analyses on neuroimaging studies of time perception, see Mondok and Wiener, 2023, Naghibi et al., 2023). This is consistent with earlier proposals of the insula cortex as, on the one hand, the neuronal substrate of subjective time perception (Craig, 2009) and the conscious feeling of presence (Seth et al., 2012); and, on the other hand, the neuronal substrate of interoceptive inference and central aspects of self-modeling (e.g. Seth, 2013, Seth and Friston, 2016). Moreover, subjective time perception and the conscious feeling of presence have already been, and by these very accounts, tied to the “material me” (Craig, 2009) and self-modeling and agency more generally (Seth et al., 2012). Psychologically, furthermore, various normal (e.g. Rey et al., 2017, Witowska et al., 2020) and altered (for a review see Wittmann, 2015) states, traits, and psychopathological conditions (e.g. Martin et al., 2014, Kent et al., 2023) are associated with covariant changes in self-modeling and temporal inference. These states and traits include, as will be reviewed and explained below, among others, varying levels of valence, impulsivity, boredom, the flow-state, near-death experiences, the intentional-binding effect, and concentrated states of deep meditative absorption. Consistently, many of these temporality-modulating constructs are related to functional activation, and structural alterations, in the insula cortex (for a review in the context of psychopathologies, for instance, see Vicario et al., 2020). While this covariation, holding across levels of description, methods, tasks, and theoretical perspectives, is too close for a coincidence to be likely, a theory explicitly relating self-modeling and temporal inference, in the sense of one being emergent from the other, has been lacking. The self-simulational theory of temporal extension (SST) is a theory about the emergence of subjective temporal extension (i.e. the “width” of subjective temporality) from processes of self-modeling, formalized with information theory and formulated on a neuronally plausible level of description (Bellingrath, 2023). In this paper, the perspective of the free energy principle (FEP) (Friston, 2010) will be assumed to cast the emergence of subjective temporal extension from first principles of the physics of self-organization. The first part of this paper will, after a section introducing the FEP and active inference, be concerned with an information-geometric and a computational description of the “width” of subjective temporality. In the second part of this paper, the described dynamics are simulated with a deep parametric generative model of temporal inference. Two empirically documented “biases” (i.e. variations) of time experience will be shown to naturally emerge from the simulation of the computational model. In the third part of this paper, empirical evidence on systematic variations of temporal experience from diverse psychological states, traits, and psychopathological conditions will be explained in a unified manner, based on active inference accounts mapping onto the respective domains. This deployment of mutual constraints between phenomenological, computational, neuronal, and behavioral levels of description is historically and methodologically in line with the neurophenomenological research program outlined by Francisco Varela (Varela, 1996), which has recently, based upon integration with the FEP and active inference, been revived in terms of the nascent field of computational (neuro-) phenomenology (Ramstead et al., 2022, Sandved-Smith et al., 2025). This improved version of SST (Bellingrath, 2023) is mathematically more principled, empirically more meaningful and predictive, and theoretically more integrated with current models in the cognitive sciences, and can be considered as SST 2.0. Introduction to the free energy principle and active inference If things exist, what must they do? Specifically, how can self-organizing systems, or things, “maintain their states and form in the face of a constantly changing environment” (Friston, 2010)? This question is addressed by the FEP (Friston, 2010; see also, Clark, 2013, Hohwy, 2013, Clark, 2015, Parr et al., 2022, Friston et al., 2023). Originally proposed within the context of theoretical neuroscience as a unified brain theory (Friston, 2010), accounting for perception, attention, action, and numerous associated phenomena, the explanatory scope of the FEP has been expanded to include every thing (Friston, 2019). But what does it mean to be a thing in the first place? A thing can be characterized via the notion of a Markov blanket (Pearl and Smets, 1998). A Markov blanket, , is the set of states that separates the internal parts of a system, , from its surrounding environment, . Intuitively, for the brain, internal (neural) states are statistically insulated from external objects by sensory receptors and muscles. This is a statement of conditional independence: given knowledge of the blanket, internal and external states are conditionally independent—any influence that or have on one another is mediated by (Parr et al., 2019): This enables us to speak about both the system we seek to characterize and its surrounding environment, as entities, or things that can be individuated from their environment. To remain in their characteristic low entropy state, however, things must, on average, minimize the model-conditional improbability of impinging sensory data, known as surprise, (Friston, 2010). This is because entropy, , is the long-term average of surprise, : Mathematically, the minimization of surprise is equivalent to the maximization of model-evidence, , enabling a formulation of systemic preservation in terms of Bayesian inference (i.e. self-evidencing (Hohwy, 2016)). However, as exact Bayesian inference is usually computationally intractable, as there are many hidden states that all need marginalizing out, the intractable quantities (the model-evidence, , and the posterior probability, ) are substituted with quantities that approximate them (the variational free energy, , and the approximate posterior, (i.e. the recognition distribution)) (Parr et al., 2022). Accordingly, the minimization of variational free energy is equivalent to the maximization of model evidence, , while simultaneously minimizing the KL-divergence, , of the approximate posterior, , from the posterior probability, . Variational free energy, , is equivalent to surprise, , if this divergence is zero (i.e. free energy is an upper bound on surprise) (Parr et al., 2022): This enables a perspective on the emergence of genuinely inferential architectures from nothing but systemic preservation itself (e.g. Friston et al. 2017a). For example, in active inference, the minimization of corresponds to perception and the maximization of corresponds to action (i.e. selectively sampling and changing the world to conform to prior expectations about it) (Parr et al., 2022). Considering a simple generative model of Bayesian perceptual inference (see Fig. 1), the likelihood mapping from hidden causes to their outcomes, , is inverted using prior beliefs, , and sensory data, , giving the perceptual inference of hidden causes given outcomes, . The likelihood mapping, (i.e. ), is equipped with a precision term, , by exponentiating each element in the th row and th column of by and normalizing (Fig. 1) (Parr et al., 2018). This precision-weighting is understood in terms of attentional processes (Feldman and Friston, 2010). A higher precision-weighting maps onto an increased attentional resource-allocation, with faster and more efficient inferences being resultant (e.g. Sandved-Smith et al., 2021). A probabilistic graphical model showing a basic generative model of precision-weighted perceptual inference. A prior over hidden states is combined with sensory data to infer the posterior over hidden states given observations. Here, Bayesian beliefs are noted in bold, bar notation represents posterior beliefs, is the softmax function (returning a normalized probability distribution), and is the Kronecker delta function (returning 1 for the observed outcome and zero for all non-observed outcomes). The precision term, , over the likelihood mapping, , is sampled from a gamma distribution with inverse temperature parameter . For the derivation of the precision belief update equation see Parr and Friston (2017) (supplementary appendix A.2); for the derivation of the latent state belief update equation, see Friston et al. (2017a) (supplementary appendix A). Figure and parts of description reproduced with permission from Sandved-Smith et al. (2021); adapted from a template given in Fig. 1a in the study by Hesp et al. (2021). While variational free energy enables inferences about the future based on past data, it does not yet account for prospective forms of inference (i.e. planning as inference (Botvinick and Toussaint, 2012)) over extended timescales based on anticipated future consequences (Fig. 2). Counterfactual simulations of sequences of actions (i.e. policies, ), associated with expected future observations under those policies, are evaluated in terms of expected free energy, (Friston et al., 2015). Mathematically, the minimization of is equivalent to the minimization of the divergence of the outcomes expected under a policy from the outcomes phenotypically expected (i.e. risk), , while simultaneously minimizing ambiguity, (i.e. the expected inaccuracy due to an ambiguous mapping between states and outcomes) (Parr et al., 2022): A probabilistic graphical model showing a deep generative model for policy selection. This model is equipped with beliefs about state transitions. Posterior state beliefs at each time step now depend on beliefs about the previous and subsequent states, mediated by the state transition matrix, B. (Figure and some parts of description reproduced with permission from Sandved-Smith et al. (2021); adapted from a template given in Fig. 2 in the study by Hesp et al. (2021)). can also be decomposed in terms of epistemic and pragmatic value, with epistemic value, , corresponding to the amount of information expected to be gained under the pursuit of a particular policy, and pragmatic value, , being an expected utility: Policies evaluated in terms of low are a priori more likely to be enacted. Equivalently, policies that have a high are a priori surprising, which leads to their avoidance (Friston et al., 2015). Essentially, free energy minimizing systems come equipped with prior preferences, : these prior beliefs assign high probabilities to observations that are preferred. For example, human beings probabilistically expect their blood-glucose level to be in a phenotypically beneficial range. By probabilistically presupposing prior preferences, , and by then acting to bring them about (as enabled by the selection of actions in terms of , influenced by ), self-organizing systems self-evidence (Hohwy, 2016) as adaptive self-fulfilling prophecies (Friston, 2011). Selected policies are also influenced by a baseline prior over policies, , which can be understood in terms of the influence of habits (i.e. what the agent would do independent of ). For detailed introductions to active inference see (Parr et al., 2022, Smith et al., 2022; see also Friston et al. 2017a). The FEP and active inference present a parsimonious framework for computational phenomenology (Ramstead et al., 2022, Sandved-Smith et al., 2025). Many phenomenologically and computationally central phenomena, such as self-modeling (Limanowski and Blankenburg, 2013), attention (Feldman and Friston, 2010), valence (Hesp et al., 2021), and meta-awareness (Sandved-Smith et al., 2021), have already been characterized in terms of the FEP and active inference. This includes excellent work on aspects of subjective time experience (Hohwy et al., 2016, Wiese, 2017, Albarracin et al., 2023, Bogotá and Djebbara, 2023), including its interrelation with episodic memory, and the influence of structured perceptual change (Fountas et al., 2022). Furthermore, there is active research on algorithmic instantiations of active inference. For instance, and relevant in the current context by virtue of the centrality of policy simulation and evaluation, branching time active inference provides a possible, and more computationally tractable, tree-search-based scheme for planning under uncertainty in active inference (Champion et al. 2022a; b). Subjective temporality as inference “In this account one’s body is processed in a Bayesian manner as the most likely to be ‘me”’ (Apps and Tsakiris, 2014) Self-modeling is a complex multidimensional process ranging across various experiential domains, comprising interoceptive, emotional, and bodily aspects, next to social, intentional, narrative, and more broadly cognitive aspects (e.g. Metzinger, 2003). The minimally sufficient form of phenomenal self-modeling is known as minimal phenomenal selfhood (MPS) (Blanke and Metzinger, 2009). MPS is the simplest form of self-consciousness. In its original formulation, MPS comprised three necessary and sufficient conditions: (i) global self-identification with the body as a whole, (ii) spatiotemporal self-location, and (iii) a (weak) first-person perspective (Blanke and Metzinger, 2009). In the following text, and even though SST is also compatible with MPS being a cluster-concept without a central individuating essence, current functional level descriptions of self-modeling will be presented in the context of MPS (for a review on minimal self-modeling and the FEP, see Limanowski and Blankenburg, 2013). This will illustrate how subjective temporal extension emerges in the context of the simplest forms of self-modeling. Self-identification (with the body as a whole; the first condition of MPS) is the consequence of a transparent inferential format (Metzinger, 2003, Blanke and Metzinger, 2009). An inference is transparent if the system is unable to attend to the process of construction of that inference. Precisely by not being able to introspectively notice that a given inference is, in fact, constructed, the inference is imbued with a transparent (i.e. “naïve realistic”) status. Conversely, if the system can direct its attention to the process of construction of a given inference, that inference can be inspected opaquely as “just a construct” (i.e. “just a thought”) (Metzinger, 2003). The opacification of a transparent inferential format is, in active inference, understood in terms of the deep parametric architecture of hierarchical generative models (Limanowski and Friston, 2018, Sandved-Smith et al., 2021), with “parametric depth” referring to a nested architecture of “beliefs about beliefs” (Hesp et al., 2021). Here, the precision-weighting, , of the likelihood, (i.e. ), corresponds to a central aspect of the process of construction of a given probabilistic inference. Accordingly, the process of construction of a hierarchically lower level probabilistic inference can be inspected by predicting its precision-weighting and inferred state via an inference on a hierarchically higher level—with an opaque inferential format of the hierarchically lower level inference being resultant (Limanowski and Friston, 2018, Sandved-Smith et al., 2021). Transparency (i.e. “naïve realism”) is the default inferential format (Metzinger, 2003), as evidenced both phenomenologically and by the computational cost opacification (i.e. the additional presence of inferences on a hierarchically higher level) engenders. Interestingly, complete intentional opacification is logically impossible, because the highest-order inference is necessarily transparent: its precision-weighting and inferred state would have to be predicted by yet another higher-order prediction, which leads to an infinite regress of hierarchical height (Limanowski and Blankenburg, 2013). Expressed in the words of Friston et al. (2012): “I can never conceive of what it is like to be me, because that would require the number of recursions I can physically entertain, plus one.” While the global self-identification with the body as a whole (the first condition of MPS) can be understood in terms of a transparent inferential format of the body-representation (Blanke and Metzinger, 2009), empirical evidence from asomatic out-of-body experiences and bodiless dreams suggests that a global self-identification with a body is not a necessary feature of MPS, as in these states, self-identification occurs without a global body-representation (e.g. self-identification as an extensionless point in space) (Alvarado, 2000, Metzinger 2013b, Alcaraz-Sánchez et al., 2022). The second condition in the original formulation of MPS, spatiotemporal self-location (Blanke and Metzinger, 2009), refers to a representation of a determinate volume in a spatial frame of reference, normally localized within the bodily boundaries as represented, alongside the represented “Now” in a temporal frame of reference (from this conceptual perspective, MPS presupposes (a minimal degree of) temporal “width,” with the latter, as will be detailed below, emerging from subpersonal (i.e. sub-MPS) dynamics of aspects of self-modeling, accordingly). A weak first-person perspective, the third and last necessary condition of MPS, is a purely geometric feature of self-modeling, which functions as the geometric origin, or center of projection, of the embodied system (Blanke and Metzinger, 2009). While the weak first-person perspective was originally tied to perceptual processing (e.g. visuospatial- or auditory-processing) (Blanke and Metzinger, 2009), the weak first-person perspective has also been understood completely independently of perceptual processing (Windt, 2010). Functionally, in terms of Bayesian inference, Metzinger (2013b) identifies the origin of the first-person perspective not with a specific inferential content, but rather with the systemic region of maximal invariance. Indeed, the very existence of the agentic model mandates the inclusion of the prior expectation that its form and internal states are contained within some invariant set (Friston, 2011). As Friston (2011) writes: “This is easy to see by considering the alternative: if the agent (model) entailed prior expectations that it will change irreversibly, then (as an optimal model of itself), it will cease to exist in its present state. Therefore, if the agent (model) exists, it must a priori expect to occupy an invariant set of bounded states (cf., homeostasis).” The hierarchically highest (i.e. maximally invariant) inference, the belief that “I am a cause in the world, or an agent” (Hohwy et al., 2017), is thereby equivalent to the belief that “I act to maintain myself in certain states” (Hohwy et al., 2017), i.e. the prior preferences. As elaborated above, prior preferences, denoted by , specify phenotypically expected outcomes, , such as the expected blood-sugar level, or the expected socio-emotional situatedness, and it is precisely the systemic presupposition of the prior preferences that engenders their adaptive actualization akin to a self-fulfilling prophecy (Friston, 2011). “Given the important links between the notion of priors and the conditions that undergird an organism’s existence, we can also say that in Active Inference, the identity of an agent is isomorphic with its priors” (Parr et al., 2022). Self-modeling, however, is not unidimensional. In counterfactual inferences (i.e. in prospectively predicting or retrodictively recalling), and for the transient time the simulated episode persists, an aspect of our self-model is being simulated (Friston, 2018, Limanowski and Friston, 2020). This aspect of our self-model “lives through” the simulated action sequence, being confronted with the “outcomes that would result if I were to act this way” (i.e. ). This is the domain of mind-wandering (Smallwood and Schooler, 2015), episodic memory (Tulving, 2002), planning as inference (Botvinick and Toussaint, 2012), and related notions that centrally comprise self-relational inferences in counterfactuality. Indeed, it has even been suggested that it is precisely the change in the unit of identification that could serve to individuate individual episodes of mind-wandering (Metzinger 2013a). Owing to the generality of the computational problem that these counterfactual inferences (i.e. policies) are engendered to solve (i.e. minimization of expected free energy), policy simulations are a virtually all-prevalent aspect of our inferential architecture, being present over nested timescales, different levels of hierarchical height, and in diverse inferential domains (e.g. Friston et al. 2017c, Pezzulo et al., 2018, Badcock et al., 2019). Motivationally, the outcomes expected under a counterfactual self-simulation, , are always evaluated in terms of a comparison (i.e. dissimilarity-relation) with phenotypically preferred outcomes, . Intuitively, “what could” is always motivationally evaluated in terms of a comparison with “what ought.” In active inference, this dissimilarity-relation between and is expressed with the KL-divergence (Kullback and Leibler 1951), a central concept in information theory. The divergence of from , , constitutes part of the expected free energy, , decomposed in terms of risk (i.e. ) and ambiguity (i.e. ), in terms of which policies are evaluated (Parr et al., 2022): As elaborated above, and are computational level descriptions of aspects of self-modeling. Expressed from the transparent perspective of self-modeling itself, corresponds to the “change that would have to happen for (a part of) me to become that other (part of) me.” The central claim of the SST is that this self-simulational dissimilarity relation is identical to the “width” of subjective temporality: the subjective extension of time is equivalent to the lived experience of the “length of the path toward another version of yourself.” Length, in this instance, has a technical meaning as we will see below; namely, it refers to the information length that is closely related to temporal expression via the accompanying information rate (i.e. the change in information length or distance per unit clock or universal time). Mathematically, by virtue of the properties of the KL-divergence, cannot be negative (i.e. ). Indeed, it is not even clear what a “negative width” of subjective temporal extension could mean in principle. Crucially, nothing in physical time, considered as described by physics, straightforwardly corresponds to an extended phenomenal Now in the subjective sense. Accordingly, because the “width” of subjective temporality cannot be perceptually inferred, must be the consequence of a counterfactual inference, entertained at a given moment of physical time. Because (i.e. “what would result, if I were to act this way”) is a counterfactual inference, it neither unfolds in the 3D space available for external action, nor in the n-dimensional space available for internal (attentional) action. Technically, unfolds on a statistical manifold or belief space, equipped with a well-defined metric that induces the information length that subtends temporal extension, as elaborated below. However, a given counterfactual self-simulation could, if selected (in terms of ), be enacted in the 3D (and n-dimensional) space of bodily and mental actions. Accordingly, because a counterfactual self-simulation could be instantiated in those spaces, and because it is evaluated in terms of a dissimilarity-relation, (as part of ), of how far it diverges from this instantiation, it extends in a distinct dimension of self-simulational counterfactuality, anchored on these spaces. Extension in a counterfactual dimension is however simply what an inference of the “width” of subjective temporality conceptually presupposes—hence the proposed mechanism. varies: for some policies, might be very small, with subjective time being barely extended at all (i.e. subjective temporal contraction). For other policies, might be larger, corresponding to a “wider” temporal Now (i.e. subjective temporal dilation). Furthermore, subjective temporality, as opposed to other inferences, such as the mental number line, is extended not only in a cognitive, or opaque, sense, but in an immediate, transparent, felt sense (Metzinger, 2003, Seth et al., 2012). Essentially, transparency is the default inferential format, being both phenomenologically evident and entailed by the computational cost opacification (i.e. the additional presence of inferences on a hierarchically higher level) engenders (Metzinger, 2003, Limanowski and Friston, 2018). Accordingly, it is precisely because introspective attention is not turned toward the process of construction of the subjective temporal Now that we simply live within it without realizing its inferential nature, with its extension varying systematically across our experience. Using the tools of information-geometry (e.g. Amari, 2016), the asymmetric KL divergence can, assuming infinitesimally small changes of the parameters, , be interpreted in terms of the symmetric Fisher information, , a Riemannian metric tensor of a statistical manifold, (i.e. a space of probability-distributions). This is possible because by measuring the divergence over an infinitesimally small change, , a second-order Taylor series expansion becomes sufficient for its characterization; and is the local Hessian matrix of the KL-divergence (Friston, 2019). After integrating over a path in , we obtain the path length, traversed by neuronal dynamics, in the statistical manifold, thus yielding a distinct information geometric formulation of subjective temporal extension (locally proportional to the divergence and globally distinct). We can take a derivative with respect to (external) time, thus deriving an instantaneous relative measure of subjective and objective time, . Technically, this is referred to as information rate: While the precise functional mapping between temporal extension and explicit duration judgment remains under-determined, the present proposal assumes that higher-level duration inference is informed by a lower-level temporal-extension signal. Locally, this signal is expressed as information rate, i.e. the rate at which information-geometric distance is traversed on the statistical manifold. Over an interval, this signal accumulates into information length. Explicit Bayesian duration inference is therefore cast as a hierarchically higher level that generates top-down predictions about expected cumulative information length. The Bayesian nature of this inference scheme directly enables an account of Vierordt’s law (the central-tendency effect): participants in temporal reproduction tasks overproduce shorter durations and underproduce longer durations (for reviews, see Lejeune and Wearden, 2009, Bon-Mi et al., 2011). In line with Bayesian frameworks of timing generally (e.g. Shi et al., 2013), this central-tendency effect can be explained by the influence of the (acquired) hierarchically higher level duration inference priors, with duration estimates being drawn toward previously inferred interval-lengths. For an example of a deep generative model equipped with duration inference states see Fig. 3. A probabilistic graphical model depicting an example of a deep parametric generative model with two hierarchical levels (perceptual and attentional). Each hierarchical level is equipped with duration inference states. A more extensive description of this kind of model-architecture can be found below. Adapted from a template given by (Sandved-Smith et al., 2021). It is important to remark that temporal extension itself, as formulated above, does not directly license an interpretation in terms of a chronometry (i.e. a direct quantification of the extension of a moment in terms of milliseconds or seconds): the self-simulational dissimilarity path length in the statistical manifold can be traversed with a varying information rate. This does not imply, however, that existing attempts to estimate the temporal interval spanned by “a moment” are empirically uninformative (see, e.g. Sauer et al., 2012, Gutierrez et al., 2019; see also Wittmann, 2011). Rather, if temporality-sensitive empirical measures are systematically shaped by self-simulational dissimilarity, then they may function as indirect empirical readouts of subjective temporal extension, even if they do not constitute direct measurements of it. Hierarchical active inference offers a parsimonious framework for duration inference across multiple interval lengths, because the same inferential architecture—including policy-simulation, evaluation, and self-simulational dissimilarities—can be realized across nested temporal scales. In this sense, duration-inference states need not be restricted to a single temporal domain, but can be defined at several hierarchical levels, with faster lower levels tracking shorter intervals and slower higher levels contextualizing longer temporal structure. This is consistent with deep temporal models in which evidence is accumulated over nested timescales and higher-level states entail sequences of transitions at the level below (Kiebel et al., 2008, Friston et al. 2017c; Pezzulo et al., 2018). Generally, empirically, aspects of self-modeling and temporal inferences covary too systematically for a coincidence to be likely (see Sections Introduction, A deep parametric generative model of temporal inference, and Variations of the ‘width’ of the subjective moment). SST explains this close covariation: subjective temporal extension emerges from self-modeling. Put metaphorically, the “width” of the subjective temporal moment is the “length of the path toward another version of yourself.” The observations that are expected if a sequence of actions were enacted, , are evaluated relative to the observations that are phenotypically preferred, , with subjective time experience emerging from their dissimilarity-relation implicated by the never-ending phenomenological pursuit of “that other me, where I would have finally reached my goal.” A deep parametric generative model of temporal inference The following simulation of a deep parametric generative model (Fig. 4) will serve to illustrate the computational dynamics of the “width” of subjective temporality, demonstrating how two “biases” (i.e. variations) of temporal inference emerge naturally from this minimal model. The model-architecture features two modifications from the deep generative model of mental action (Sandved-Smith et al., 2021), from which the use of the oddball-perceptual task (consisting of the presentation of either a “standard” or a “deviant” stimulus, with the former being presented 24/25 and the latter 1/25 times) has also been adopted. The generative model is composed of different hierarchical levels. The lowest level of the model is concerned with perceptual inference, specifically, with inferring whether the presented stimulus is a deviant (i.e. oddball) stimulus or not, being equipped with a state-transition matrix . As described above, the precision-weighting, , of the likelihood, , is understood in terms of attention (Feldman and Friston, 2010). In this generative model, the precision-weighting of the likelihood of the perceptual inference level is conditioned by hierarchically higher level attentional states, which are in turn conditioned by attentional policies (Sandved-Smith et al., 2021). This is in accord with work on attentional control as precision deployment (Brown et al., 2011, Kanai et al., 2015). Faster and more reliable perceptual inferences are induced by a higher precision-weighting. The hierarchically higher level attentional states are “focused” and “distracted.” These attentional states are, just as the perceptual states, inferred via variational inference and equipped with a state transition matrix, . The attentional actions on the second hierarchical level are either “stay” or “switch.” “Stay” leads the agent to remain in its current attentional state and “switch” flips the attentional state. The attentional actions are chosen from attentional policies (also comprising “stay” or “switch”) based on their expected free energy, . Actions are exclusively defined on this second hierarchical level, conditioning the state-transition matrix . The prior preferences over attentional states, , are set such that the agent prefers being in the state “focused.” This generative model was originally deployed as a generative model of mental action, modeling the dynamic attentional cycle in meditative practices (for a more detailed description see Sandved-Smith et al., 2021). The simulated deep parametric generative model. Policies are only defined on the second hierarchical level. The simulation of policies, and their evaluation in terms of expected free energy, is omitted at some time-steps (see the text). For the derivation of the precision belief update equation see Parr and Friston (2017). Adopted from a template given by (Sandved-Smith et al., 2021). This simulation was built using Python and the code is openly available. The original model architecture (Sandved-Smith et al., 2021) has been modified in two distinct ways. These changes concern (i) the sequentialization of policy-simulation, and (ii) the incorporation of resource-sensitive habitual control mechanisms. Specifically, while in the original model, the two distinct policies (”stay” and “switch”) were both simulated at each time-step, here, policy-simulation is sequentially unfolded in time, such that, at each time-step, only one policy is (at most) simulated. Consistent with previous extensions of active inference to resource-sensitive habitual control (Maisto et al., 2019), policy-simulation is entirely absent at some time-steps. Specifically, first, if, at the end of a simulation epoch (comprising the two policies “stay” and “switch” once, each), a policy is evaluated in terms of a sufficiently low (and thereby deemed sufficiently “good”), the next simulation epoch is skipped entirely, with the selected policy being enacted for the next two time-steps. Intuitively, “once a sufficiently good plan has been found, there is no need to reconsider for a while.” Second, if during each simulation epoch (the order of the two policies is alternated), the of the first simulated policy is sufficiently low (and, more specifically, also below the of both policies from the previous simulation epoch), the rest of the simulation epoch (i.e. the other policy) is skipped, with the first policy being immediately enacted. This, too, amounts to the implicit assumption that contingencies do not change significantly in the absence of policy-simulations (as deliberative control would be more advantageous otherwise (Maisto et al., 2019)), and can also be understood as the agent refraining from additional policy simulation, given that a sufficiently adaptive policy has already been found, thereby saving significant resources. This is consistent with current perspectives on the importance of resource-control for brain functioning (e.g. Barrett et al., 2016). From another perspective, this can be considered as simply another consequence of free-energy minimization: analogously to model-selection, non-necessary computations are avoided by the removal of excessive parameters (FitzGerald et al., 2014, Pezzulo et al., 2015, Friston et al. 2017a, Maisto et al., 2019). Furthermore, the decreased frequency of (sequential) policy simulation on the hierarchically higher level, with selected policies thereby being (on average) enacted for a longer time, mirrors empirical evidence on the increase of intrinsic neuronal timescales with the hierarchical depth of cortical processing (Kiebel et al., 2008, Murray et al., 2014, Friston et al. 2017c). Strikingly, given the minimality of this generative model, two “biases” (i.e. variations) of temporal inference naturally emerge from the simulation. Well known in cognitive psychology, the intentional binding effect (sometimes known as the temporal binding effect) refers to the compression of the temporal interval between voluntarily initiated actions and subsequent sensory consequences, often measured in terms of a direct interval estimation procedure (for a review see Moore and Obhi, 2012). This covariation of temporal inference and self-modeling has been observed so reliably that intentional binding has been proposed as an implicit measure for the sense of agency (Haggard et al. 2002a; b, Suzuki et al., 2019 for another perspective on intentional binding in terms of multisensory causal binding). As elaborated above, and due to the free-energy minimizing resource-sensitive habitual control mechanism, if a policy is evaluated in terms of a sufficiently low , the rest of the simulation epoch (comprising the (alternating) second policy, “stay” or “switch”) is skipped. mathematically constitutes, together with ambiguity, expected free energy, , in terms of which the simulated policies are evaluated. Accordingly, given the low (implying a contracted ) of the selected policy, and the absence of policy-simulation (and -evaluation) for the rest of the epoch, the temporal interval between voluntarily initiated actions and resulting sensory consequences is underestimated, explaining the intentional binding effect. The emergence of the intentional binding effect is demonstrated in Fig. 5 in terms of an increased number of “congruent” episodes—episodes where the simulation and selection of a policy occur at the same time-step. This congruence implies, on average, and because the simulated policy is “already good enough,” a relatively decreased and relatively fewer subsequent policy simulations. In contrast, “incongruent” episodes (here, a policy is not selected immediately upon its simulation) imply, because of the relatively increased , an increased , and more subsequent policy simulations. Simulation of a deep parametric generative model. A First hierarchical level: inference over perceptual states (standard/deviant). B Second hierarchical level: inference over attentional states (focused/distracted). C Subjective temporal extension (continuous line) and its absence (dashed line). The colored circles indicate simultaneity (congruency) or non-simultaneity (incongruency) between the simulation of a policy and its selection, if a policy is selected. The colored boxes indicate simulated policies (stay/switch/no policy-simulation) in the sequential simulation-epoch. Illustration inspired by Sandved-Smith et al. (2021). Code openly available at https://github.com/JanBellingrath/deep_parametric_generative_model_of_temporal_inference. This generative model enables a computational account of alternating episodes of concentration and distraction (Sandved-Smith et al., 2021), characteristic of mindfulness meditation: starting in a state in which the agent both is focused and infers itself to be focused, the agent, reflecting the intrusion of all-too-frequent mind-wandering-episodes (Schooler et al., 2011, Smallwood and Schooler, 2015), becomes distracted without inferring this very fact. Then, after a period of being distracted without having noticed, the agent infers that it is distracted, leading to a reallocation of attentional focus (Fig. 5). Mindfulness meditation has traditionally (e.g. Anālayo, 2003) and more recently been associated with attenuations of self-referential processing (Dahl et al., 2015, Tang et al., 2015, Milliere, 2020), being mirrored on the neuronal level of description in terms of a reduced activity of a network of regions implicated in mind-wandering (e.g. Brewer et al., 2011). From the perspective of active inference, the effects of meditation, and “selfless” experiences generally, have been understood in terms of a reduced frequency and a reduced counterfactual depth of policy-simulations (i.e. self-simulations) (Limanowski and Friston, 2020, Laukkonen and Slagter, 2021). In line with a rich tradition in eastern philosophy, alterations of temporal experience are attested in deep states of meditative absorption (Metzinger, 2024): a recent factor-analytic analysis of 1403 qualitative reports from meditators, directed at the experience of “pure awareness,” an entirely non-conceptual and contentless form of awareness as such (i.e. awareness of nothing but awareness), converged on a 12-factor model to describe the phenomenological character of these experiences in a fine-grained way (Gamma and Metzinger, 2021). The factor that explained most of the variance (Time, Effort, and Desire) integrated goal-directedness, effort, and temporal experience. Describing the phenomenological profile of this factor in meditative experience, Gamma and Metzinger (2021) write: “Typically, there will be an attentional lapse, followed by the phenomenology of noticing, remembering the goal state, and refocusing. As a result, temporal experience is preserved: for example, we find the phenomenology of duration and of time passing.” This factor is positively correlated with mind-wandering, memories, the simulation of future events, and the arising of thoughts and feelings generally (Factor 7: “Thoughts and Feelings”). This association of temporal experience with goal-directed counterfactual inferences is also evident in the generative model: the “width” of subjective temporality (i.e. ) varies characteristically, dependent on the counterfactual observations expected under a given policy, , evaluated in terms of systemic goal states, . Strikingly, evidence from these qualitative reports attests to alterations of temporal experience that culminate in phenomenal states that are described as atemporal (Metzinger, 2024). However, because these states nevertheless often feature experiences of change, it may be asked: “How can a physical system like the human brain create conscious models of reality that, on a conceptual level, seem to necessitate paradoxical descriptions like ‘timeless change’?” (Metzinger, 2024). As can be seen in Fig. 5, which depicts the dynamics of the generative model of temporal inference, perceptual states are inferred at each time-step, with the perceptual state changing over time. However, while the perceptual state may change, policies are not computed at each time-step. Crucially, because the existence of depends on the simulation of a policy (specifically, its evaluation in terms of ), is entirely absent if no policy is computed. In the generative model, furthermore, and consistent with the target phenomenon, the probability for policies to not be computed increases in the focused attentional state. As elaborated above, this reduced frequency of policy-simulation is consistent with active inference accounts of the effects of meditation and “selfless” experiences generally (Limanowski and Friston, 2020, Laukkonen and Slagter, 2021). Furthermore, this absence of deliberative control (i.e. the absence of policy-simulation), and the concomitant absence of a feeling of effort (Factor 1 integrated effort and time (Gamma and Metzinger, 2021)), is consistent with an active inference account of effort, with effort being constitutively dependent on the presence of policy-simulations (i.e. effort being absent in the absence of policy-simulations) (Parr et al., 2023). In actual neuronal systems, a drastic reduction in the frequency of policy-simulation over hierarchical levels might suffice for explaining the phenomenological reports (i.e. the binary presence/absence distinction in the generative model is an idealization). Accordingly, as the computation of counterfactual policies recedes more and more, the subjective temporal Now ceases to be constructed. A continuous flow of perceptual states unfolds in an effortless and atemporal phenomenal state: dynamic succession without temporal extension—the specious present (James, 1890) is not specious anymore. ”The early Dzogchen scholar–practitioners in Tibet knew all of this very well, but through their own meditation practice: “Nowness” is empty” (Metzinger, 2024) Variations of the “width” of the subjective moment Psychologically, much is known concerning systematic variations of subjective time perception. While probably everyone has experienced time slowing down in highly emotional moments, or speeding up during pleasurable episodes of absorption, numerous systematic variations of temporal inference have been described and experimentally tested in detail. This empirical evidence ranges from transient situational factors to stable dispositional constructs, including variations of temporal inference in diverse psychopathological conditions. Methodologically, and building on current computational (active inference) accounts of the various temporality-modulating psychological constructs, diverse branches of empirical evidence concerning variations of temporal inference will be explained by domain-specific modulations of the information-theoretic structure corresponding to the “width” of subjective temporality (for a summary, see Table 1). As mentioned above, varying reports on the felt passage of time and duration estimations are assumed to correspond to varying degrees of temporal extension, with a “faster” felt passage of time and smaller estimated durations corresponding to a shorter temporal extension, and vice versa. In this sense, subjective temporal extension is proposed as a unifying latent computational construct, which, while not having been empirically measured directly, is informed and constrained by systematic empirical evidence across diverse methods, states, and psychological dispositions. As mentioned in Section Subjective temporality as inference, active inference supplies a natural scaffold for self-simulational dissimilarities and duration inference states across varying hierarchical levels and associated timescales, thus accounting for time perception over varying temporal domains. Table 1Psychological effect classes computationally explained by SST Effect class Empirical finding Mechanism Agency (intentional binding) Contraction/underestimationThe comparatively low expected free energy of policies corresponding to subsequently enacted actions maps onto a low , thus producing the intentional binding effect. Meditation/minimal phenomenal experience “Atemporality”/ “timeless change”The reduced frequency of counterfactual policy simulation implies that is transiently not (or less) constructed; perceptual inference can still evolve over time, yielding dynamic succession without temporal extension (“timeless change”). Attention-to-time effects Dilation/overestimationPrecision as attention: attention to time is, by SST, simply attention to the (weak) first-person perspective, which, if outcomes under policies, , do not similarly contract, implies an increased temporal extension. Central tendency effects Central tendency effectsA direct consequence of Bayesian duration inference (i.e. the influence of the duration inference prior). States and dispositions: Negative valence/threat/pain Dilation/overestimationIncreased mismatch between outcomes expected under simulated policies and phenotypically preferred outcomes increases . Dispositional impulsivity (ADHD, borderline) Dilation/overestimationTime-discounted and/or narrowed prior preferences increase . Boredom Dilation/overestimationA comparatively reduced pragmatic value in a given situation mathematically implies an increased . Flow Contraction/underestimationThe expectation that actions fulfill prior preferences implies a comparatively small . Addiction/withdrawal Dilation/overestimationNarrowed, time-discounted preferences and frustrated goal-states increase the divergence . Psychological effect classes computationally explained by SST The extension of the subjective moment increases in negative experiences, with temporal durations being overestimated. Experimentally, for instance, temporal durations are overestimated when negative stimuli are expected (Ogden et al., 2015), fear-inducing videos are seen (Pollatos et al. 2014a), or participants recall a free-fall from a height of 31 m (Stetson et al., 2007). This association is so reliable that even the reverse inference can be empirically demonstrated: For (illusorily) shortened time-periods, an equally intense painful stimulus is reported to be decreased in magnitude (Maia et al., 2023). This emerging picture, in which the presence of negative valence covaries with temporal overestimations, has recently been meta-analytically supported (Xiaobing Cui et al., 2023). To give another example, participants overestimate time when their hand is immersed in water at painful, as opposed to neutral, temperatures, with increases in pain perception covarying with increases in temporal overestimation (Rey et al., 2017). Put in the words of the authors: “Being in pain reflects a deviation from homeostasis, which threatens integrity and prompts repeated conscious access to one’s self” (Rey et al., 2017). Whether it is a painful physical sensation, a situation constraining possible behavioral options to non-desired outcomes, or whether the constraints are largely internal and self-imposed, the dissimilarity between the outcomes expected under simulated policies (“what could come true, if a sequence of actions were performed”), , and the outcomes expected under prior preferences (“what you prefer to observe”), , increases. Indeed, negative states might be conceptually individuated by nothing but a certain difference between expectations of possible and preferred outcomes, holding across policy-simulations, and over a given episode, as, by definition for negative experiences, “you just cannot get what you want.” Essentially, if SST is correct, the “width” of the subjective temporal moment, , is itself nothing but this difference-relation. Accordingly, it is precisely this increased dissimilarity-relation between possible and preferred outcomes, characterizing negative influences, that explains the increased extension of the subjective moment in negatively connoted circumstances, with impressions of “moments feeling like eternities” being resultant. This is also consistent with a perspective offered by research on functional asymmetries in valence-processing, documenting a deeper processing for negative stimuli across various domains (Baumeister et al., 2001, Alves et al., 2017). Computationally, these processing asymmetries might be manifested in terms of an increased policy-simulation frequency and depth, necessitated because a policy evaluated in terms of sufficiently low is yet to be simulated, with an increased temporal “width” being resultant ( is constitutive of ). Consistently, time is perceived to be “slower” in depression (for a meta-analysis see Thönes and Oberfeld, 2015) and in advanced cancer patients (Van Laarhoven et al., 2011); and temporal overestimations occur in schizophrenia (for a meta-analysis, see Ueda et al., 2018) and PTSD (Vicario and Felmingham, 2018). While variations in valence span across many psychological domains in which temporal processing is altered, temporal inference is also, more specifically, modulated by impulsivity. This can be parsimoniously explained by virtue of a domain-specific variation of the information-theoretic structure corresponding to the “width” of subjective temporality under a current active inference account of impulsivity (Berk Mirza et al., 2019). First, empirically, a meta-analysis of temporal inference in attention-deficit hyperactivity disorder (ADHD), a condition closely associated with increased levels of impulsivity (e.g. Patros et al., 2015), has shown that affected individuals overestimate time (Zheng et al., 2022). Borderline personality disorder, too, being also coupled to heightened levels of impulsivity (e.g. Links et al., 1999), is related to a negatively felt expansion of time (Mioni et al., 2020) and temporal overestimations (Berlin and Rolls, 2004). Consistently, time is overestimated during abstinence in tobacco use disorder (Miglin et al., 2017), with time passing more slowly during withdrawal (Sayette et al., 2005). A similar trend has been shown in heroin users during the withdrawal period (Aleksandrov, 2005). Cocaine- and methamphetamine-dependent individuals, too, overestimate time to a degree that is correlated with their level of impulsivity (Wittmann et al., 2007) (for reviews on the relationship between time perception and impulsivity disorders, see Moreira et al., 2016, Paasche et al., 2019; see also Wittmann and Paulus, 2008). Computationally, one perspective on impulsivity is in terms of time-discounted prior-preferences (i.e. “selling the future for the present”) (for an active inference account, see Berk Mirza et al., 2019). Essentially, the steeper the discounting of the prior preferences, the stronger the preference for rewards in the present, discarding the future (for a review on smokers devaluing the future more than nonsmokers, for instance, see Barlow et al., 2017; for a more general review on temporal discounting see Story et al., 2014). Furthermore, the prior preferences may not only be more strongly focused on the temporal proximity of the rewards, but also on a specific subset of outcomes (e.g. being overly focused on obtaining a given addictive substance, relative to other rewards). Given this “narrowing” of the prior preferences, , is not mirrored (on average) by the outcomes expected under simulated policies, , the latter are (on average) more divergent from the former. Accordingly, the “width” of subjective temporality, , is, on average, increased for impulsive states and traits, with temporal overestimations being resultant, thus explaining the empirical evidence. Furthermore, the passage of time is judged to be slower in acute episodes of boredom (e.g. Droit-Volet et al., 2023, Danckert and Allman, 2005). For example, for participants shut in an empty room alone for 7.5 min, the experienced extent of boredom is associated with the feeling of time passing more slowly (Witowska et al., 2020). Computationally, boredom has been understood in terms of switching between exploitation and exploration: Tasks that are too easy provoke boredom, with boredom signifying the transition from pragmatic action to novelty seeking (Gomez-Ramirez and Costa, 2017, Danckert and Ros Velasco, 2019, Darling, 2023, Parvizi-Wayne et al., 2024). Mathematically, if the pragmatic value (i.e. ) currently afforded by a given situation decreases, (i.e. ) increases, explaining empirical evidence on an overly extended temporal moment in acute states of boredom. On the other hand, the flow-state (Csikszentmihalyi, 1990), characterized both by an increase in concentration on a task that is subjectively perceived as meaningful and challenging, and by a decrease of narrative self-modeling, has repeatedly been associated with a faster temporal experience (e.g. Im and Varma, 2018, Rutrecht et al., 2021). Participants even reverse infer having been in the flow-state, given that they are (falsely) informed about much time having passed (Christandl et al., 2018). Computationally, a recent active inference account explains the attenuation of narrative self-modeling in flow-experiences by virtue of an exclusive attentional focus mandated by the challenging nature of the task (Parvizi-Wayne et al., 2024). By this account, because the agent is, in the flow-state, because of experience and training, already confident about the consequences of actions, action-selection is primarily influenced by pragmatic (as opposed to epistemic) value. The flow-state is characterized further: “The individual has high prior expectations (i.e. preferences) about the states it expects to occupy and low uncertainty about how to fulfill those expectations” (Parvizi-Wayne et al., 2024). Because expecting to fulfill prior preferences mathematically implies a small (i.e. a low ), this explains the faster temporal experience empirically attested in the flow-state. Consider another well-known phenomenon: paying attention to time leads to longer estimations of durations, a classic result described as “the most robust, well-replicated finding in the time perception literature” (for a review see Brown et al., 2010). This effect is often studied in terms of dual-task interference tasks, wherein the additional presence of a non-timing task, concurrently performed with a timing task, leads to decreases of the estimated durations, interpreted in terms of a relatively decreased attention allocated to the timing task. Consistently, if the difficulty of the non-timing task increases, the magnitude of the interference effect increases (e.g. Zakay et al., 1983, Brown, 1985), and the interference effect reduces if the participants are trained in either one of the two tasks (e.g. Brown and Bennett, 2002). A review on the family of dual-task interference effects (i.e. smaller estimated durations, higher variability & higher inaccuracy) found that 91% out of 77 studies conducted between 1924 and 2008 showed this interference effect (Brown and Grondin, 2008, Brown, 1997). Paying increased attention to time entails, under the usual mapping of precision to attention, and the perspective offered by SST, a transiently increased precision of prior preferences: if temporal extension is a consequence of self-modeling, anchored on the first person perspective (see Section Subjective temporality as inference), attention to time simply is attention to the first person perspective (for a generalized active inference model of mental actions, see Sandved-Smith et al., 2021). If, however, the precision of prior preferences increases without a corresponding contraction of the outcomes expected under counterfactual policies, increases. Thus, SST explains why increased duration estimates ensue when attention is being paid to time (for a review on attentional effects on time see Brown and Grondin, 2008); for the classic attentional-gate model, often referenced as an explanation in this context, see (Zakay and Block, 1997). Furthermore, there is an interesting connection between temporal extension and cognitive effort, as defined in (Parr et al., 2023). In this—active inference—setting, cognitive effort is associated with the divergence between prior beliefs about policies based upon expected free energy and those based upon habits (i.e. based upon and , respectively). On this view, temporal extension and cognitive effort are related via their common dependence upon expected free energy and, implicitly, information length. However, this relationship depends upon habitual priors, namely, how habitual or automatic policy selection has become. Put in other words, temporal extension and cognitive effort are jointly implicated in a less habitual computational regime, wherein policies are simulated, and evaluated in terms of (for a recent perspective on the connection between effort, perceived time, and interoception, see Wolff et al., 2025). As a last empirical constraint, consider near-death experiences. In near-death experiences, such as car-crashes or other accidents, the subjective temporal Now is reported to be extended to such a degree that time subjectively “stands still” (Russell and Kletti, 1976). Indeed, how could the predicted outcomes of a simulated policy be more divergent from prior preferences (with the subjective temporal Now, , extending accordingly) than the expected outcomes of a policy where one is predicted to be dead, or potentially so? Conclusion Nothing in the physics of time corresponds to an extended moment in the subjective sense. This paper has described how the “width” of subjective temporality emerges as a transparent and counterfactual inference from the dynamics of self-modeling, cast in terms of free energy minimization. Intuitively, because a given counterfactual self-simulation could, if selected, be instantiated in the 3D (and n-dimensional) space of bodily (and mental) actions, and because it is associated with a dissimilarity-relation, of how far it diverges from this instantiation, it extends in a distinct dimension of self-simulational counterfactuality, anchored on these spaces. But this transparent counterfactual extension simply is what an inference of subjective temporal “width” conceptually presupposes—hence the proposed mechanism. Technically, this self-simulational dissimilarity relation was expressed as an actively inferred temporal extension, and cast in the parlance of information geometry, wherein unfolds on a statistical manifold or belief space, the intrinsic metric of which induces precisely the information length that subtends temporal extension. Following a simulation of a deep parametric generative model illustrating the emergence of two empirically attested alterations of temporal inference (i.e. the intentional binding effect and alterations of temporal experience in deep meditative states), numerous systematic variations of temporal experience and biases of temporal inference have been explained via the proposed mathematical structure. This concerns classic psychological effects such as attentional and central-tendency effects, as well as variations of subjective temporal experience across varying levels of valence, boredom, impulsivity, flow-states, near-death experiences, and various psychopathologies, among others. While there are also many psychological effects and features of temporal experience not addressed by the current proposal, the explanatory strategy has been the same across all the addressed constructs: Demonstrating that current active inference accounts mapping onto the respective constructs mathematically imply systematic variations of precisely that information-theoretic structure that is—by SST—identical to subjective temporal “width” in precisely the direction that the psychological evidence respectively indicates. Accordingly, here, temporal extension functions as a latent computational construct, the dynamics of which, while not being empirically measured directly, are informed and constrained by virtue of duration estimations and passage-of-time judgments. In this sense, these temporality-sensitive empirical measures are assumed to instantiate a systematic mapping to temporal extension (for interesting examples of alternative, and potentially more direct, temporality-sensitive empirical measures, see, e.g. Sauer et al., 2012; Gutiérrez et al., 2022). Currently there is little work exploring how, or if, variations in temporal extension, suitably operationalized, systematically covary with duration estimates and passage-of-time judgments across psychological states and traits, with a fruitful avenue of future work being the characterization of their interrelation. The information-geometric formalism proposed here does not, by itself, yield a direct chronometric quantity in milliseconds or seconds. Rather, it characterizes the information-geometric distance traversed by trajectories of probabilistic states, i.e. an informational geometry of self-simulational dynamics. In the absence of a fully specified bridge from these quantities to explicit chronometric judgments, the present account is best understood as supporting comparative predictions about temporal dilation and contraction across psychological states and dispositions, rather than absolute chronometric readouts. This systematic covariation between the dynamics of self-modeling and temporal inference is consistent with neuroimaging meta-analyses revealing both self-modeling and temporal inference as centrally implicated in insula functioning (for a meta-analysis on neuroimaging studies of self-modeling, see e.g. Qin et al., 2020; for meta-analyses on neuroimaging studies of time perception, see (Mondok and Wiener, 2023, Naghibi et al., 2023)). Furthermore, the proposed self-simulational dissimilarity-relation is epistemically parsimonious, not only because the computation of a dissimilarity-relation between “what could” and “what ought” is psychologically all-prevalent but also because the corresponding mathematical structure is implied by the minimization of expected free energy. Concerning states of minimal phenomenal experience (i.e. awareness of nothing but awareness) (Metzinger, 2024), entirely atemporal (without a subjective temporal Now) and non-dual (without a phenomenal self-model), SST describes how atemporality emerges as a consequence of non-duality, thus potentially paving a way toward a deeper understanding of minimal phenomenal experience itself. ”I measure myself as I measure time” (Heidegger, 1992) Acknowledgements I am grateful to Daniel Friedman, Karl Friston, Thomas Metzinger, Lars Sandved-Smith, Marc Wittmann, and members of the Minimal Phenomenal Experience Project for their constructive comments and feedback. Author contributions Jan Erik Bellingrath (Conceptualization [lead], Formal analysis [lead], Investigation [lead], Methodology [lead], Project administration [lead], Resources [lead], Software [lead], Visualization [lead], Writing—original draft [lead]) Conflict of interest None declared. Funding This study received support from the Open Access Publication Funds of the CerCo (Université de Toulouse). Data availability The code is open-source and can be accessed at https://github.com/JanBellingrath/deep_parametric_generative_model_of_temporal_inference. References Associated Data Data Availability Statement The code is open-source and can be accessed at https://github.com/JanBellingrath/deep_parametric_generative_model_of_temporal_inference.