The temporal structure of interoceptive inference concerns how the brain anticipates, tracks, and updates bodily states across time, rather than merely reacting to them as they occur. In the framework often described as the bayesian brain, interoceptive systems maintain generative models that predict the evolution of internal states—such as heart rate, blood pressure, gut motility, and blood glucose—over multiple time scales. These models do not just encode what the current bodily state is likely to be; they also encode when certain interoceptive events are expected to occur, and how quickly deviations should be corrected to maintain homeostasis and support allostasis. The result is a temporally extended inference process where past experiences and expected future demands shape present physiological regulation.
At the core of this temporal organization are priors that specify not only the most probable values of bodily variables but also their expected trajectories and rhythms. For instance, circadian and ultradian cycles, respiratory patterns, and cardiac cycles all provide reliable temporal scaffolds that the brain can learn and exploit. When the heart is expected to accelerate with physical exertion and decelerate during rest, the generative model incorporates these regularities as structured expectations about change over time. Interoception, in this sense, is not limited to snapshot-like sensing; it is a continuous Bayesian updating process where moment-to-moment afferent signals are compared against temporally specific predictions that integrate context, past history, and anticipated events.
The temporal dimension becomes especially salient when considering delays and lags in physiological systems. Many regulatory actions—such as the release of hormones, changes in vascular tone, or adjustments to immune responses—unfold on time scales that are much slower than the neural computations that initiate them. To be effective, the generative model must account for these delays by embedding temporal kernels that capture how causes (for example, nutrient intake) lead to consequences (for example, changes in blood glucose) after characteristic time intervals. The brain thus learns causal-temporal mappings between actions, environmental conditions, and ensuing interoceptive outcomes, allowing it to issue predictions that are temporally aligned with expected bodily responses.
Anticipatory control mechanisms illustrate how temporal structure is critical for efficient regulation. Before a meal is fully digested, the body already begins to prepare for an influx of nutrients through cephalic-phase responses. Similarly, heart rate and blood pressure often increase in advance of physically demanding actions, driven by predictions about impending energetic needs rather than by current metabolic deficits. This anticipatory engagement depends on temporally precise priors: the brain must infer not only what will happen internally but also when it will happen, to minimize interoceptive prediction errors and avoid costly deviations from homeostatic set points. Temporal precision here is not absolute; the generative model encodes probability distributions over onset times and durations, and these distributions are continually refined through experience.
Different time scales of interoceptive inference can be distinguished conceptually and mechanistically. Fast time scales involve millisecond-to-second adjustments such as baroreflexes and respiratory gating, where prediction and error-correction loops operate almost continuously. Intermediate time scales, spanning seconds to hours, govern processes like gastric emptying, thermoregulation, and acute stress responses, where predictive mechanisms must integrate contextual cues and recent history. Long time scales, from days to years, encompass circadian and seasonal changes, developmental trajectories, and long-term metabolic regulation. At each scale, the generative model incorporates temporal priors about typical patterns, volatility, and stability, enabling the organism to predict not merely instantaneous bodily states but evolving trajectories under varying conditions.
Temporal discounting and horizon length are also integral to interoceptive inference. The brain must decide how far into the future to project its predictions when regulating internal states. A short temporal horizon may be adequate for acute reflex-like responses but insufficient for planning around extended energetic demands, such as endurance exercise or fasting. Conversely, an excessively long or inflexible horizon could lead to maladaptive over-preparation, consuming resources in anticipation of events that may not materialize. The generative model therefore includes hierarchically organized temporal layers, where higher levels encode slow-changing contexts and long-term expectations, while lower levels track rapid fluctuations. Interoceptive signals ascend this hierarchy, constraining and being constrained by temporally structured predictions at each tier.
Accuracy in interoceptive prediction depends not only on the shape of temporal priors but also on their precision, that is, the degree of confidence the brain assigns to them relative to incoming signals. Precision weighting determines how strongly prediction errors at different time lags influence belief updating. For example, the system may treat short-latency deviations in heart rate as noise during vigorous exercise, while assigning higher weight to sustained deviations over longer periods. By modulating the precision of temporally specific priors, the brain can flexibly adapt to contexts where rapid fluctuations are expected versus contexts where stability is the norm. This dynamic precision control allows the organism to filter out irrelevant variability while remaining sensitive to meaningful shifts in internal state.
Temporal structure is also reflected in the way interoceptive signals are sampled and integrated. Many bodily signals are rhythmic or quasi-rhythmic, such as heartbeat and respiration, and neural systems appear to synchronize with these rhythms. Oscillatory coupling between interoceptive afferents and cortical processes can provide a temporal reference frame that segments continuous bodily input into discrete windows for prediction and error evaluation. For instance, the phase of the cardiac cycle has been shown to modulate perceptual and affective processing, suggesting that the brain leverages repeated temporal landmarks to organize both interoceptive and exteroceptive inference. In this way, the periodicity of bodily signals becomes an internal clock that structures the flow of predictions and updates.
Contextual cues from the external environment further inform the temporal organization of interoceptive inference. Environmental regularities—such as time-of-day, social routines, and habitual activity patterns—serve as anchors for anticipating internal demands. The generative model learns that certain external events reliably precede particular interoceptive states, like fatigue toward the end of a workday or increased arousal in social settings. Over time, these learned temporal contingencies allow the brain to schedule regulatory actions so they align with expected bodily needs. This coupling of exteroceptive and interoceptive timelines supports allostasis, where prediction of future demands guides proactive adjustments of internal parameters long before deviations from homeostatic ranges occur.
Crucially, the temporal structure of interoceptive inference is not static. It adapts as the organism encounters new environments, routines, and physiological conditions. Learning mechanisms reshape temporal priors when external schedules change, when illness alters the latency and magnitude of bodily responses, or when lifestyle modifications transform patterns of sleep, diet, and activity. Such plasticity ensures that the generative model remains calibrated to the actual temporal dynamics of the body in its current ecological niche. When the mapping between time, context, and internal state is accurately represented, interoceptive prediction can pre-empt large errors with minimal effort, enabling efficient and flexible maintenance of homeostasis across ever-changing temporal landscapes.
Predictive coding and homeostatic regulation
Within the predictive coding framework, homeostatic regulation is conceived as the continuous minimization of interoceptive prediction errors. Rather than merely reacting to deviations from set points, the brain issues top-down predictions about expected internal states and compares them with bottom-up afferent signals. When there is a mismatch, prediction errors arise that can be resolved either by updating beliefs about the body or by engaging autonomic, endocrine, and behavioral effectors to change the body itself. Homeostasis, and more broadly allostasis, thus emerges as a consequence of the brain acting to reduce the discrepancy between predicted and actual physiological conditions over time.
In this view, the generative model encodes not only preferred ranges for bodily variables but also the dynamics by which these variables fluctuate around their targets. For example, a model may specify that blood pressure should increase transiently when standing up and then decay back to baseline with a characteristic time constant. The brain predicts this trajectory in advance and adjusts sympathetic outflow accordingly. When the observed interoceptive trajectory deviates from this expectation—either rising too slowly, overshooting, or failing to stabilize—prediction errors signal that the current model of vascular control is inaccurate or that corrective actions must be intensified. Regulation thus depends on priors about plausible temporal evolutions of bodily states, not just on static set points.
A central feature of predictive coding is its hierarchical organization, which maps naturally onto different levels and time scales of bodily regulation. At lower levels, fast reflex arcs and local circuits implement simple predictive loops, such as baroreflex adjustments or chemoreflex control of respiration. These circuits encode highly precise short-term predictions about rapid fluctuations and correct small, frequent errors with minimal delay. Higher levels of the hierarchy represent more abstract and slower-varying aspects of internal milieu, such as metabolic reserves, circadian phase, or chronic inflammatory tone. They generate coarse-grained predictions over longer horizons and constrain lower-level dynamics by specifying what counts as a tolerable pattern of fluctuation within broader homeostatic objectives.
Homeostatic set points themselves can be understood as preferred hidden states in the generative model. However, in a predictive coding scheme, these set points are not fixed values but distributions reflecting acceptable ranges and their context dependence. For instance, an optimal blood glucose level is different during rest, intense exercise, or prolonged fasting. Higher-level systems encode contextual priors about which range is currently desirable, while lower levels implement predictions and controls to keep glucose within that context-specific band. Prediction errors that persist over extended periods drive learning at higher levels, gradually reshaping priors so that what is expected comes to reflect what is statistically typical for that individual in their environment.
Allostasis refines this picture by emphasizing that the brain is often engaged in predicting future demands rather than merely correcting present deviations. Predictive coding supports allostasis by allowing generative models to forecast impending changes in energetic, thermal, or osmotic requirements and mobilize resources beforehand. For example, anticipatory increases in heart rate and ventilation before starting a familiar exercise routine reflect higher-level predictions about pending metabolic load. These anticipatory signals cascade down the hierarchy, setting new short-term set points for lower-level regulators. When exercise begins, the actual interoceptive input closely matches the predicted state, resulting in minimal error and efficient regulation.
The role of precision weighting is crucial in understanding how predictive coding governs homeostasis. Precision reflects the confidence assigned to predictions or to sensory inputs and effectively controls their influence on belief updating and action. In volatile or noisy environments, the system may increase the precision of certain interoceptive priors, dampening the impact of transient fluctuations that are expected and uninformative. Conversely, when the internal milieu becomes unpredictable—such as during acute illness—the system may down-weight prior expectations and increase the gain on incoming signals to detect and respond to genuine perturbations. Dysregulation of precision, rather than of the predictions themselves, can thus lead to maladaptive homeostatic responses, including exaggerated stress reactions or blunted awareness of bodily threat.
Predictive coding also clarifies how behavioral actions participate in homeostatic control. From this perspective, actions are selected to realize predicted interoceptive states that minimize expected future error. Eating, drinking, resting, seeking warmth, or engaging in social contact can all be construed as policies chosen because their predicted bodily consequences align with higher-level homeostatic and allostatic goals. The brain evaluates alternative actions by simulating their likely interoceptive and exteroceptive outcomes within the generative model. Policies that are expected to keep bodily variables within preferred ranges over relevant time scales are favored, while those predicted to produce large or prolonged errors are suppressed. Regulation therefore extends beyond autonomic reflexes to encompass complex, goal-directed behaviors embedded in rich temporal contexts.
Importantly, the generative model incorporates causal structure linking external events, internal states, and regulatory actions. It learns that certain cues reliably precede specific bodily changes, and uses these contingencies to generate timely predictions. For instance, recurrent experience teaches that exposure to cold air leads to a delayed drop in skin and core temperature, which in turn activates shivering and vasoconstriction. The predictive model comes to anticipate this sequence, initiating countermeasures before the drop becomes pronounced. Homeostatic regulation in this setting is inherently predictive: the system uses learned causal-temporal relationships to keep internal variables within bounds with minimal delay, minimizing the metabolic and functional costs of large corrections.
The interaction between interoception and exteroception is integral to predictive homeostatic control. External sensory cues often provide earlier and more reliable information about upcoming challenges than interoceptive feedback alone. By aligning interoceptive predictions with exteroceptive context—for example, linking the sight and smell of food with future nutrient inflow, or associating workplace cues with upcoming cognitive and social demands—the system can orchestrate preparatory adjustments in endocrine, immune, and autonomic systems. This cross-modal integration effectively extends the temporal horizon of homeostatic regulation, allowing the organism to maintain internal stability in anticipation of, rather than in reaction to, environmental changes.
On slower time scales, predictive coding supports the maintenance of long-term bodily integrity by integrating history and trends in interoceptive signals. The generative model can track gradual drifts in baseline levels, such as progressive weight gain, changes in blood pressure, or shifts in sleep architecture, and treat these as evidence that prior expectations about the body’s typical state are no longer accurate. Persistent prediction errors at these scales drive structural updates of priors, altering what the system considers normal. In healthy adaptation, these adjustments track reversible changes—such as seasonal variations in activity or temporary training effects—while preserving an overarching stability that protects vital functions. When these longer-term inferences fail, homeostatic systems can become locked into maladaptive equilibria, such as chronically elevated stress or altered metabolic set points.
The predictive coding perspective implies that homeostatic failures often reflect inference problems rather than purely mechanical defects of the body. Distorted or overly rigid priors about interoceptive states, maladaptive precision assignments, or impaired hierarchical coordination can all lead to persistent prediction errors and compensatory regulatory patterns that are themselves harmful. Understanding homeostasis through the lens of the bayesian brain and predictive coding thus foregrounds the role of temporal priors, learning, and belief updating in shaping how organisms sustain internal balance across continually unfolding bodily and environmental changes.
Learning temporal priors from bodily signals
Temporal priors over bodily states are not innately fixed; they are constructed and continually reshaped through experience with the body’s own dynamics. From the perspective of the bayesian brain, learning these priors involves estimating the statistical regularities that govern how interoceptive signals unfold in time and how they relate to actions and environmental cues. Each heartbeat, breath, gastric contraction, or fluctuation in blood glucose provides data from which the system can infer characteristic latencies, durations, and patterns of variability. Over repeated exposures, these samples allow the organism to build an internal catalogue of temporal contingencies—what typically happens, in what order, and over what time scales—forming the scaffold for anticipatory control and allostasis.
One key source of information for learning temporal priors is the intrinsic rhythmicity of many bodily processes. Cardiac and respiratory cycles, gastric slow waves, and endocrine pulses offer relatively stable periodic structures that the nervous system can entrain to. Through repeated synchronization between neural activity and these bodily rhythms, the system estimates not only mean cycle lengths but also how much phase jitter and amplitude variability to expect. For example, through exposure to varied activity levels, the organism learns that heart rate accelerates following certain motor commands and decelerates with rest, with typical rise and decay constants. These learned temporal profiles allow the system to anticipate transient accelerations or decelerations without treating them as surprising, thereby preventing unnecessary corrective responses.
Learning is not confined to periodic processes; it also applies to event-related dynamics, such as the time course of postprandial glucose changes or the evolution of stress hormone levels following a threat. Each episode of eating, exertion, or psychosocial stress becomes an experiment in which the brain compares predicted trajectories of interoceptive variables with observed outcomes. Discrepancies—prediction errors—are used to adjust expectations about onset times, peak magnitudes, and recovery rates. If, for instance, a new dietary pattern consistently shifts the latency and amplitude of glucose spikes, the generative model gradually incorporates these changes, refining its priors about how long after food intake a metabolic response will occur and how extended the ensuing return to baseline is likely to be.
Crucially, temporal priors are learned in a context-sensitive fashion. The same interoceptive event can unfold with different timing depending on factors such as time of day, posture, ambient temperature, or emotional state. Through hierarchical learning, higher levels of the generative model come to represent these contextual moderators, encoding that, for example, sympathetic arousal during social performance ramps up earlier and recovers more slowly than arousal during solitary exercise of similar intensity. Over time, the system infers conditional priors: not just that a given bodily change will occur, but that its temporal profile depends systematically on exteroceptive circumstances, internal motivational states, and prior history of similar episodes.
Experience-dependent plasticity in the neural circuits subserving interoception underlies these learning processes. Structures such as the insula, anterior cingulate cortex, and brainstem nuclei are repeatedly exposed to coupled patterns of ascending interoceptive signals and descending predictions. Hebbian and error-driven learning mechanisms adjust synaptic strengths so that neural ensembles come to encode expected sequences and delays. For instance, insular representations of cardiac and respiratory signals can become tuned not only to instantaneous states but also to specific phase relationships and cross-correlations, capturing how one physiological variable typically leads or lags another. This tuning effectively embeds temporal kernels within the interoceptive model, assigning probability mass to particular delays and transitions.
On longer time scales, learning temporal priors requires integrating information over extended histories of bodily regulation. Sleep-wake cycles, seasonal changes in activity, and developmental shifts in metabolism are all discovered through long-term tracking of repeating patterns. The system aggregates interoceptive data across days, weeks, and years, inferring slow regularities such as circadian phase relationships between core temperature, cortisol levels, and subjective alertness. When lifestyle or environment changes—for example, with shift work, jet lag, or chronic illness—the resulting persistent prediction errors drive revisions of these higher-level temporal priors. The organism gradually recalibrates its sense of when it is appropriate to feel hungry, sleepy, or fatigued, updating expectations about the timing of homeostatic needs.
Learning temporal priors also entails estimating the volatility of bodily processes—the likelihood that their temporal characteristics will change over time. In a relatively stable environment with consistent routines, the system can justifiably assign high precision to its temporal expectations, leading to tight priors about when certain interoceptive states should emerge and subside. By contrast, in contexts where routines or bodily conditions fluctuate widely, the model must accommodate higher uncertainty, maintaining broader priors over onset times and durations. Adaptive learning therefore involves not only inferring mean temporal parameters but also tracking higher-order statistics that reflect how rapidly those parameters themselves drift.
Actions and policies play an active role in this learning. When the organism selects behaviors—such as eating at a new time of day, altering exercise routines, or experimenting with relaxation practices—it effectively probes the temporal response properties of its own body. Each policy generates predicted interoceptive trajectories, and the observed outcomes refine beliefs about which temporal patterns are controllable, how action-dependent they are, and what delays typically separate action from interoceptive consequence. Over repeated cycles of policy selection and evaluation, the generative model becomes better equipped to simulate future bodily trajectories under different behavioral strategies, supporting more precise planning of when to initiate regulatory actions to pre-empt deviations from homeostatic ranges.
Exteroceptive cues serve as powerful training signals for temporal interoceptive priors. Visual, auditory, and social signals often precede bodily changes by reliable intervals, forming predictive chains that can be learned. The sight and smell of food, time cues associated with work and rest, or social signals of impending conflict provide early indicators of future internal demands. By tracking the typical lags between such cues and ensuing interoceptive responses, the system learns to align predictions about internal states with the temporal structure of the external world. This cross-modal temporal binding extends the predictive horizon of interoception, allowing the organism to initiate regulatory adjustments before interoceptive deviations even begin.
Development offers a particularly informative window into how temporal priors from bodily signals are acquired. Infants are born with immature regulatory systems and limited prior exposure to their own physiological rhythms. Early caregiving environments introduce regularities—feeding schedules, sleep routines, soothing practices—that impose structure on interoceptive experiences. Repeated cycles of need, caregiver response, and subsequent bodily recovery shape the infant’s expectations about how quickly distress is resolved, how long satiety lasts, and when fatigue is likely to arise. These early-formed temporal priors can have lasting consequences, influencing later patterns of stress responsivity, energy regulation, and subjective interoception in adulthood.
The learning of temporal priors is inherently bidirectional: as priors develop, they shape the sampling of bodily signals, which in turn constrains further learning. When the generative model expects certain patterns to occur at specific times, it can modulate attention to interoceptive channels accordingly, enhancing sensitivity around critical periods and down-regulating it when little change is expected. This selective sampling creates an efficient but biased learning regime, in which the organism becomes especially adept at predicting well-practiced temporal patterns but may remain relatively insensitive to rare or unexpected ones. Only when unexpected interoceptive events generate large, persistent prediction errors does the system reallocate attention and update temporal priors to capture the new structure.
Learning temporal priors from bodily signals is constrained by energetic and computational costs. Tracking fine-grained temporal detail across multiple interoceptive modalities and time scales is resource intensive. The bayesian brain therefore compresses temporal information, learning abstracted representations such as typical phase relationships, coarse time windows, and canonical event sequences. These compressed priors are sufficiently rich to support effective homeostasis and allostasis in familiar contexts, while remaining flexible enough to be retuned when environmental or bodily conditions change. The result is a continuously evolving set of temporally structured expectations that allow the organism to anticipate, rather than merely react to, the unfolding of its own physiological states.
Disruptions of temporal priors in psychopathology
Disruptions of temporal priors in psychopathology can be understood as failures of the bayesian brain to correctly learn, maintain, or deploy expectations about how bodily states unfold in time. When the temporal structure of interoception is misrepresented, prediction errors become chronically elevated or misallocated across time scales, and the systems that implement homeostasis and allostasis are forced into inefficient or maladaptive patterns of regulation. These disturbances may not manifest as obvious mechanical defects of the body, but rather as dysregulated inference about when physiological changes should begin, peak, and resolve. Many forms of psychopathology can therefore be reframed as disorders of temporal interoceptive inference, in which priors about bodily trajectories are either too rigid, too volatile, or incorrectly coupled to exteroceptive context.
Anxiety disorders offer a paradigmatic case. Here, priors about the timing and duration of autonomic arousal tend to exaggerate the likelihood of imminent bodily threat and underestimate the expected rate of recovery. Interoceptive sensations such as palpitations, breathlessness, or gastric unease are predicted to escalate quickly and to persist, even when previous episodes have resolved without harm. This temporal pessimism biases prediction toward catastrophic trajectories, increasing the precision assigned to early warning signals and shortening the temporal horizon within which safety is inferred. As a result, minor fluctuations that would otherwise be treated as transient are interpreted as the onset of prolonged dysregulation, amplifying anxiety and promoting compensatory behaviors such as avoidance and safety seeking that prevent corrective updating of these maladaptive priors.
Panic disorder illustrates how distortions of temporal precision can become self-reinforcing. Individuals often hold highly precise priors that bodily sensations will reach peak intensity rapidly and with little warning. The predicted latency between initial cues (for example, a slight change in heart rate) and full-blown panic is shortened, and the anticipated recovery window is extended, leading to an overestimation of both immediacy and duration of threat. In predictive terms, the system expects a steep trajectory of escalating interoceptive disturbance that will not quickly resolve, which magnifies the impact of any early deviations. This compressed temporal window for threat detection undermines the ability to accumulate disconfirming evidence over time; panic becomes more likely, and each episode further entrenches the belief that rapid, uncontrollable bodily spirals are the norm.
In generalized anxiety and worry, temporal priors tend to stretch far into the future, with an enlarged anticipation window and difficulty in letting predictions decay once a putative threat has passed. The generative model overestimates the persistence of bodily and situational dangers, holding the interoceptive system in a state of sustained readiness. Allostasis is chronically engaged in preparing for anticipated stressors that seldom materialize or that, when they do occur, are less severe than expected. This long-horizon threat prediction prolongs sympathetic activation and blunts the normal transition back to parasympathetic dominance, contributing to fatigue, somatic tension, and sleep disruption. Because the organism rarely experiences extended periods of genuine physiological safety, there is little opportunity to revise priors about how quickly and reliably the body can return to baseline.
Depressive disorders often involve the opposite problem: temporal flattening and pessimistic expectations about change. Priors may encode that positive bodily and affective shifts are slow to emerge, fragile, and short-lived, whereas negative states are expected to be protracted and recurrent. From a temporal inference perspective, the system assigns higher prior probability to trajectories in which low energy, anhedonia, and fatigue persist over long intervals and are only weakly influenced by contextual improvements. This skewed temporal structure reduces the expected value of initiating regulatory or reward-seeking actions, because the generative model predicts that their bodily consequences will be either delayed or minimal. In turn, diminished behavioral engagement curtails opportunities to experience rapid or robust interoceptive improvements, providing little evidence to challenge pessimistic priors about the sluggishness of recovery.
Circadian and ultradian dysregulation in depression further indicates a disruption of temporal scaffolds that normally support interoception. When sleep-wake cycles, hormonal rhythms, and temperature profiles lose coherence or shift phase, higher-level priors about daily bodily trajectories become less reliable. The system may respond by assigning lower precision to circadian cues and higher precision to momentary interoceptive fluctuations, making the internal milieu feel erratic and unanchored in time. This temporal disorganization can manifest subjectively as a loss of diurnal structure to mood and energy, and clinically as irregular sleep, appetite disturbances, and variable psychomotor activity. Without stable temporal anchors, the generative model struggles to forecast when relief or activation should be expected, perpetuating a sense of interminable dysphoria.
Somatic symptom and functional disorders highlight another form of temporal disruption: miscalibrated priors about the typical time course of bodily responses to minor perturbations or injuries. In these conditions, interoceptive sensations that would usually be transient, such as localized pain or gastrointestinal discomfort, are predicted to be more prolonged and more tightly coupled to exteroceptive cues that have lost their original biological relevance. The model may encode that certain activities or environmental triggers almost invariably lead to long-lasting or escalating symptoms, even when objective physiological evidence does not support such trajectories. Because patients often alter their behavior to avoid putative triggers or to rest extensively after mild symptoms, they generate experiential data that are consistent with slow recovery, reinforcing priors that minor bodily deviations are precursors to extended dysfunction.
In chronic pain, the temporal contour of nociceptive expectation is particularly distorted. The generative model becomes biased toward predictions of persistence: once pain is present, it is expected to remain or worsen rather than to subside. At the same time, the delayed beneficial effects of activity, rehabilitation, or analgesic practices are underweighted, whereas short-term exacerbations are given high precision. This leads to an inference landscape in which any early increase in pain following movement is taken as confirmation that further activity will extend or intensify suffering over long time scales. The resulting avoidance reduces exposure to trajectories in which short-term discomfort is followed by longer-term relief or functional gain, limiting the corrective evidence needed to broaden temporal priors and restore more adaptive expectations about the dynamics of pain.
Post-traumatic stress disorder (PTSD) reveals how acute violations of temporal predictions can have enduring consequences. Traumatic events are often characterized by sudden, overwhelming bodily changes—extreme arousal, freezing, dissociation—that violate prior expectations about the upper bounds and time course of interoceptive states. The generative model may respond by expanding the expected volatility of internal states and encoding that such catastrophic shifts can occur with little or no warning. Temporal priors about the latency between threat cues and bodily upheaval become shortened, while priors about recovery stretch out, implying that once the body enters a high-arousal state, it will be slow to return to equilibrium. Intrusive memories and flashbacks can be seen as maladaptive replay of earlier interoceptive trajectories, keeping alive the expectation that similar sequences could recur at any moment.
In PTSD, contextual modulation of temporal priors is also impaired. Normally, exteroceptive cues indicating safety should lengthen the inferred time to possible threat and compress the expected duration of arousal episodes. After trauma, safety cues lose their regulatory power; the model may generalize threat across contexts and times, treating neutral stimuli as if they were temporally proximal to danger. This generalization collapses distinctions between past and present, so that bodily responses appropriate to the original event are redeployed in response to remote triggers. The sense that time has stopped or that the trauma is endlessly recurring may reflect a deep failure of hierarchical temporal inference, in which the brain cannot confidently assign the traumatic interoceptive sequence to a bounded segment of the past.
Obsessive–compulsive disorder (OCD) provides another window into disrupted temporal prediction. Compulsions such as repeated checking or washing can be interpreted as attempts to manually reset temporal priors about threat resolution and bodily safety. The generative model doubts the durability of safety: it predicts that contamination, harm, or incompleteness will either re-emerge quickly or may not have been resolved in the first place. Interoceptive signals that would normally mark the end of a regulatory episode—such as relief after washing or checking—are either not registered as sufficiently precise evidence or are predicted to decay rapidly. This short-lived confidence in safety drives repeated actions to refresh the sense of resolution, keeping the system locked in short temporal loops that prevent the consolidation of longer-term priors about lasting security.
Eating disorders involve complex alterations of interoceptive timing, particularly around hunger, satiety, and reward. Restrictive patterns and binge–purge cycles distort the usual contingencies between food intake, gastric distension, metabolic shifts, and affective responses. In restrictive disorders, priors may encode that hunger sensations can be indefinitely tolerated without serious consequences, and that the aversive interoceptive and emotional responses associated with eating will be prolonged. As a result, the system expects delayed or attenuated relief after food consumption and overestimates the persistence of guilt or discomfort, making avoidance more likely. In binge–purge presentations, the temporal coupling between consumption and anticipated relief or punishment is compressed and intensified: ingestion is predicted to rapidly produce both immediate reward and rapid onset of negative interoceptive states that can be alleviated only through compensatory behaviors, narrowing the window for more measured regulatory responses.
Substance use disorders further demonstrate the remodeling of temporal priors around interoceptive and reward signals. Repeated exposure to fast-acting pharmacological effects teaches the generative model that desired bodily and affective states can be induced quickly and reliably by specific actions, often with less delay than natural rewards. The expected onset of relief or euphoria shifts earlier, while the anticipated duration of positive states becomes increasingly tied to continued use. At the same time, priors about the time course of withdrawal and craving may become exaggerated, predicting early, intense, and long-lasting discomfort if use is discontinued. This skewed temporal calculus makes alternative regulatory strategies—which typically offer slower and more moderate benefits—seem comparatively ineffective, reinforcing the selection of short-latency, high-intensity policies despite their long-term physiological costs.
Psychotic disorders highlight disruptions not only in the content of interoceptive inference but also in the temporal coordination between interoception and exteroception. Aberrant precision may be assigned to spontaneous bodily fluctuations, which are then misattributed to external agents or retrocausality-like influences. The timing of internal events is no longer coherently bound to environmental cues or to plausible causal sequences. For example, a benign change in heart rate may be inferred to precede and cause an external event, or to be controlled by distant actors, violating usual temporal orderings. This breakdown of temporal alignment undermines the generative model’s ability to maintain a stable narrative linking bodily states with events in the world, fostering delusional explanations and hallucinated control over internal sensations.
Autism and related neurodevelopmental conditions may involve atypical development of temporal priors across both interoceptive and exteroceptive domains. Some accounts suggest that priors about the stability and predictability of bodily states are under-specified or assigned low precision, leading the system to rely heavily on moment-to-moment sensory data without strong expectations about longer-term trajectories. This can make internal states feel abrupt, surprising, or difficult to anticipate, especially in contexts with rapid sensory change. Alternatively, in certain individuals, priors may be overly rigid, with narrow expectations about timing that are easily violated, resulting in strong prediction errors and distress in response to minor deviations from routine. In either case, the calibration of temporal expectations about bodily and environmental regularities is compromised, complicating the coordination of internal regulation with external demands.
Across these diverse conditions, a recurring theme is the misallocation of precision to temporal priors versus sensory evidence. When priors about the timing and persistence of interoceptive states are granted excessive precision, the system becomes inflexible, discounting real-world evidence that bodily changes are briefer, less intense, or more context-dependent than expected. Conversely, when temporal priors are too imprecise, interoception becomes dominated by noisy, short-term fluctuations, making bodily experience feel chaotic and undermining confidence in the body’s long-range predictability. In both extremes, homeostasis and allostasis suffer: regulatory actions are launched either too early, too late, for too long, or not long enough, as the organism struggles to align internal dynamics with an inaccurately inferred temporal landscape.
These disruptions can be traced to multiple levels of the hierarchical generative model. At lower levels, local circuits may misestimate the time constants of specific reflexes and autonomic loops, leading to overcorrections, undershoots, or oscillatory instabilities. At intermediate levels, representations of typical episode structure—for example, the temporal profile of a stress response—can be biased toward rapid escalation or incomplete recovery. At higher levels, contextual signals that should modulate temporal priors may lose their influence, or global beliefs about the volatility and controllability of bodily states may become pessimistically skewed. The resulting coordination failures cascade down the hierarchy, manifesting as persistent prediction errors that are experienced phenomenologically as anxiety, fatigue, pain, dysphoria, or a pervasive sense that one’s body is temporally out of sync with the environment.
Implications for computational and clinical models
Viewing interoceptive regulation through the lens of temporal priors has direct implications for how computational models are constructed and how clinical interventions are designed. In computational terms, models must move beyond static representations of bodily set points to explicitly encode trajectories, delays, and hierarchical time scales. In clinical terms, many symptoms can be reinterpreted as manifestations of miscalibrated temporal expectations about internal states, suggesting that treatment should aim not only to change what people believe about their bodies, but also when they expect bodily changes to occur, how long they think they will last, and how quickly recovery should follow.
For computational neuroscience, this perspective encourages the explicit formulation of generative models in which time is a core latent dimension. Rather than representing interoception as a mapping from hidden causes to instantaneous sensory outcomes, models should specify differential equations or state-space dynamics that capture predicted trajectories under different contexts and actions. This entails encoding temporal kernels—parameterized by latencies, rise and decay constants, and periodicities—that express priors about how bodily variables evolve. Model inversion then becomes a problem of recovering not only current states but also their most likely futures, given past trajectories and control policies. Such formulations can naturally be implemented in active inference or predictive coding schemes, where temporal structure is encoded in the transition matrices and precision schedules of the generative model.
Incorporating hierarchical time scales is particularly important. Computational models should reflect at least three interacting layers: fast controllers (for example, baroreflex, respiratory feedback) that operate on sub-second to second scales; mid-range controllers (for example, thermoregulatory and stress systems) that operate over minutes to hours; and slow contextual layers (for example, circadian phase, developmental stage, chronic inflammatory tone) that evolve over days to years. Each level carries its own temporal priors and precision assignments, constraining the others. Simulations in which perturbations are applied at different levels can reveal how local changes in time constants or precision propagate through the hierarchy to produce the prolonged dysregulation characteristic of clinical conditions.
Another modeling implication concerns volatility and uncertainty about temporal structure. Computational accounts should allow priors over time constants and periodicities themselves to be updated, rather than treating them as fixed. This can be achieved by placing higher-order priors over parameters that govern the speed and regularity of bodily dynamics, and allowing these parameters to drift in response to persistent prediction errors. Such meta-learning of temporal structure is crucial for explaining adaptation to new environments (for example, shift work) or to chronic illness, as well as maladaptive plasticity in disorders where the system overestimates the volatility or persistence of certain bodily states.
Bridging computational work with physiology requires models that can be quantitatively linked to measurable signals such as heart rate variability, skin conductance, hormone levels, or inflammatory markers. By fitting temporally explicit generative models to longitudinal physiological data, one can infer latent temporal priors and precision parameters that characterize individual differences in interoceptive regulation. This approach opens the possibility of computational phenotyping, where patients are classified not solely by symptoms but by the inferred structure of their temporal interoceptive models: for example, short anticipatory horizons with high precision in anxiety, flattened recovery priors in depression, or exaggerated volatility estimates in PTSD.
These computational insights translate into several concrete directions for clinical assessment. First, diagnostic procedures could more systematically probe temporal expectations about bodily events. Instead of merely asking whether a symptom is present, clinicians can inquire about perceived onset, escalation, and recovery: How quickly do people expect palpitations or pain to worsen? How long do they anticipate that fatigue or sadness will last? Do they believe that rest or coping strategies will produce rapid or delayed benefits? Such questions can reveal the shape of patients’ temporal priors and may uncover distortions that are not evident from symptom checklists alone.
Second, behavioral tasks can be designed to experimentally measure temporal prediction and learning in interoception. For example, paradigms could manipulate the delay between an exteroceptive cue and a mild interoceptive perturbation (such as a brief respiratory load or heartbeat feedback), tracking how quickly participants learn the cue–body timing and how they generalize across contexts. Variants in which the timing becomes more or less predictable can assess sensitivity to volatility and the plasticity of temporal priors. Errors in predicting when bodily sensations will begin, peak, or resolve may serve as quantitative biomarkers for disorders characterized by particular temporal distortions.
Neuroimaging and electrophysiological methods can be leveraged to identify neural signatures of temporal interoceptive inference. Time-resolved analyses of insular, anterior cingulate, and brainstem activity can test whether these regions encode predictions about future bodily states or merely track current inputs. Measures of cross-frequency coupling or phase–amplitude relationships between interoceptive rhythms (such as heartbeat or respiration) and cortical oscillations may reveal how temporal landmarks from the body are used as reference frames for prediction and error correction. Aberrant coupling could indicate disrupted temporal scaffolds, helping to link computational parameters (such as priors over phase relationships) with observable neural dynamics.
On the intervention side, conceptualizing psychopathology as involving disrupted temporal priors suggests that treatment can be targeted at recalibrating the timing of interoceptive expectations. Many existing therapies may already be acting on these priors implicitly, and computational models can help to make these mechanisms explicit and optimizable. For instance, exposure-based interventions for anxiety and panic often involve repeated, prolonged contact with feared bodily sensations until they naturally subside. Viewed through a temporal-prior lens, such protocols are designed to lengthen the predicted delay to catastrophe and shorten the expected duration of arousal, teaching the system that spikes of discomfort are self-limiting and that escalation is neither as rapid nor as enduring as previously believed.
Similarly, behavioral activation and graded activity programs in depression, chronic pain, and fatigue syndromes can be understood as experiments in updating pessimistic priors about sluggish or absent recovery. By systematically scheduling actions and tracking their interoceptive consequences over time, patients accumulate evidence that positive bodily shifts can occur more quickly and last longer than expected. Therapists can explicitly frame these exercises as tests of temporal predictions about homeostasis and allostasis: the goal is not merely to “feel better” in an undifferentiated way, but to revise beliefs about how fast and how reliably the body can return toward equilibrium following stress, exertion, or emotional challenge.
Drug treatments also have temporal dimensions that can be incorporated into computational and clinical models. Pharmacotherapies alter not only the intensity of interoceptive signals but also their time course—onset, half-life, and rebound. Patients often develop expectations about how rapidly relief will occur, how long it should last, and when side effects may arise. Miscalibrated medication-related priors can contribute to non-adherence (for example, stopping a treatment because improvement is slower than expected) or to overreliance on fast-acting substances. By modeling and communicating the temporal profiles of medications, clinicians can help align patients’ priors with pharmacokinetic and pharmacodynamic realities, reducing prediction errors that might otherwise be misinterpreted as treatment failure.
Training methods that directly target interoceptive awareness—such as mindfulness, biofeedback, and slow-breathing practices—may owe part of their efficacy to reshaping temporal priors. Mindfulness exercises that involve sustained, nonjudgmental attention to fluctuating bodily sensations can increase the resolution with which temporal patterns are perceived, revealing that many sensations wax and wane more quickly than previously noticed. This richer sampling of interoception provides data that can support broader, less catastrophic priors about the duration and volatility of internal states. Biofeedback tools can go further by making hidden dynamics (for example, heart rate variability, respiratory sinus arrhythmia) explicitly visible, allowing individuals to experiment with how particular actions or mental states alter trajectories in real time, and thus update their generative models about controllability and delay.
Designing interventions around temporal prediction also opens possibilities for “timing-based prescriptions.” For example, in insomnia, protocols can specifically target priors about the onset of sleep and the expected time to return to sleep after awakenings. In eating disorders, structured meal plans can be framed as opportunities to relearn the timing of hunger, fullness, and affective responses to food. In chronic pain, time-contingent rather than pain-contingent activity schedules can teach that movement-related discomfort is transient and followed by stable or improved functioning, thereby adjusting priors about the short- and long-term consequences of exertion.
From a systems perspective, temporal-prior models suggest that clinical care should attend to the coherence of multiple bodily rhythms rather than isolated symptoms. Restoring circadian regularity, for example through light therapy or structured daily routines, may re-establish reliable temporal scaffolds for interoception, thereby improving the predictability of mood, appetite, and energy. Interventions that synchronize sleep, feeding, social interaction, and physical activity can help rebuild higher-level priors about when different physiological states should occur, which in turn can stabilize lower-level regulatory loops. Computational models that simulate how circadian and ultradian processes constrain faster autonomic and affective dynamics can guide the design of multi-modal interventions that act synergistically across time scales.
Personalization is a further implication. If individual differences in psychopathology partly reflect differences in temporal priors and precision allocations, then treatment can be tailored by estimating these parameters for each person. One patient with panic might primarily overestimate the speed of escalation, while another might chiefly overestimate the duration of episodes. Their therapeutic exercises would then differ: the first might focus on repeatedly observing the gradual build-up of arousal, the second on documenting relatively rapid recovery. Computational tools that fit generative models to each person’s physiological and self-report data can provide clinicians with individualized maps of temporal distortions, informing nuanced, mechanism-based treatment planning.
The temporal-prior framework has implications for prevention. Early developmental environments that provide consistent, contingent responses to bodily needs are likely to foster accurate and flexible temporal expectations about relief and regulation. Public health and caregiving interventions can therefore be evaluated not only in terms of immediate symptom reduction but also in terms of how they support or disrupt the acquisition of reliable temporal priors about interoception. By promoting regular routines, predictable caregiving, and constructive experiences with recovery from stress, it may be possible to build foundational generative models that confer resilience against later perturbations, reducing vulnerability to disorders in which the timing of bodily change is persistently misjudged.
