The brain’s predictive capacities unfold along nested temporal hierarchies, in which faster and slower processes interact continuously to generate coherent perception and action. Within the framework of predictive processing, the nervous system is understood as a layered architecture of generative models, each level operating at a characteristic timescale. Lower levels encode rapidly changing sensory details, such as the flicker of light on the retina or the fine-grained acoustic features of speech, while higher levels capture slowly varying, more abstract structure, such as object identity, situational context, or social norms. These layers are not isolated; they exchange information through recurrent message passing, so that predictions descend from higher to lower levels and prediction errors ascend from lower to higher ones. Crucially, this exchange is temporally organized: the speed at which each level updates its expectations is tuned to the degree of volatility it typically encounters in the environment.
Temporal hierarchies emerge because environmental regularities themselves are stratified in time. Some aspects of the world are relatively stable, like the fact that walls tend not to move rapidly, while others fluctuate on intermediate scales, such as the position of a moving object, and still others change very quickly, like the shimmer of specular reflections. A hierarchical system can minimize overall prediction error more efficiently by assigning slowly changing causes to higher levels and more ephemeral causes to lower levels. Under predictive processing, this arrangement amounts to a form of structured bayesian updating spread across time: each level maintains priors about the likely causes operating on its own timescale, and these priors are revised in light of incoming prediction errors that have already been partially explained away at lower levels.
At the fastest end of the temporal hierarchy are processes that track moment-to-moment sensory fluctuations. For example, neurons in early visual cortex respond on the order of tens of milliseconds to changes in contrast or orientation. Their generative models specify how simple features are expected to evolve over very short intervals, such as the continuous motion of an edge across the retina. Because the environment at this granularity is highly dynamic, low-level priors must be relatively flexible, allowing rapid adjustments in response to prediction errors. This high update rate enables the system to swiftly accommodate unexpected changes, like a sudden flash of light or a brief occlusion. However, these quick updates are constrained by the expectations passed down from slower, higher levels that encode more enduring aspects of the scene.
Intermediate levels of the hierarchy operate over longer windows, integrating information over hundreds of milliseconds to seconds. In auditory perception, these levels might encode phonemes, syllables, and words, each unfolding at a characteristic rate. In motor control, they may represent submovements, gestures, or action chunks. The generative models at these levels track how sequences of fast sensory events cluster into meaningful units, using predictions about temporal structure—such as rhythm, prosody, and habitual movement patterns—to interpret noisy or incomplete inputs. Because the relevant causes change more slowly than raw sensory signals, the priors here are more stable, and updating proceeds at a correspondingly slower pace. Prediction errors at these intermediate timescales signal not just transient noise but potentially significant deviations from expected temporal patterns, prompting recalibration of the model when discrepancies persist.
At the top of the temporal hierarchy lie the slowest, most abstract generative models, which span seconds, minutes, or even longer. These levels encode contextual features such as the current task, the social situation, or the broader environmental setting, and they may underwrite enduring expectations about one’s body, identity, and interpersonal world. For example, when entering a quiet library, high-level priors set the expectation of low ambient noise and subdued movement, shaping the interpretation of every subsequent sensory input. Because the causes represented here are assumed to be highly stable, their priors are correspondingly resistant to change, and updating unfolds over long timescales. Yet when prediction errors remain large and systematic—such as discovering that the “library” is actually being used as a venue for a noisy event—these higher-level models can undergo substantial revision, reorganizing how lower-level processes interpret the same sensory flux.
These nested temporal layers support a form of dynamic neural inference that is intrinsically multi-timescale. Each level maintains beliefs about hidden causes at its preferred temporal resolution while also modeling how causes at slower timescales modulate those at faster ones. For instance, the expectation that a conversation is occurring in a crowded café can influence how intermediate auditory levels parse ambiguous sounds as speech rather than random noise, while fast sensory levels track the micro-variations in pitch and intensity. Prediction errors flow upward when observed dynamics diverge from these expectations, but the rate at which beliefs are adjusted depends on how volatile each level judges the world to be. Temporal hierarchies thus naturally implement something akin to two-time updating: rapid corrections for transient discrepancies at lower levels and more gradual accommodation of persistent discrepancies at higher levels.
The organization of temporal hierarchies also hinges on precision weighting, namely the assignment of confidence to predictions and prediction errors at each timescale. Precision reflects the estimated reliability of signals: highly reliable prediction errors are weighted strongly in updating, while unreliable ones are discounted. Typically, the system treats fast sensory fluctuations as relatively noisy and high-level contextual information as more stable, but this pattern is not fixed. Under conditions of sudden change, lower levels may temporarily increase the precision of their prediction errors, forcing faster updates at intermediate and higher levels. Conversely, when the environment is judged to be stable, high-level priors exert stronger influence, effectively smoothing over transient anomalies. The dynamic regulation of precision across temporal hierarchies allows the brain to flexibly navigate between rapid responsiveness and conservative stability.
Temporal stratification is also evident in neural oscillations and their coupling across frequency bands. Fast oscillations, such as gamma rhythms, are associated with local processing and rapid feature binding, while slower rhythms, such as theta and alpha, support integration over longer intervals and coordination across distant brain areas. Cross-frequency coupling provides a mechanistic substrate for temporal hierarchies in predictive processing: slow oscillations can gate the timing and excitability of faster ones, aligning windows of sensory sampling with higher-level expectations. In this way, slow dynamics at higher levels set the temporal stage for fast, bottom-up evidence accumulation, and the interplay between different oscillatory bands mirrors the interaction between multiple timescales of prediction and error correction embedded in the hierarchical architecture.
Temporal hierarchies play a central role in explaining how perception can be both stable and flexible. Stability arises because higher levels, with their slow-changing generative models and strong priors, provide continuity across time, anchoring our experience despite rapidly shifting sensory inputs. Flexibility arises because lower levels can still react quickly to prediction errors, allowing sudden changes in the environment to trigger immediate adjustments in perception and action. When the environment persistently contradicts high-level expectations, the resulting accumulation of prediction errors forces a reconfiguration of those expectations, gradually revising the temporal hierarchy itself. This interplay between stability and flexibility is intrinsic to hierarchical predictive processing and is essential for coping with the layered temporal structure of the world.
Dual timescales of neural updating
To understand two-time updating in the brain, it is useful to distinguish between two complementary kinds of temporal dynamics in neural inference. On one hand, there are rapid, trial-by-trial adjustments that occur whenever a prediction error is registered; on the other, there are slower, cumulative revisions of the assumptions that govern how quickly those errors should matter. The predictive processing framework casts these as distinct but interacting timescales of bayesian updating: a fast timescale that tunes specific predictions about the current input, and a slow timescale that tunes hyperparameters such as volatility estimates and precision weights. This is what gives neural systems the capacity to be simultaneously responsive to momentary surprises and conservative with respect to more enduring beliefs about the world.
At the fast timescale, updating primarily concerns the current state estimates of hidden causes: where an object is located right now, which phoneme is being heard, whether a tactile stimulus indicates a harmless touch or a threat. In formal terms, these are updates to the posterior beliefs at a given hierarchical level, driven by incoming prediction errors from the level below. The relevant learning rates or step sizes are high, so beliefs can shift substantially within tens to hundreds of milliseconds. For example, in a noisy room, the auditory system must constantly adjust which frequencies belong to a speaker’s voice from one moment to the next; prediction errors about mismatched spectral content prompt swift corrections in the inferred speech stream without overhauling the broader model of the conversation or the social setting.
The slower timescale, by contrast, governs changes in the parameters that shape how these fast inferences unfold. Rather than adjusting specific expectations about an ongoing event, slow updating revises beliefs about the statistics of the environment: how volatile it is, how noisy certain sensory channels are, and how trustworthy high-level contextual information should be. In predictive processing terms, this largely concerns the dynamics of precision weighting. When the environment is inferred to be stable, precision on high-level priors is increased and precision on fast sensory prediction errors is reduced, producing conservative responses to fleeting anomalies. When sustained discrepancies accumulate, the system gradually infers higher volatility, lowering the effective precision of entrenched priors and increasing the gain on incoming error signals. This recalibration tends to unfold over multiple events or exposures, rather than within a single perceptual moment.
Two-time updating therefore hinges on a division between “state” and “meta-state” inference. Fast neural updates track the state of the world at the timescale of perception and action, while slower updates infer properties of the environment that shape how those fast inferences should evolve. A simple illustration comes from learning a new speaker’s accent. Initially, individual words may be misheard, prompting quick, local corrections as prediction errors arise with each syllable. Over repeated interactions, however, the brain gradually alters its generative model of the mapping from acoustic features to phonemes, effectively changing the parameters of the speech model itself. That second, slower process—reparameterizing the generative model—modifies the future landscape of prediction errors and thus the character of fast updating.
Neurally, these dual timescales can be implemented through differences in synaptic dynamics and circuit organization. Fast timescale updating may be realized through transient changes in neuronal firing rates and short-term synaptic plasticity, which allow populations to track prediction errors over tens to hundreds of milliseconds. Slow updates, in contrast, are associated with longer-lasting synaptic modifications and neuromodulatory influences that can adjust excitability and gain over seconds to minutes and beyond. For instance, tonic levels of neuromodulators such as norepinephrine and acetylcholine can globally regulate precision weighting by altering the signal-to-noise ratio in sensory cortices, effectively tuning how strongly prediction errors drive fast belief revisions. Thus, the same circuitry can support two-time prediction by combining quickly reversible dynamics with slower, more persistent changes in connectivity and gain control.
Dual timescales are also manifest in how prediction errors are integrated over time. Single violations of expectation at the fast timescale do not necessarily trigger an immediate revision of slow priors. Instead, the brain appears to accumulate evidence about environmental change, effectively averaging prediction errors across time. Only when a pattern of mismatches persists does the slower updating process become engaged. In a volatile environment, this accumulation window shrinks: fewer consistent errors are needed to infer that the world has changed, leading to faster adjustments at the slow timescale. In a stable environment, the window expands, and the system demands many repeated discrepancies before revising high-level beliefs. This adaptive integration window is central to how two-time updating balances sensitivity to genuine shifts against robustness to random fluctuations.
Another key component of two-time updating involves top-down modulation of lower-level dynamics. Slow-timescale beliefs about context and volatility do not merely sit on top of fast processes; they actively shape them by setting expectations about which prediction errors should be heeded. For example, when the brain infers that sensory input is likely to be unreliable—such as in foggy visual conditions—precision on low-level visual errors is reduced, and the system leans more heavily on stable priors from higher levels. This effectively slows down fast updating in that modality. Conversely, when a sudden, salient change is detected—such as a loud unexpected sound in a quiet room—slow-timescale mechanisms can transiently boost the precision of sensory errors, momentarily accelerating fast updates across multiple levels. In this way, the slow system continuously regulates the tempo of the fast system.
These dual timescales of neural updating have a direct bearing on how stability and flexibility are achieved in perception and action. Fast updating provides flexibility: it ensures that transient events are tracked with high temporal resolution and that motor responses can be rapidly adapted to immediate demands. Slow updating provides stability: it prevents the system from overreacting to every fluctuation, instead reserving substantial model revisions for changes that are consistent and enduring. The apparent smoothness of conscious experience reflects this negotiated balance. While moment-to-moment contents of consciousness may shift quickly in response to fast prediction errors, the underlying sense of a stable world and enduring self is anchored by slow-changing priors that are only gradually reshaped by long-run patterns of discrepancy.
Two-time updating can thus be understood as a division of labor within a single predictive architecture. Rapid error-correction cycles continuously refine detailed estimates of current states, while slower cycles recalibrate the rules by which such refinements occur. Formal bayesian updating occurs at both levels, but the relevant variables and time horizons differ: fast updates concern the here-and-now configuration of hidden causes, whereas slow updates concern the evolving assumptions about noise, volatility, and structure that govern future perception and learning. This layered temporal organization enables the brain to treat each experience both as an occasion for immediate adjustment and as a data point in a much longer process of model revision.
Computational models of two-time prediction
Computational accounts of two-time prediction typically formalize neural inference as bayesian updating in hierarchical generative models with explicitly separated timescales. In these models, the brain is treated as estimating both rapidly changing latent states and more slowly changing parameters or hyperparameters that govern those states. Fast variables encode the immediate configuration of hidden causes—such as current object positions, phonemic categories, or limb trajectories—while slow variables represent contextual factors, volatilities, and precisions that evolve on longer horizons. Two-time updating emerges when the system simultaneously infers both sets of variables, but with different learning rates and temporal integration windows, thereby mirroring the distinction between rapid perceptual adjustments and gradual belief revision.
A canonical formulation uses state-space models in which hidden states evolve according to stochastic differential equations, and observations are generated from these states through probabilistic mappings. The key computational move for two-time prediction is to add a second layer of dynamics that governs how the parameters of these equations themselves drift or switch over time. For example, in a volatile environment, the variance of state transitions may increase, signaling that predictions should be less confident and more responsive to incoming data. Estimating both the current state and the current volatility requires two coupled updating processes: a fast one that adjusts the state estimates on each time step, and a slower one that revises beliefs about volatility based on longer-term patterns of prediction error.
Hierarchical Gaussian filters (HGFs) offer a concrete example of such models. In HGFs, each level of the hierarchy encodes a latent variable that evolves as a Gaussian random walk, with higher levels modulating the step size or variance of lower levels. The lowest level tracks the current cause of observations; the next level tracks the tendency or bias of those causes; higher levels track how quickly those tendencies themselves change. Two-time updating is implemented by assigning different learning rates to each level, derived from their inferred uncertainties. Fast learning at lower levels allows rapid correction of momentary mismatches between predictions and sensory input, while slower learning at higher levels adjusts beliefs about the stability or volatility of the environment. This architecture reproduces behavior in tasks where agents must infer both immediate contingencies and changing rules, such as probabilistic reversal learning.
Another family of models emphasizes the role of precision weighting in two-time prediction. In these schemes, prediction errors at each level are multiplied by a dynamically estimated precision, which acts as a gain factor controlling how strongly they influence belief updates. Slow variables encode expectations about precision—essentially, priors over the reliability of different sensory channels or contextual cues—while fast variables encode current state estimates. Two-time updating arises because precision expectations change more slowly than state estimates: a single surprising observation might temporarily shift fast beliefs about what is happening now, but only sustained discrepancies will lead the model to revise its assumptions about how noisy or volatile the world is. This distinction between rapid state updates and slower precision updates underlies computational explanations of how agents distinguish random noise from genuine environmental change.
Predictive coding implementations of predictive processing provide a neurally plausible scheme for two-time prediction. In predictive coding networks, each layer contains separate populations of “representation” units that encode expectations about hidden causes and “error” units that encode mismatches between expectations and input. Continuous message passing between these units yields approximate bayesian updating in real time. To capture dual timescales, modelers introduce additional dynamics in which synaptic weights, gain parameters, or the variances associated with prediction errors evolve more slowly than the activity of representation units. Fast activity dynamics correspond to rapid adjustment of current state estimates, while slow synaptic or gain dynamics correspond to gradual changes in model structure and precision weighting. This separation enables the same network to both track immediate stimuli and adapt its internal model to persistent changes in environmental statistics.
Bayesian reinforcement learning models also embody two-time updating by distinguishing between fast value estimation and slow adaptation of learning rates or policies. For instance, in meta-learning frameworks, an agent updates expected values of actions on each trial according to a prediction error rule, but simultaneously learns, on a slower timescale, how large those updates should be in different contexts. The fast process corresponds to adjusting expectations about immediate rewards; the slow process corresponds to inferring environmental volatility or task structure and tuning the learning rate accordingly. When cast within hierarchical generative models, this structure aligns closely with two-time predictive processing: value signals at lower levels are rapidly updated, while higher levels adjust the volatility priors that determine how quickly those value expectations should change.
Computational treatments of two-time updating often deploy switching or mixture models to capture abrupt contextual changes. In hidden Markov models (HMMs) and switching linear dynamical systems (SLDS), a discrete latent variable encodes which regime or context is currently operative, while continuous variables encode the state within that context. Two-time prediction appears as the coexistence of rapid within-context state estimation and slower, evidence-accumulating inference about context switches. The system maintains a posterior over possible regimes, updating it gradually as evidence accrues; only when this posterior crosses a threshold does the model “decide” that a new context has begun. This approach accounts for human tendencies to resist reinterpreting a situation after a few anomalous events but to reframe it when anomalies become consistent and prolonged.
Approximate inference methods are central to making two-time predictive models computationally tractable. Variational inference, in particular, provides a way to derive coupled update equations for fast and slow variables by minimizing a single free energy or variational objective. Under mean-field assumptions, the joint posterior over states and parameters factors into separate distributions, each with its own dynamics. The update rule for fast states is driven chiefly by current prediction errors and their precisions; the update rule for slow parameters is driven by longer-term statistics of those errors, such as their variance and autocorrelation. Because parameters influence the shape of the future error landscape, slow updating effectively reshapes the terrain on which fast inference operates, implementing two-time adaptation within a unified optimization framework.
Some models take inspiration from control theory and adaptive filtering, such as the Kalman filter and its adaptive variants. In a standard Kalman filter, there is already a separation between state estimation and covariance estimation, but in adaptive filters this separation becomes dynamic: the process and observation noise covariances are themselves estimated online. Two-time updating appears as a fast loop that produces state estimates on every time step and a slower loop that adjusts noise covariances based on accumulated discrepancies between predicted and observed errors. These adaptive filters have been used to model human tracking behavior, where participants follow a moving target whose motion statistics may change unpredictably. Fast updates keep the cursor aligned with the target, while slow updates allow participants to adjust how much they trust recent movements versus past history.
Computational models that aim to link two-time prediction to neural data often incorporate biophysical constraints. For example, models may assign fast inference to recurrent interactions among cortical columns, operating on tens-of-milliseconds timescales, while assigning slow inference to synaptic plasticity and neuromodulatory control, unfolding over seconds to minutes. In such schemes, prediction error units integrate inputs on short time windows but their gain is modulated by slowly varying neuromodulatory signals encoding precision expectations. Simulations show that this arrangement can reproduce key signatures of hierarchical predictive coding, such as transient error bursts when stimuli deviate from predictions, followed by gradual adaptation of baseline activity and connectivity when deviations persist. This provides a mechanistic bridge between abstract bayesian formulations of two-time updating and concrete neurophysiological processes.
Some theorists have explored whether two-time updating has implications for how time itself is represented in the brain’s generative models. Certain formulations introduce explicit temporal basis functions or kernels at multiple scales, allowing the system to predict future inputs by linearly combining contributions from different temporal windows. Fast kernels capture immediate dynamics, while slow kernels track long-range dependencies and contextual drifts. Two-time prediction then consists in jointly inferring the coefficients on these different kernels, with rapid adjustments for short-horizon predictions and slower reweighting of long-horizon components. This approach can approximate complex temporal structures without committing to a single fixed timescale, and it aligns with empirical findings of neurons and networks tuned to distinct temporal receptive fields.
Computational accounts of two-time updating also intersect with models of memory. In dual-store or complementary learning systems, fast-learning mechanisms encode specific episodes quickly, while slow-learning mechanisms integrate information across episodes to build more abstract knowledge. When recast in predictive processing terms, episodic traces support rapid correction of current predictions, whereas semantic structures correspond to slowly changing priors that shape expectations across similar situations. Two-time prediction emerges from the interaction between these systems: recent experiences can temporarily bias fast inference, yet only consistent patterns across many episodes will significantly shift entrenched semantic priors. This framework can explain why immediate surprises can alter momentary perception without instantly restructuring deep-seated beliefs, and why extended exposure is needed to transform local learning into durable conceptual change.
Across these different computational approaches, the unifying theme is that two-time updating is not an add-on to bayesian updating but a natural consequence of inferring both states and the conditions under which those states evolve. By building models in which some variables change rapidly and others only slowly in response to enduring patterns of prediction error, theorists capture how a single predictive architecture can maintain stable high-level priors while remaining agile at the level of moment-to-moment perception and action. The resulting simulations not only reproduce observed behavioral dynamics in volatile tasks but also offer testable predictions about the neural signatures of fast and slow inference, setting the stage for empirical work on multi-timescale prediction in the brain.
Empirical evidence for multi-timescale inference
Empirical support for multi-timescale inference comes from a wide range of behavioral paradigms in which environmental statistics change at different rates. One classic line of evidence involves probabilistic reversal learning, where subjects learn which of two options is more rewarding while the underlying contingencies occasionally flip. Behavioral modeling consistently reveals that a single fixed learning rate cannot capture participants’ choices; instead, models with dynamically adjusted learning rates—or explicit hierarchical structures that estimate both choice values and environmental volatility—fit better. These data suggest that human learners track not only fast fluctuations in reward outcomes but also slower changes in the likelihood that the rules themselves might shift, consistent with two-time mechanisms of bayesian updating embedded in hierarchical generative models.
Perceptual decision-making tasks provide converging evidence. In dynamic random-dot motion paradigms, where coherence levels or motion directions change unpredictably, observers integrate sensory evidence over time to decide on motion direction. Analyses of choice patterns show that the effective integration window shortens when the environment is volatile and lengthens when it is stable, implying a slow-timescale inference about volatility that modulates fast evidence accumulation. Drift–diffusion models augmented with a volatility-sensitive component, or hierarchical Gaussian filter fits, capture these adaptive changes in integration timescales, indicating that participants infer how quickly the environment changes and adjust their moment-by-moment decision policies accordingly.
Change-point detection tasks make the separation between fast and slow inference particularly transparent. Participants observe a sequence of stimuli—such as tones, visual locations, or reward magnitudes—generated by a distribution whose mean or variance occasionally jumps. Behaviorally, participants do not update their expectations in a uniform, incremental fashion. Instead, they exhibit long periods of conservative updating punctuated by rapid shifts in belief when sustained discrepancies accumulate. Bayesian change-point models, which maintain a posterior over the hazard rate of change, reproduce this pattern by implementing a slow-timescale estimate of how likely it is that a change has occurred, superimposed on fast tracking of the current mean. Empirical fits show that such models outperform simple delta-rule learners, supporting the notion that human subjects exploit multi-timescale neural inference to parse stable regimes from transient noise.
Trial-by-trial physiological data align closely with this behavioral structure. In oddball paradigms, where rare deviant stimuli are embedded among frequent standards, the mismatch negativity (MMN) component of the EEG is interpreted as a neural signature of prediction error. When the probability of deviants changes over blocks, MMN amplitude does not merely follow the local frequency of deviants; it also reflects participants’ inferred statistical structure. For example, when blocks are more volatile—containing frequent reversals in deviant probability—MMN amplitude adapts more quickly to changes, consistent with an upregulation of precision on fast sensory errors. In more stable blocks, MMN adaptation is slower and more conservative, indicating that high-level priors about stability dampen the impact of individual surprises. Source localization studies implicate a network spanning auditory cortex and frontal regions, with frontal generators exhibiting slower dynamics that likely encode longer-term statistical expectations.
Hierarchical auditory processing offers further evidence through “roving standard” and “many-standards” paradigms. Here, what counts as a “standard” tone is periodically redefined, forcing participants’ brains to relearn the underlying auditory regularities. Early evoked responses in primary auditory cortex adjust rapidly to the new standard—indicating fast sensory-level updating—while later components and frontal responses recalibrate over a longer series of repetitions. This temporal differentiation mirrors the distinction between fast inference about current tone probabilities and slower updates to the inferred structure of the auditory environment, supporting the idea that predictive processing in sensory systems is layered across distinct temporal scales.
In the visual domain, adaptation and serial dependence phenomena also reflect multi-timescale inference. Short-term adaptation to motion or orientation occurs within seconds, leading to repulsive aftereffects in perceived direction or tilt. At the same time, prolonged exposure across minutes or longer can shift baseline perceptual biases in more persistent ways. Psychophysical modeling indicates that immediate perception reflects a weighted combination of recent sensory history and long-standing priors about natural scene statistics. The relative contributions of these components change with environmental stability: in rapidly changing conditions, perception leans more heavily on the last few stimuli; in steady contexts, long-run priors exert stronger influence. This interplay between quick sensory adaptation and slower recalibration of structural expectations exemplifies two-time updating in perception.
Measurements of eye movements and sensorimotor control likewise reveal distinct temporal layers of learning. In smooth pursuit and manual tracking tasks, participants follow targets whose motion statistics can abruptly or gradually change. Kinematic analyses show that tracking responses adjust rapidly to immediate deviations in target position or velocity, while parameters such as gain and anticipatory timing evolve more slowly across many trials. Adaptive Kalman filter models fitted to the data require both fast correction of prediction errors and slower updates to process noise estimates to account for behavior. Neurophysiological recordings from oculomotor regions, such as the frontal eye fields and cerebellum, show parallel patterns: neurons encoding current error signals respond on fast timescales, while others reflect gradual tuning of prediction and gain, indicating distinct but interacting layers of adaptation.
Neuroimaging studies using fMRI have sought to dissociate neural substrates of fast and slow inference explicitly. In tasks where subjects must learn changing reward contingencies, striatal activity often tracks immediate reward prediction errors, while prefrontal areas, including dorsolateral and orbitofrontal cortex, code for inferred volatility or higher-order beliefs about task structure. When computational models with separate state and volatility estimates are fitted to behavior, trial-by-trial regressors derived from state prediction errors and volatility updates map onto distinct brain networks: striatal and sensory regions correlate most strongly with the former, whereas anterior cingulate cortex and lateral prefrontal cortex correlate with the latter. This division supports the view that slow-timescale inferences about environmental change are implemented by higher-order control circuits that modulate the learning dynamics of lower-level systems.
Magnetoencephalography and intracranial electrophysiology provide a more temporally resolved picture consistent with multi-timescale predictive coding. In auditory and visual tasks, gamma-band activity in early sensory cortices tracks fine-grained prediction errors, unfolding within tens of milliseconds after stimulus onset. Concurrently, slower oscillations in theta and alpha bands, often localized to frontal and parietal regions, show gradual shifts that correlate with blockwise changes in task context or volatility. Cross-frequency coupling analyses reveal that the phase of slower rhythms modulates the amplitude of fast gamma responses, suggesting that slow-timescale contextual inferences gate the gain of rapid error processing. This coupling offers a mechanistic substrate for two-time updating, where slow dynamics tune the sensitivity of fast perceptual processes on a moment-to-moment basis.
Pharmacological manipulations provide causal evidence that neuromodulatory systems support multi-timescale adaptation. Agents that influence noradrenergic or cholinergic transmission alter how participants balance slow priors against fast sensory information. For instance, increasing noradrenergic tone tends to enhance sensitivity to environmental volatility, leading to faster adjustments in learning rates and increased weighting of recent outcomes. Conversely, dampening these systems promotes more rigid behavior, with participants clinging to outdated expectations despite accumulating evidence to the contrary. These effects map neatly onto theoretical proposals that neuromodulators encode precision or uncertainty, thereby regulating the relative strength of fast and slow components in predictive processing architectures.
Developmental trajectories and aging effects offer a natural experiment on how multiple timescales of inference emerge and degrade over the lifespan. Young children often exhibit high learning rates and pronounced responsiveness to individual outcomes, suggesting a dominance of fast updating with relatively weak or unstable higher-level priors about environmental stability. As children mature, they become better at distinguishing random fluctuations from true rule changes, aligning more closely with models that include a robust slower layer estimating volatility. In older adults, evidence points to a partial breakdown of this calibration: some studies report overly conservative updating despite clear environmental change, while others find excessive sensitivity to noise. Both patterns can be interpreted as disruptions to slow-timescale inference or its capacity to modulate fast learning appropriately.
Disorders of perception and cognition provide especially salient evidence for the functional importance of multi-timescale inference. In schizophrenia, for example, participants often overweight recent sensory evidence relative to prior expectations in certain tasks, yet cling to idiosyncratic high-level beliefs in others. Computational modeling of reversal learning, auditory mismatch, and sensory integration tasks frequently points to abnormal volatility estimates and precision weighting across levels. Neuroimaging studies reveal altered engagement of frontal and temporal regions associated with higher-order belief updating, alongside atypical sensory prediction error signals. These findings support accounts on which hallucinations and delusions emerge from imbalances in two-time updating, where either fast error signals are granted excessive influence in some contexts or slow priors about the world and self become pathologically resistant to revision.
Autism spectrum conditions provide another window into disrupted multi-timescale predictive mechanisms. Empirical work suggests that autistic individuals may rely less on long-range priors and more heavily on immediate sensory input, leading to superior performance in certain low-level tasks but difficulties with contextual integration and social inference. In volatile learning tasks, some studies report atypical adjustments of learning rates and reduced sensitivity to inferred environmental change. At the neural level, differences in MMN adaptation, reduced habituation in sensory cortices, and altered connectivity between sensory and frontal regions have been reported. These patterns can be interpreted as reflecting an altered balance between fast sensory updating and slower contextual calibration, consistent with predictive processing theories that emphasize atypical precision weighting across temporal hierarchies.
Empirical investigations into the subjective side of experience hint that multi-timescale inference shapes consciousness as well. Temporal integration windows in perception, such as the time frames over which stimuli are bound into a single conscious event, appear to depend on both fast sensory dynamics and slower contextual expectations. For instance, in the perception of ambiguous figures or bistable motion, moment-to-moment fluctuations in interpretation are constrained by longer-term priors about plausible configurations of the world. Brain imaging studies of such phenomena often reveal rapid competition between alternative representations in sensory regions, modulated by slower shifts in frontoparietal activity that reflect changes in task set or attentional context. These data suggest that the flow of conscious experience may itself be a product of two-time updating, with fast representational changes unfolding within a scaffold of slow, high-level beliefs that confer continuity.
Collectively, these empirical strands—from behavior and physiology to pharmacology and psychopathology—converge on the idea that the brain implements predictive processing through temporally layered mechanisms. Fast processes track immediate states of the world and correct errors on short timescales, while slower processes infer the stability, volatility, and reliability of those states, reshaping how future errors are treated. Multi-timescale inference thus finds support not only in abstract computational models but in the measurable signatures of neural and behavioral adaptation observed across diverse experimental contexts.
Implications for cognition and psychopathology
Two-time updating has pervasive implications for cognition because much of what is called thinking, remembering, deciding, and attending can be cast as the regulation of interactions between fast and slow inferential loops. At the fast end, cognition involves on-line tracking of task-relevant variables—such as which option is currently best, which cue is informative right now, or which interpretation of a sentence is most likely. At the slow end, it involves learning which kinds of patterns tend to recur, which strategies are generally effective, and which perspectives on a situation are worth maintaining despite short-term setbacks. Within a predictive processing framework, these facets of cognition instantiate bayesian updating at different temporal depths: fast loops revise immediate hypotheses, while slow loops adjust structural priors about tasks, domains, and self-relevant contingencies. The alignment—or misalignment—between these loops is central to whether cognition appears flexible, stable, or dysfunctional.
Attention is a primary example of this dynamic. On one view, selective attention just is the controlled allocation of precision across levels of a predictive hierarchy. Fast attentional shifts prioritize current prediction errors in particular sensory channels or feature maps, enabling quick reorientation to salient events. Slower, goal-directed control sets the background policy for these shifts, encoding beliefs about which kinds of signals will be informative in a given context (for instance, faces in a crowded room or numerical information in a financial report). Two-time updating ensures that attention can be rapidly captured by unexpected events yet gradually re-tuned by experience: if a particular cue repeatedly turns out to be misleading, its associated precision is downgraded over time. Failures of attention—such as distractibility or perseveration—can be interpreted as imbalances between the influence of fast, bottom-up precision signals and slow, top-down guidance about what should matter.
Working memory and executive control can be similarly reframed as mechanisms that buffer slow-timescale priors against the volatility of ongoing sensory input. Representations held “in mind” over seconds to minutes correspond to high-level states that are prevented from updating too quickly in response to noise. Fast inference continuously integrates new evidence, but executive systems impose inertia on certain variables, treating them as task-set or goal priors that should only change when prediction errors accumulate in a sustained way. From this perspective, cognitive control problems—such as difficulty maintaining a plan in the face of temptation, or conversely, difficulty abandoning an obsolete strategy—reflect miscalibration of the timescale on which task-related priors are updated. If slow priors are too labile, behavior becomes erratic and context-insensitive; if they are too rigid, behavior becomes stereotyped and unresponsive to important changes.
Memory systems naturally map onto the two-time architecture. Episodic encoding supports fast updating, capturing specific events and their immediate contexts, while semantic and procedural memory embody slower, cross-episode learning that reshapes more abstract generative models. Each new experience introduces prediction errors that can modify expectations about particular people, places, or actions in the short run; only when similar patterns recur across many episodes do they significantly revise deeper priors about social scripts, causal regularities, or one’s own capacities. This temporal filtering explains why single traumatic or intensely rewarding experiences can transiently bias perception and choice, yet enduring cognitive change typically requires repetition and consolidation. When consolidation mechanisms are compromised, fast inference may remain intact—people can still track recent events—but slow restructuring of priors is impaired, leading to brittle cognition that fails to integrate experience over long horizons.
The relationship between two-time updating and consciousness is more speculative but increasingly explored. One influential idea is that contents of consciousness correspond to intermediate-level hypotheses that win the competition for high precision at a given moment, stabilized by slower priors that define the current “scene” or task context. Fast neural inference continuously proposes candidate interpretations of sensory input, while slower dynamics provide a scaffold that constrains which of these interpretations can cohere into a unified, reportable experience. This view helps explain how consciousness can feel both rapidly changeable and deeply continuous: fleeting contents (such as the exact phrasing of a sentence just heard) are nested within a more slowly evolving sense of situation, self, and goal. Disruptions to slow-timescale priors—such as unstable beliefs about bodily integrity or social reality—may lead to fragmented or incoherent experiences, even if fast perceptual processing is relatively preserved.
Time perception itself likely depends on multi-timescale neural inference. The brain cannot directly sense time; instead, it infers temporal structure from the unfolding of events and internal dynamics. Fast timescales support interval timing and ordering of closely spaced stimuli, while slower timescales code the estimated rate of environmental change and the typical duration of situations or activities. When the environment is judged to be stable and predictable, slow priors may favor longer integration windows, yielding the impression of smooth, continuous time. Under high volatility or arousal, precision may shift toward fast error signals, compressing subjective integration windows and making time feel fragmented or dilated. Abnormalities in this calibration can contribute to cognitive symptoms in several disorders, such as the sense that time is “stopped” or “rushing” without corresponding changes in the external world.
Language and narrative cognition illustrate how two-time updating supports complex, structured thought. Parsing a sentence involves fast, incremental predictions about upcoming words and syntactic constructions, coupled with slower integration into a discourse model that spans multiple sentences or episodes. Misleading garden-path sentences show that initial fast inferences can be overturned when later words introduce persistent prediction errors, forcing a slower reinterpretation of the syntactic or semantic structure. At the narrative level, readers and listeners maintain high-level priors about characters, motives, and plots that are revised only when contradictions are strong and sustained. Cognitive difficulties in language comprehension can thus be viewed as disruptions in coordinating these timescales—for instance, overreliance on local word-level cues without adequate use of discourse-level priors, or excessive rigidity in high-level interpretations that resist correction by accumulating anomalies in the text.
Decision-making and valuation processes offer another domain where two-time mechanisms are evident. On short timescales, agents track recent outcomes and adjust expectations about which actions are currently advantageous. On longer timescales, they infer properties of the environment such as volatility, reward richness, and the reliability of different information sources. These slow inferences influence risk preferences, exploration strategies, and the willingness to switch away from established habits. If slow priors underestimate environmental volatility, individuals may persist in suboptimal behaviors even as negative outcomes accumulate. If they overestimate volatility, they may abandon promising options prematurely or exhibit unstable preferences. Many cognitive biases—such as overgeneralizing from recent streaks of success or failure—can be understood as distortions in how fast outcome signals are weighted relative to slow beliefs about stability and change.
Psychopathology can be recast within this framework as a family of disorders in which the calibration between fast and slow updating is compromised, often in domain-specific ways. Anxiety disorders, for example, may involve slow priors that overestimate the base rate and persistence of threat, causing the system to assign high precision to danger-related prediction errors and to interpret ambiguous cues pessimistically. Fast inference then becomes biased toward detecting and confirming threat, with minimal opportunity for disconfirming evidence to accumulate. Even when many safe experiences occur, they may be relegated to fast, local corrections that never substantially revise entrenched high-level priors about vulnerability or danger. This can generate chronic hypervigilance and difficulty extinguishing fear responses, consistent with findings of impaired safety learning and context-dependent extinction in anxiety.
Depressive disorders have been modeled as states in which slow-timescale priors about reward availability, self-efficacy, and social value become pathologically negative and resistant to change. In such conditions, positive prediction errors—unexpected rewards or successes—are treated as low-precision anomalies, leading to minimal updates of high-level beliefs about the future. Fast inference still registers transient pleasure or relief, but these experiences are not allowed to accumulate into a durable revision of pessimistic expectations. Over time, this imbalance between fast hedonic responses and slow belief revision can produce the characteristic combination of anhedonia, hopelessness, and cognitive bias toward negative information. Therapeutic interventions that encourage behavioral activation and re-engagement with rewarding activities may work partly by increasing exposure to consistent positive prediction errors, gradually forcing a shift in slow priors.
Obsessive–compulsive disorder (OCD) can be interpreted as a disruption in how uncertainty and precision are managed across timescales. One proposal is that individuals with OCD have difficulty trusting slow priors about the completion of actions or the absence of threat, causing them to repeatedly seek fast, sensory-level confirmation (for instance, checking that a door is locked). Because fast confirmations do not substantially update slow beliefs—either due to underweighting of confirming evidence or to overly conservative volatility estimates—the system never reaches a stable state of “enough certainty.” This leads to repetitive behaviors and intrusive doubt, even when the objective situation is secure. Cognitive treatments that target beliefs about responsibility and threat may help recalibrate slow-timescale expectations, breaking the loop between momentary relief from checking and rapid re-emergence of uncertainty.
In psychotic disorders, the imbalance between timescales can be more extreme and internally inconsistent across domains. Some accounts suggest that overly high precision is assigned to certain fast sensory or interoceptive prediction errors, making ordinary fluctuations feel unusually salient and in need of explanation. At the same time, slow priors about the self, others, and the structure of reality may become aberrant—too easily revised in response to idiosyncratic experiences in some contexts, yet excessively rigid in others. This combination can support the formation of delusions: unusual perceptual experiences are interpreted through increasingly specific and implausible high-level hypotheses, which are then shielded from disconfirmation by reinterpreting subsequent events in their light. Hallucinations can likewise be modeled as cases where strong top-down predictions, perhaps reinforced by slow maladaptive priors, dominate weak or noisy sensory evidence, causing internally generated content to be experienced as externally imposed.
Autism spectrum conditions highlight another pattern of disruption: comparatively weak or underutilized slow priors in certain domains, coupled with strong reliance on fast sensory evidence. This can yield advantages in detail-focused tasks where local information suffices, but difficulties in domains that depend heavily on long-range structure and context, such as intuitive social cognition, pragmatic language use, and flexible generalization. In two-time terms, the system may update vigorously at fast levels, accurately tracking immediate contingencies, yet form less robust or less influential slow beliefs about others’ mental states, social norms, or conventional meanings. As a result, each new situation may feel relatively novel, requiring effortful reconstruction rather than effortless reuse of abstract generative models. Interventions that scaffold the acquisition of explicit, stable rules and regularities can be seen as external supports for slow-timescale learning that the system does not perform as automatically.
Post-traumatic stress disorder (PTSD) offers a complementary picture, in which slow priors are powerfully reshaped by a subset of highly salient episodes. Traumatic events may produce extremely large prediction errors concerning safety, agency, or the predictability of others’ behavior. If these errors are assigned high precision and not subsequently counterbalanced by equally consistent and precise evidence of safety, they can produce enduring priors that treat many otherwise neutral contexts as threatening. Fast inference then becomes biased toward reactivating trauma-related interpretations, generating flashbacks, hyperarousal, and avoidance. Importantly, ordinary safe experiences may fail to update these priors because they are processed in low-precision modes (for example, during dissociation or emotional numbing), preventing them from being registered as strong counter-evidence. Successful therapies often create conditions in which safe re-exposure to trauma cues occurs under high precision for corrective prediction errors, promoting slow restructuring of the relevant generative models.
Substance use disorders can also be viewed through the lens of two-time updating. Repeated drug use establishes slow-timescale priors that overvalue drug-related cues and undervalue natural rewards, while fast prediction errors within drug contexts are heavily weighted. This combination biases attention, memory, and choice toward substance-seeking behaviors. Even when individuals consciously intend to abstain, slow priors about the expected relief or pleasure from use, and about the difficulty of alternative coping strategies, can dominate, undermining top-down goals. Episodes of abstinence that contradict these expectations may produce fast positive prediction errors (for instance, discovering that certain activities are enjoyable without the substance) but, particularly early in recovery, these signals often fail to induce large shifts in entrenched reward priors. Treatment strategies that emphasize prolonged, structured exposure to alternative sources of value can be seen as attempts to accumulate sufficient counter-evidence to reshape slow value models.
Across these diverse conditions, a common theme emerges: psychopathology often involves not simply “wrong beliefs” or “distorted perceptions,” but systematic mis-timing of belief revision across hierarchical levels. Fast neural inference may be too noisy, too precise, or poorly gated; slow neural inference may be too rigid, too labile, or insufficiently connected to immediate experience. Because cognition and consciousness are built from the ongoing interplay of these timescales, disturbances in their coupling can manifest as aberrant attention, memory, affect, and sense of self. This perspective encourages diagnostic and therapeutic approaches that focus on how individuals learn from prediction errors across time, and how interventions might restore a healthier balance between rapid adaptation to the present and slow integration of the past into guiding priors for the future.
