Protracted prediction: Neurodevelopment of reward processing in the adolescent cerebellum

Abstract

Adolescence is characterized by heightened reward sensitivity and risky decision-making. Prevailing neurodevelopmental frameworks typically attribute these behavioral trends to a maturational imbalance between rapidly developing motivational brain regions, and slower-maturing prefrontal cognitive control circuitry. However, these models largely overlook the cerebellum—a structure that demonstrates protracted development across adolescence and reciprocal connectivity with both striatum and prefrontal cortex. Computational models also highlight the cerebellum’s central role in reinforcement learning and error-based model updating, making it a potentially critical region for understanding adolescent reward processing. To evaluate this, we conducted a systematic literature search and coordinate-based meta-analysis of functional magnetic resonance imaging (fMRI) studies examining reward anticipation and receipt in healthy adolescents (19 studies; 55 cerebellar peaks). Results demonstrate a striking functional dissociation. During reward anticipation, adolescent cerebellar activation mirrors adult topographies—demonstrating widely distributed activation patterns across the cerebellar lobules and Vermis, localized to cerebellar regions that are functionally connected with salience, somatomotor, and frontoparietal cortico-cerebellar networks. Conversely, while reward receipt also elicits widespread cerebellar activation in adolescents, this stands in contrast to highly focal feedback-locked reward activity seen only in the Vermis in adult studies. We interpret these findings through the lens of cerebellar reinforcement learning. We argue that widespread reward outcome-locked BOLD activity in adolescents may reflect broader parallel fiber recruitment, supporting the active maintenance of short-timescale eligibility traces required for credit assignment while internal forward models are being constructed during development. Ultimately—rather than a biological epiphenomenon—it is hypothesized that this active cerebellar computation during adolescence may contribute to the developmental shaping of prefrontal networks necessary for normative regulation of motivation and decision-making in adulthood.

Introduction

Adolescence is associated with a heightened sensitivity to rewards. Adolescent fMRI studies suggest this increased sensitivity to rewards is driven by enhanced recruitment of motivational brain regions including ventral striatum, amygdala, and orbitofrontal cortex (Cohen et al., 2010; Ernst et al., 2005; Galván and McGlennen, 2013; May et al., 2004). In parallel, cognitive control circuits anchored in lateral prefrontal cortex (PFC) undergo protracted development in adolescence, and are delayed in establishing the long-range connections with motivational regions critical for regulating motivated behavior (Crone and Steinbeis, 2017; Luna, 2009; Luna et al., 2010). Accordingly, ‘dual systems models’ of neurodevelopment have argued that an imbalance between hyperactivation of motivational brain regions in response to rewards, alongside hypoactivation of cognitive control circuits, drives the canonical increase in emotional and risky decision-making in adolescence (Casey et al., 2008; Fareri et al., 2008; Luna et al., 2015; Somerville et al., 2010; Steinberg, 2007). While these models have offered a critical heuristic for understanding adolescence, they fail to consider the potential relevance of another region that undergoes protracted maturation and plays a functional role in reward processing—the cerebellum.

Despite being one of the first structures to emerge in fetal development, the cerebellum is among the last brain regions to fully mature, a protracted developmental trajectory necessitated by the establishment of its complex reciprocal cerebellar-cerebral circuit anatomy (Dobbing, 1974). Similar to the prefrontal cortex, cerebellar structural volume follows an inverted “U-shaped” developmental trajectory (Tiemeier et al., 2010). This volumetric growth occurs later than in much of the neocortex, with total cerebellar volume peaking between 12 and 16 years of age (≈12 years in females, ≈16 years in males; (Tiemeier et al., 2010)). Importantly, this maturation is not spatially uniform, and “cognitive” regions of the cerebellum including Crus I-II appear to develop more slowly than more anterior sensorimotor areas (Bernard et al., 2016, 2015). Subsequently, the cerebellum’s pruning and shrinkage developmental trajectory is accelerated relative to the prefrontal cortex—while the PFC continues to undergo significant synaptic pruning throughout early adulthood (Petanjek et al., 2011), cerebellar grey matter volumes remain largely stabilized after the late teens / early 20s (Bernard and Seidler, 2014). As a result, it is hypothesized that the cerebellum may help to scaffold neurodevelopmental processes in connected cortical structures, by shaping activity-dependent myelination and cortical pruning in remotely connected regions like the PFC (Moberget et al., 2015; Wang et al., 2014).

In adulthood, the cerebellum likely plays a more central role in cognitive control and reward-guided behavior than traditionally realized. Recent evidence suggests the cerebellum functions as a critical modulatory hub that bridges motivational and control circuitry. It forms reciprocal loops with the prefrontal cortex which support high-level cognitive control (Buckner et al., 2011; Schmahmann, 2019), while simultaneously maintaining reciprocal connectivity with the basal ganglia (Bostan et al., 2010). Furthermore, circuit-manipulation studies suggest the cerebellum exerts direct influences on striatal reward processing (Chen et al., 2014) and dopaminergic tone in the ventral tegmental area (Carta et al., 2019). This architecture leaves the cerebellum well-positioned to integrate reward signals into ongoing cognitive control processes. Therefore, it is necessary to consider expanding existing dual systems models to better account for the role of the cerebellum in adolescent neurodevelopment. To address this need, we conducted a systematic search of the fMRI literature, synthesizing the current evidence on the cerebellum’s involvement in reward anticipation and receipt during adolescence.

Cerebellar Circuit Anatomy

To integrate the cerebellum into models of adolescent neurodevelopment, we must first understand this region’s circuit makeup and functional involvement in reward processing (Figure 1). The physical architecture of the cerebellar circuit relies on relaying widespread neocortical data via a vast granule cell network, whose parallel fibers extend horizontally to synapse onto Purkinje cells where those signals intersect with sparse feedback from climbing fibers originating in the inferior olive of the medulla (Houck and Person, 2014; Tomasch, 1969). This structural convergence allows the Purkinje cells to rapidly adapt their synaptic weights, ultimately transforming high-dimensional information into the precise, timed release of deep cerebellar nuclei, which send excitatory outputs back to the neocortex via the thalamus (Diedrichsen and McDougle, 2026).

Figure 1. Anatomy and Functional Architecture of the Human Cerebellum.

(A) Anatomical lobules of the human cerebellum, displayed on coronal and axial slices alongside a 2D flatmap projection of the cerebellar cortex (Diedrichsen et al., 2009). (B) Cortico-cerebellar functional resting-state networks mapped onto the cerebellum. Note that the majority of the human cerebellum is densely and reciprocally connected to non-motor regions of parietal, prefrontal, and temporal cortices (Buckner et al., 2011).

The cerebellum is essential for motor control. The temporally-precise cerebellar circuit is thought to play a role in predicting and updating the sensory consequences of planned voluntary actions, ensuring that complex movements are smoothly executed and adjusted online to minimize errors (Wolpert et al., 1998). This “internal model” function of the cerebellum is necessary in the sensorimotor domain, with extensive evidence that damage to this region causes impairments in the execution of complex movements (Kawato, 1999; Therrien and Bastian, 2015). However, converging lines of anatomical (Middleton and Strick, 1998), clinical (Schmahmann and Sherman, 1998), and functional imaging (Buckner et al., 2011; King et al., 2019) evidence have now firmly established that the cerebellum plays important functional roles in non-motor domains as well (for recent reviews, see (Diedrichsen and McDougle, 2026; Sokolov et al., 2017)). In fact, the majority of the human cerebellum is densely and reciprocally connected to regions of the parietal, prefrontal, and temporal cortices that contribute to a wide array of non-motoric cognitive and affective functions (Amore et al., 2021; Diedrichsen and McDougle, 2026; Schmahmann, 2019). The regions of the cerebellum that project to PFC, lobules Crus I and Crus II, appear to be particularly expanded structurally in humans relative to nonhuman primates, suggesting that these regions and their cortico-cerebellar circuit architecture may play a role in high-level cognition (Balsters et al., 2010). This new frontier in the science of the cerebellum includes strong evidence that it plays a functional role in reward processing (e.g. (Kruithof et al., 2023)).

Cerebellar Reward Processing

Prominent among the non-motor domains that have recently been linked to cerebellar circuitry is the processing of rewards and the computations underlying reinforcement learning. In nonhuman animal model studies, the cerebellum has been shown to modulate dopaminergic tone in ventral tegmental area, and also forms a loop with the basal ganglia—projecting to striatum through the thalamus, and receiving return projections from striatum via the subthalamic and pontine nuclei (Bostan et al., 2010; Carta et al., 2019). At the circuit level, granule cells providing cortical inputs to the cerebellum appear to encode the expectation of rewards, whereas climbing fibers from inferior olivary nucleus to cerebellum encode post-receipt differences between what was expected and what was received (i.e., prediction errors; (Heffley and Hull, 2019; Kostadinov et al., 2019; Wagner et al., 2017)). These reward anticipation and feedback-locked computations are thought to contribute to error-driven learning and updating of an affective-motivational internal model, analogous to what the cerebellum contributes in the sensorimotor domain (cf., (Manto et al., 2024; Sokolov et al., 2017)).

Rather than representing a redundant byproduct of cortico-striatal value computations, cerebellar reward signals appear to play a necessary functional role in reinforcement learning and decision-making. Indeed, patients with cerebellar degeneration show specific reinforcement learning deficits that highlight the cerebellum’s critical role in this credit assignment process (McDougle et al., 2016). Model-based fMRI in healthy volunteers suggests the cerebellum encodes an unsigned prediction error—indexing outcome “surprise” or salience rather than a valenced reward value signal—following action execution (McDougle et al., 2019). These cerebellar reinforcement signals appear to be highly temporally sensitive. The region encodes not just the occurrence of an unexpected outcome, but the precise timing of that outcome relative to the internal forward model’s predictions (Trach et al., 2025). These cerebellar computations are reflected at the macro/functional brain network level in human neuroimaging studies, where several recent meta-analyses have identified cerebellar clusters in adult task fMRI studies involving the anticipation and/or receipt of monetary incentives (Jauhar et al., 2021; Kruithof et al., 2023; Oldham et al., 2018; Wilson et al., 2018). To reconcile these macro-level findings with circuit-level computations, we propose that the expansive cortico-cerebellar network may help support reinforcement learning credit assignment by maintaining short-timescale eligibility traces—lingering neural representations of recent states or actions—via parallel fiber afferents. This allows subsequent climbing fiber prediction errors to precisely update internal forward models. This eligibility trace framework serves as a central organizing lens for the current developmental review.

Of special interest, a recent meta-analysis of fMRI studies on reward anticipation or receipt in adults focused exclusively on the cerebellum (Kruithof et al., 2023). This study found a diffuse set of clusters involved in reward anticipation, comprising several dorsal cerebellar lobules (lobules I-IV, V, VI), crus I, and anterior portions of the vermis (vermis VIIIa, VIIIb, and IX; (Kruithof et al., 2023)). For reward outcome, Kruithof et al. found a more focal set of clusters involved in reward receipt, primarily in the cerebellar vermis (vermis VI and vermis crus II) with some common activation in left lobule VI (Kruithof et al., 2023). From a mechanistic perspective, this functional divergence may reflect the distinct computational and physiological demands of these different reward processing stages. Specifically, while reward anticipation activity in cerebellum is likely driven by an established cortical prediction about the future outcome via dense parallel fiber inputs evoking a large BOLD response, reward outcome processing in cerebellum likely relies on more sparse reafferent climbing fiber teaching signals that produce a reduced BOLD response (cf., (Diedrichsen et al., 2024)). Crucially, the majority of findings on cerebellum and its involvement in reward processing come from animal models or human adults. To fully integrate the cerebellum into existing models of adolescent neurodevelopment, a central question remains: What does the cerebellum contribute to reward processing in adolescence?

Methods

Overview: A Meta-Analysis of Adolescent Reward Processing and the Cerebellum

To fully integrate the cerebellum into models of adolescent neurodevelopment, we conducted a systematic literature search and coordinate-based fMRI spatial conjunction meta-analysis of the adolescent functional neuroimaging literature on reward processing. Included studies were not restricted to those focused explicitly on the cerebellum, as many whole-brain analyses report relevant cerebellar coordinates despite few studies directly focusing on the region. We aimed to compare and contrast the results of the current adolescent meta-analytic analysis with Kruithof et al.’s recent meta-analysis of adult cerebellar fMRI studies on reward processing. The degree of overlap and/or dissociation between functional maps observed in the adult and adolescent literatures will allow us to adjudicate between three possibilities regarding adolescent neurodevelopment and the cerebellum:

1. MATURITY: If adolescent cerebellar activation during both reward anticipation and receipt closely mirrors adults, it suggests that cerebellar development does not likely impact the heightened reward-seeking characteristic of adolescence.

2. IMMATURITY: If adolescent cerebellar activation differs from adults during both reward anticipation and receipt, it suggests a generalized, ongoing maturation of cerebellar reward processing in adolescence.

3. PHASE-SPECIFIC MATURATION: Lastly, if either reward anticipation or receipt exhibits a selectively distinct pattern in adolescence, this would provide unique insight into the maturation of efferent versus re-afferent reward activity in the cerebellum.

Therefore, regardless of the findings in the current meta-analysis, understanding how they do or do not overlap with extant meta-analyses in adults will help to resolve whether and how the cerebellum should be integrated with existing neurodevelopmental frameworks of maturational imbalance in the adolescent brain.

Study Selection

Figure 2 depicts a PRISMA diagram of studies included in the current coordinate-based meta-analysis. To identify eligible studies, PubMed was searched with the following criteria: (“reward” OR “prediction error” OR “reinforcement learning”) AND (fmri OR “functional magnetic resonance imaging”) AND (development OR adolesc* OR child*) AND (healthy OR typical OR trajectory). This identified N=613 total records when the search was conducted in Spring 2024. Google Scholar was also searched, identifying 10 unique records that were not on PubMed. The search was then filtered to human articles, in English with population 10-18 years. Studies were excluded if they did not have event-related task fMRI designs, did not include task contrasts to isolate the anticipation or receipt of rewards, did not include a group of neurologically-healthy subjects, or did not report at least one cluster with peak activation coordinates within the cerebellum.

Figure 2. PRISMA diagram for study inclusion.

The systematic literature search identified 623 initial records. After screening titles and abstracts against predefined exclusion criteria, 59 full-text reports were assessed for eligibility. 40 were excluded due to missing cerebellar coordinates or falling out of scope. A final total of 19 studies were included in the current meta-analysis.

Data Extraction and Meta-Analytic Approach

Standardized coordinates from all MNI or Talairach spaces in existing studies were transformed into a common space to match the “MNI152 T1 2mm brain” distributed with the FSL software library (Smith et al., 2004). Peak cerebellar coordinates in each study’s results were extracted, and a 5mm sphere was created at these peaks in MNI standard space using FSL. These spheres were then plotted on a standard cerebellar anatomical atlas (Diedrichsen et al., 2009) using a flat map to directly visualize the degree of overlap or divergence from Kruithof et al.’s recent adult meta-analysis. Lastly, a spatial conjunction analysis was used to determine the areas that had multiple studies with converging overlap in the cerebellum (cf., (Hobson et al., 2018; Hogeveen et al., 2019)).

This scoping fMRI meta-analytic approach was adopted rather than a probabilistic whole-brain approach (e.g., multilevel kernel density analysis (MKDA) or activation likelihood estimation (ALE)). These methods rely on the ability to generate a null or chance distribution across the entire brain to test for statistical spatial convergence. Because the cerebellum is systematically under-reported or physically cropped out of the field of view in much of the older adolescent fMRI literature, the whole-brain baseline required for an ALE or MKDA approach would be inherently skewed.

Results

Initially, 613 total records were identified through PubMed with 10 additional unique records identified via Google Scholar (N=623 total studies). Title and abstract screening excluded 564 articles that were not within the scope of our systematic search. Fifty-nine reports were full text screened for inclusion and 19 articles met the current inclusion criteria (Figure 2). These articles included 55 fMRI clusters with peaks in the cerebellum with N=18 cerebellar clusters in task fMRI contrasts involving reward anticipation, and N=37 for reward receipt. See Table 1 for a full results table with all included clusters. Overall, the spatial convergence across studies is broadly consistent with a reward phase-specific developmental dissociation, with reward anticipation more closely mirroring adult activation patterns than reward receipt in adolescence.

Table 1.

List of included cerebellar clusters in the current coordinate-based fMRI meta-analysis. Full study list is included in Supplemental References.

First Author	Contrast	Clust	Task	Age	N	Peak MNI Coordinate
						X	Y	Z
Reward Anticipation
Bjork	Loss v. None	1	MID	12 to 17	24	−75	−46	4
Forbes	Anticipation Early Adolescence	1	card guessing	M≈11	77	−7	−41	−5
Forbes	Anticipation Mid Adolescence	1	card guessing	M≈11	77	−39	−5	−5
Kappel	Gain v. Neutral	1	Incentive delay (child version)	M≈11	10	4	−66	−27
Kappel	Gain v. Neutral	2	Incentive delay (child version)	M≈11	10	−30	−66	18
Lorenz	Gain v. No Gain	1	Slot machine task	13 to 16	34	−2	0	−55
Nees	Big v. None	1	MID	Mean≈14	54	−61	−26	−3
Nees	Big v. Small	1	MID	Mean≈14	54	−5	−58	3
Nees	Small v. None	1	MID	Mean≈14	54	−3	−70	−26
Nees	Anticipation Win	1	MID	Mean≈14	54	−23	−73	0
Nees	Anticipation Loss	1	simple guessing	Mean≈14	54	−16	−82	−6
Nees	Anticipation Win	1	simple guessing	Mean≈14	54	−76	−22	−2
Telzer	Pumps v. Control	1	BART	14 to 16.5	48	−50	−26	26
Telzer	Pumps v. Control	2	BART	14 to 16.5	48	−36	−56	−26
Waltmann	Negative Choice Probability	1	Probabilistic reversal learning task	M≈14.8	40	−32	−38	−56
Waltmann	Negative Choice Probability	2	Probabilistic reversal learning task	M≈14.8	40	−38	−52	−44
Waltmann	Negative Choice Probability	3	Probabilistic reversal learning task	M≈14.8	40	−30	−56	36
Waltmann	Negative Choice Probability	4	Probabilistic reversal learning task	M≈14.8	40	−78	−26	−8
Reward Receipt
Cao	Time 1	1	MID variant	M≈14	13 0	−49	−20	0
Cao	Time 1	2	MID variant	M≈14	13 0	−23	−52	24
Cao	Time 2	1	MID variant	M≈19	12 4	−21	−55	−20
Cao	Time 2	2	MID variant	M≈19	12 4	−82	−17	15
Crowley	Decision Making	1	Colorado balloon game	14 to 18	20	−14	−52	−48
Crowley	Decision Making	2	Colorado balloon game	14 to 18	20	−36	−60	6
Crowley	Decision Making	3	Colorado balloon game	14 to 18	20	−38	−64	2
Crowley	Decision Making	4	Colorado balloon game	14 to 18	20	−34	−64	−38
Crowley	Wins	1	Colorado balloon game	14 to 18	20	−68	−44	−40
Crowley	Wins	2	Colorado balloon game	14 to 18	20	−66	−28	38
Ernst	Win v. No Win	1	Wheel of fortune	9 to 17	16	−78	−16	18
Harle	High-Stake v. Baseline	1	gambling task with high and low stakes	13 to 17	42	−11	−71	−9
Hawes	Gain table s3	1	incentivized antisaccade task	10 to 18	13 9	−80	28	−24
Helfinstein	Contingent Pos. FB v. Neg FB	1	reward	14 to 18	32	−72	−24	11
Helfinstein	Contingent Pos. FB v. Neg FB	2	reward	14 to 18	32	−46	−41	−26
Insel	Gain Greater than Loss	1	magnitude tracking task with low and high stakes condition	13 to 18	74	−66	−38	42
Insel	High Greater Than Loss Gain	1	magnitude tracking task with low and high stakes condition	13 to 18	74	−34	−68	−50
Insel	High Greater Than Loss Gain	2	magnitude tracking task with low and high stakes condition	13 to 18	74	−24	−66	−30
Insel	High Greater Than Low Gain	1	magnitude tracking task with low and high stakes condition	13 to 18	74	−24	−66	−30
Insel	High Greater Than Low Gain	2	magnitude tracking task with low and high stakes condition	13 to 18	74	−28	−62	36
Nees	Win	1	MID	M≈14	54	−61	−41	−3
Nees	Loss	1	simple guessing	M≈14	54	−58	−40	0
Nees	No Loss	1	simple guessing	M≈14	54	−82	−16	−6
Nees	No win	1	simple guessing	M≈14	54	−14	−78	−6
Nees	Win	1	simple guessing	M≈14	54	−82	−16	−6
Op de Macks	Boys: Reward Greater Than Loss	1	Jackpot Gambling Task	10 to 16	50	−12	−72	−3
Vaidya	Amount Gain v. Failure Gain	1	modified MID	12 to 15	18	−36	−66	42
Vaidya	Amount Gain v. Failure Gain	2	modified MID	12 to 15	18	−36	−68	−34
Van Leijenhorst	Reward 14-15yo	1	Slot machine task	10 to 15	18	−15	−63	0
Waltmann	Negative DU PES	1	Probabilistic reversal learning task	M≈14.8	40	−28	−10	−80
Waltmann	Negative DU PES	2	Probabilistic reversal learning task	M≈14.8	40	−36	−64	−28
Waltmann	Negative DU PES	3	Probabilistic reversal learning task	M≈14.8	40	−38	−62	−46
Waltmann	Negative DU PES	4	Probabilistic reversal learning task	M≈14.8	40	−32	−72	−50
Waltmann	Positive Choice Probability < 50	1	Probabilistic reversal learning task	M≈14.8	40	−34	−32	−56
Waltmann	Positive Choice Probability < 50	2	Probabilistic reversal learning task	M≈14.8	40	−40	−52	−48
Waltmann	Positive Choice Probability < 50	3	Probabilistic reversal learning task	M≈14.8	40	−12	−78	−28
Waltmann	Positive Choice Probability < 50	4	Probabilistic reversal learning task	M≈14.8	40	−32	−56	34

Reward anticipation.

Cerebellar clusters associated with reward anticipation were identified bilaterally in cerebellar area IV, VI, Crus I, Crus II, and VIIb, additional right-lateralized clusters were found in the dentate and lobule V. Vermal activations were also identified during reward anticipation, in Vermis VI, Crus I and II, VIIb, and VIIIa. See Figure 3 for a visualized surface plot of the reward anticipation peaks identified in the current meta-analysis, depicted alongside results from a recent adult meta-analysis of reward anticipation contrasts in the cerebellum (Kruithof et al., 2023). Notably, the results from the current adolescent meta-analysis and the previous adult meta-analysis both identified similar diffuse patterns of cerebellar activation during reward anticipation. These cerebellar activations during reward anticipation were mostly observed in regions that have cortico-cerebellar functional connectivity with the salience / ventral attention and frontoparietal networks (Buckner et al., 2011). When examining cluster-level overlap, vermis lobule VI (frontoparietal control network) contained the most common overlap between clusters.

Figure 3. Developmental differences in cerebellar reward processing.

(A) Adolescent reward anticipation recruits a diffuse set of regions with some reasonable overlap with (B) the reward anticipation activity observed in a recent adult meta-analysis (Kruithof et al., 2023). Conversely, (C) during reward receipt adults exhibited a highly focal pattern primarily within the cerebellar vermis (Kruithof et al., 2023), while (D) the current meta-analysis found evidence for more diffuse activation of the cerebellum during reward receipt in adolescents. Activation peaks are depicted on coronal slices in MNI space and also projected onto SUIT flatmaps in SUIT space.

Reward receipt.

Reward receipt associated activity in the cerebellum was more diffuse in the current adolescent meta-analysis relative to what has previously been observed in adults. Specifically, reward receipt was associated with cerebellar peaks in lobules I-IV, V, VI, VIIb, and Crus I and II in the left hemisphere, with similar patterns observed in right hemisphere. Vermal activation during reward receipt was also observed in lobule VI, VIIIa/b, and IX. When examining the location of these activations, the most overlap in clusters was observed for reward receipt in vermis lobule VIIIb, Crus I/II, and lobule VI. Functionally, these areas are connected to frontoparietal control, somatomotor, limbic, and salience/ventral attention cortico-cerebellar networks (Buckner et al., 2011).

Discussion

The prevailing neurodevelopmental models for understanding risky, reward-guided decision-making in adolescence center on a maturational imbalance between prefrontal cognitive control and motivational brain regions (e.g. (Casey, 2015; Luna et al., 2015)). While the current meta-analysis does not diminish these essential models, it highlights the need to incorporate cerebellar maturation into this framework. The cerebellum demonstrates a protracted development similar to the prefrontal cortex, alongside strong reciprocal connectivity with the specific prefrontal and striatal regions involved in reward-guided decision-making (Balsters et al., 2010; Bostan et al., 2010; Buckner et al., 2011; Tiemeier et al., 2010). By synthesizing 19 task fMRI studies, the current results adjudicate between competing models of cerebellar development, and provide evidence for a clear divergence in cerebellar spatial topologies during adolescence. Specifically, the spatial clustering of BOLD responses during reward receipt—but not reward anticipation—differs markedly between adolescents and adults, suggesting a potentially important functional difference in the adolescent brain. Future studies evaluating continuous age-related changes longitudinally within youth will be necessarily to determine whether this diffuse reward receipt-locked activation represents a truly phase-specific maturational stage.

Adolescent Maturity in Reward Anticipation

Reward anticipation reveals substantial functional overlap between adolescents and adults. Several cross-sectional comparisons of reward anticipation fMRI BOLD responses during the monetary incentive delay (MID) task between adolescents and adults have found surprisingly consistent cortical and subcortical recruitment across age groups (Bjork, 2004; Forbes et al., 2010; Luciana, 2013). Within the cerebellum, both adolescents and adults demonstrate consistent activation of lobules I-V, which are functionally connected to the cortical sensorimotor network (Buckner et al., 2011). These findings likely reflect motor preparation, wherein the cerebellum receives an efference copy of planned motor commands to update the brain’s internal model of action value (Ito, 2008).

Several regions of overlapping reward anticipation activation in adolescents and adults were found in cerebellar lobule VI, in regions functionally connected to the cortical salience / ventral attention network (Buckner et al., 2011). The pattern was more robust and widespread in adult relative to adolescent lobule VI. Reward has a robust behavioral and neural impact on bottom-up, exogenous attention (Ivanov et al., 2012; Le Pelley et al., 2022). The impact of reward on attention is also known to shift across adolescent neurodevelopment (Smith et al., 2011), which may explain the more widespread activations of lobule VI in adults (Kruithof et al., 2023) relative to adolescents during reward anticipation.

Beyond sensorimotor and salience networks, adolescents and adults demonstrate overlapping anticipation activity in posterior cerebellar regions, particularly Crus I. According to both resting-state functional connectivity and task-based co-activations, the Crus I region is primarily linked to frontoparietal networks and involved in complex cognitive tasks (e.g. working memory maintenance and divided attention; (Buckner et al., 2011; King et al., 2019)). In the context of reward anticipation, this overlap may reflect the active maintenance of prior knowledge about the state space of the task and the computation of an expected value of future outcomes. However, mirroring the pattern seen in the salience network, adults appear to recruit a more broadly organized set of Crus I regions during reward anticipation than adolescents, perhaps reflecting a greater degree of complexity in their affective/motivational internal model of the task compared to adolescents.

Diffuse Cerebellar Responses During Reward Receipt in Adolescence

In stark contrast to anticipation, reward receipt in adolescents was characterized by widespread, diffuse activation across multiple functional cerebellar networks. In adults, reward receipt activations are highly focal and restricted to the cerebellar vermis (Kruithof et al., 2023). These regions overlapped with the salience and visual/occipital intrinsic cortico-cerebellar functional networks, and linked to saccades and visual recognition in task-based cerebellar assays (Buckner et al., 2011; Diedrichsen et al., 2009; King et al., 2019). This is consistent with the possibility that, in adults, reward receipt preferentially engages cerebellar networks involved in responding to simple visual cues indicating the incentive value of the feedback.

In contrast, adolescent reward receipt responses encompassed extensive bilateral clusters spanning Crus I, Crus II, lobule VIIb, and lobule VIII (Diedrichsen et al., 2009). This diffuse topology maps onto a diverse set of intrinsic cortico-cerebellar functional networks, including frontoparietal, default mode, and somatomotor circuitry (Buckner et al., 2011). These activations also span multiple task-based domains, including those involved in complex, cognitively demanding processes like working memory maintenance and divided attention (King et al., 2019). Here, we propose that this widespread cerebellar recruitment during reward receipt in adolescents—but not adults—could capture the distinct computational demands of reinforcement learning during this developmental window.

Internal Model Construction in Adolescent Reinforcement Learning

Before contextualizing the current results in the framework of adolescent reinforcement learning, the spatial extent of the BOLD signal should be contextualized through the structural constraints of cerebellar microcircuitry. The magnitude and spatial extent of the BOLD signal are fundamentally constrained by cerebellar circuitry. Climbing fibers, which originate in the inferior olive, act as discrete “teaching” signals that encode reward prediction errors; while they powerfully depolarize Purkinje cells via complex spikes, they fire at extremely low rates (Raymond and Medina, 2018). In contrast, high-dimensional contextual information arriving to the cerebellum from the cortex via dense parallel fiber networks tend to fire at higher rates than the climbing fiber pathways (Wagner et al., 2017). These circuit-level differences might contribute to distinct metabolic demands in the cerebellum. Specifically, the fMRI BOLD signal is primarily thought to reflect efferent signals from the neocortex arriving as dense parallel fiber afferents, rather than sparse re-afferent teaching signals arriving from the brainstem via the climbing fiber path (Diedrichsen et al., 2024).

In adults—where internal forward models of task contingencies are learned rapidly and stabilize quickly—reward receipt fMRI contrasts primarily involve visual recognition of the salience of a cue indicating reward, nonreward, or loss. With optimized internal models, adults may only need to maintain highly focal, short-timescale eligibility traces to accurately assign credit. This may not require very complex predictions about the task’s state space from the cortex, therefore driving a weaker and more focal functional response in the cerebellum during reward receipt. In adolescents, however, the internal models mapping actions to reward value are still under construction. Because these stimulus-action-reward contingencies are not yet stabilized, the adolescent brain may rely on a broader array of short-timescale eligibility traces across distributed cognitive and sensorimotor networks to successfully solve the credit assignment problem. The diffuse BOLD signal observed here during reward receipt in youth could therefore reflect more widespread recruitment of parallel fiber pathways from the cortex, as more effort is required to maintain these traces and process the task relevant information necessary for building and updating an affective/motivational internal model. This physiological interpretation aligns with behavioral evidence that adolescents engage in more exploratory, effortful model-building during reinforcement learning tasks, as opposed to exploitative reinforcement learning policies that can be acquired rapidly by adults (Davidow et al., 2016; Nussenbaum and Hartley, 2019).

The protracted maturation of the cerebellum occurs in parallel with the maturation of prefrontal cognitive control circuitry, and both are associated with profound changes in reward-guided behavior in youth. The current results suggest that existing fronto-striatal models of impulsive and risky reward-guided behavior in adolescence should be updated to incorporate the cerebellum. Importantly, while differences in sample size and spatial smoothing between adult and adolescent meta-analyses could partially contribute to the focal versus diffuse spatial topologies observed, the widespread adolescent activations reported here are not randomly distributed. Rather, they systematically map onto functionally relevant cortico-cerebellar networks. Thus, the widespread cerebellar activity seen during reward receipt in adolescence in the current meta-analysis may represent a more metabolically-demanding construction of cortico-cerebellar forward models during reinforcement learning in youth relative to adults. This may therefore represent an important neurocomputational component of the normative development of more efficient, adaptive motivated decision-making in adulthood.