Much of what we currently know about human reward processing and large-scale neural dynamics comes from non-invasive techniques such as magnetoencephalography (MEG) and scalp electroencephalography (scEEG) [30, 31]. These methods provide millisecond temporal resolution and have been used to study a wide range of cognitive and affective processes while participants remain relatively sedentary [32, 33]. MEG and scEEG have offered insight into decision-making, value computation, sensory processing, and motor planning, but are limited in resolving activity from deep structures such as the amygdala, hippocampus, ventral striatum, and thalamus [34].

The skull substantially attenuates high-frequency activity, particularly broadband high gamma (70–200 Hz), a signal closely correlated with local neuronal spiking [35, 36]. As a result, high-frequency signatures of reward computation, memory operations, and rapid cognitive transitions often remain inaccessible to MEG or scEEG. Spatial smearing and depth biases further limit their ability to distinguish fine-grained circuits or heterogeneous subpopulations within the cortex [37]. Because human cognition relies on rapid, high-frequency, and spatially distributed computations across both neocortical and subcortical networks, invasive electrophysiology offers a window into neural dynamics that non-invasive methods cannot reliably capture [5–7, 38].

Invasive intracranial recordings-including electrocorticography (ECoG), sEEG, microelectrode arrays, and depth electrodes-provide high spatial and temporal resolution with direct access to deep structures such as the hippocampus, nucleus accumbens, amygdala, basal ganglia, and thalamus [12, 39, 40]. These structures support reward learning, emotional processing, episodic memory, and goal-directed behavior, yet remain sparsely accessed by non-invasive recordings. Through high-density intracranial recordings, hierarchical organization within the human prefrontal cortex has been demonstrated. Anterior prefrontal regions encode long-horizon predictions, whereas posterior regions encode immediate action-outcome contingencies [41, 42]. Similarly, hippocampal-prefrontal circuits exhibit grid-like and map-like representations during predictive and memory-guided behavior [43, 44]. Specialization is also evident in motor and language cortices. Dorsal-ventral gradients in the precentral gyrus support error monitoring, articulatory sequencing, and phonological encoding [45, 46].

Within reward-related regions, including dorsomedial prefrontal cortex (dmPFC), anterior cingulate cortex (ACC), OFC, and insula, invasive recordings reveal small, intermingled populations selective for reward prediction errors (RPEs), value signals, and belief-state representations, patterns difficult to detect non-invasively [47, 48]. Such fine-grained circuit resolution is essential for understanding reinforcement learning, credit assignment, and human decision-making.

Invasive electrophysiology includes both non-penetrating and penetrating modalities:

ECoG uses subdural grids to record local field potentials with excellent signal-to-noise and broad cortical coverage [38, 49].

sEEG and depth microelectrodes target deep structures and support multi-site sampling across distributed networks [50, 51].

Although penetrating electrodes can cause mechanical injury, edema, or hemorrhage, placements are clinically driven, such as for epilepsy localization, often yielding rich datasets spanning multiple functional networks simultaneously [52, 53]. These recordings provide unparalleled opportunities to link neural dynamics with complex human cognition, behavior, and psychiatric dysfunction [7, 54]. It is important to acknowledge that invasive intracranial recordings are typically obtained from clinical populations such as patients with epilepsy or Parkinson’s disease. While these studies provide critical mechanistic insights, factors including underlying pathology, medication effects, and electrode placement can influence observed neural signals. This context highlights the value and inherent limitations of intracranial research in clinical populations.

Adaptive, intelligent human behavior depends on learning from outcomes, evaluating context-dependent value, and updating beliefs about uncertain environments [10, 32, 55]. Reward learning in humans is shaped by internal state, motivational context, episodic memory, social goals, and future planning, making it more complex than in animals [56, 57]. Computational models such as Q-learning, actor-critic, and Bayesian belief-state models have been widely applied to characterize human reward learning [25, 58].

Intracranial recordings in humans reveal rich spatiotemporal signatures of value, RPEs, uncertainty, and expectation across OFC, medial prefrontal cortex, ACC, hippocampus, insula, striatum, and midbrain structures [13, 15, 59]. Widespread cortical representations of reward expectation signals have been observed beyond canonical dopaminergic pathways, and direct recordings of dopaminergic neurons encoding reward expectation and prediction errors have been reported [57]. For example, Gueguen et al. (2021) [21] used iEEG in epilepsy patients and showed broadband gamma activity scaled with outcome valence and inversely with reward expectation, with anatomically dissociable coding of reward and punishment across prefrontal subregions. Collomb-Clerc et al. (2023) [25] found that low-frequency thalamic oscillations tracked expected value during reward and punishment learning, demonstrating thalamic involvement in human value computation. Aquino et al. (2023) [24] recorded single units in pre-supplementary motor area (preSMA) and showed neurons encoded value estimates for each option, linking medial frontal motor areas with computational decision variables. Foundational studies include Li et al. (2016) [19] showing distinct coding of expected value, risk, and experienced value across medial versuslateral OFC, Rutishauser et al. (2015) [60] demonstrating medial temporal lobe (MTL) neurons encoding surprise and conflict, and Saez et al. (2018) [13] observing RPEs in thalamus and OFC.

Theory of mind (ToM) is essential for understanding how humans learn from and about others’ rewards, intentions, and actions [61–63]. Social reward learning, such as predicting others’ behavior, evaluating fairness, or updating beliefs about partners, requires inferring hidden mental states and mapping them onto reward outcomes. Intracranial recordings demonstrate that ToM-related regions, such as temporoparietal junction (TPJ) [61], encode belief states, social outcomes, and inequity, often using signals associated with reward valuation. Jamali et al. (2021) [22] performed single-unit recordings in dmPFC and identified neurons encoding others’ beliefs, including false beliefs, with selective model-switch decoding. Marciano et al. (2023) [27] recorded OFC high-frequency activity during a repeated dictator game, showing encoding of self-reward, partner reward, and inequity type. Together, these findings indicate that social inference engages neural mechanisms closely aligned with reward learning, extending value-based computations into social domains.

Beyond language, intracranial recordings reveal the fine-grained neural dynamics that orchestrate movement. Action selection, motor vigor, and movement adaptation are shaped by RPEs and dopaminergic signals, particularly within cortico-basal ganglia circuits [64–66]. Intracranial recordings during movement in humans, primates, and rodents show how reward-related neuromodulatory dynamics interact with motor control signals, revealing how learning signals influence not just what decisions are made, but how they are executed [67, 68]. These studies demonstrate how value signals guide behavior in real time.

Faces are among the most important social stimuli humans encounter and serve as key sources of reward-related information, signaling trust, threat, affiliation, and social value [69–71]. Efficient reward learning in social environments depends on rapidly extracting and evaluating facial cues. Invasive recordings reveal highly selective neural responses to facial identity, emotion, and social salience in ventral temporal cortex, amygdala, and OFC, regions also central to reward valuation [72]. These findings show how perceptual systems interface with reward circuitry to support learning from socially relevant cues, with implications for disorders involving disrupted social reward processing.

Reward-related signals play a role in shaping learning and decision making under uncertainty, delay, and social context [3, 74]. Single-neuron and population-level recordings demonstrate that reward-related learning signals are expressed across prefrontal, cingulate, striatal, and thalamic circuits. For example, Gueguen et al. (2021) [21] showed that broadband gamma activity in the prefrontal cortex scaled with outcome valence and inversely with reward expectation, with anatomically dissociable coding of reward and punishment. Collomb-Clerc et al. (2023) [25] reported that low-frequency thalamic oscillations tracked expected value during reward and punishment learning. Aquino et al. (2023) [24] identified neurons in preSMA encoding option-specific value estimates, linking medial frontal motor areas with computational decision variables. These studies illustrate how reward-guided learning is implemented across distributed neural circuits in humans.

Naturalistic behaviors such as language, movement, and navigation provide rich contexts in which reward signals interact with cognition. Interestingly, language can facilitate abstract reward learning by allowing humans to represent and communicate value information symbolically rather than solely through direct reinforcement [69, 75–77]. Intracranial recordings reveal how cortical populations encode phonological, syntactic, and semantic structure with high temporal precision, providing the substrate through which information can be parsed and integrated. High-gamma activity in the superior temporal gyrus encodes phonetic features [75]. The motor cortex reflects articulatory and phonological encoding [45], and temporal lobe single units represent semantic categories and word meanings. Distributed population codes further support context-dependent interpretations of words [78]. These mechanisms support higher-order processing, linking symbolic representations to motivational and decision-making systems.

Reward-related signals also shape motor control and action execution. Intracranial recordings during movement reveal how cortico-basal ganglia circuits integrate value and learning signals to guide action selection, vigor, and adaptation. McCrimmon et al. (2018) [79] demonstrated modulation of broadband gamma activity in human primary motor cortex by stride phase and walking speed. Hell et al. (2018) [80] showed phase-dependent modulation of subthalamic nucleus beta activity during gait, and Louie et al. (2022) [81] reported that chronic ambulatory deep brain recordings encode locomotor state and freezing severity in Parkinson’s disease. These studies highlight how reward-related neuromodulatory dynamics influence not only decisions but also their execution.

In spatial navigation, invasive recordings have revealed neural representations that support goal-directed behavior, memory, and reward-guided learning. Studies have demonstrated place- and view-selective firing in human MTL neurons during virtual navigation [82] and grid-like representations in the entorhinal cortex [43], highlighting conserved spatial coding principles across species. These spatial maps provide a scaffold for integrating reward information, allowing the brain to encode the location and value of goals. More recent work extends these findings to naturalistic settings: medial temporal theta oscillations track both memory and spatial location during immersive virtual navigation [83], linking spatial representations with goal-directed and reward-related behavior.

A key strength of human intracranial recordings is the ability to directly relate neural dynamics to behavior at millisecond timescales. By combining invasive electrophysiology with naturalistic tasks and computational modeling, recent studies demonstrate how reward-related signals interact with internal goals, beliefs, and contextual information to guide behavior in real time. Rather than acting in isolation, reward emerges as a coordinating signal that links perception, learning, decision making, and action across distributed neural circuits.

Intracranial recordings reveal that psychiatric symptoms emerge from dynamic dysregulation within cortico-striatal-thalamo-cortical loops rather than static dysfunction of isolated regions. In depression, medial prefrontal and orbitofrontal neurons show blunted reward prediction signals and altered theta-gamma coupling, reflecting impaired integration of outcome expectations with ongoing behavior. In OCD, disrupted corticostriatal processing biases behavior toward habitual responding, leading to repetitive actions that persist despite reduced sensitivity to reward outcomes [84]. Continuous recordings allow these patterns to be tracked over time, showing how network interactions fluctuate with mood states or compulsive urges [85]. This level of precision links specific temporal and spectral neural signatures to symptom expression, guiding individualized interventions such as DBS targeting nodes where reward and action signals converge.

Chronic pain involves dynamic fluctuations in both sensory and affective neural circuits, which can be leveraged to improve neuromodulation strategies. DBS has been shown to provide short-term relief for refractory pain disorders. However, pain circuits can adapt to stimulation or fluctuate naturally in chronic pain states. Conventional continuous stimulation that does not adapt to these fluctuations has limited long-term efficacy. Advances in adaptive DBS are critical to appropriately modulate maladaptive responses to painful stimuli. By capturing natural, spontaneous fluctuations in chronic pain using intracranial recordings, Shirvalkar et al. (2023) [85] distinguished transient, acute, and sustained pain states represented in frontal cortical regions, including the ACC and OFC. Recent machine-learning approaches have further demonstrated that chronic pain states can be decoded from sEEG signals, offering a framework for closed-loop, personalized interventions.

Across studies over the past 15 years, human intracranial recordings reveal that components of reward, error, and belief processing operate on timescales of milliseconds to seconds, while learning unfolds over hours to weeks. Similarly, speech comprehension and production operate on millisecond timescales, whereas language acquisition occurs across multiple exposures and social interactions. Motor planning, integration, and execution operate over seconds, while motor learning and rehabilitation span hours to weeks. In disorders such as treatment-resistant depression, chronic pain, or movement disorders such as Parkinson’s disease, neural activity can fluctuate transiently, persist, or progress over longer timescales.

ECoG recordings are typically collected between six and 14 days with continuous video and audio monitoring, which allows for the study of cognitive dynamics across extended timescales. Recording durations may be shorter or longer depending on clinical need. Signals are commonly sampled at ∼1,000 Hz, allowing analysis of frequencies up to ∼500 Hz [86]. In practice, lower sampling rates are often sufficient for many cognitive questions, whereas higher rates are sometimes used to better capture fast transients and high-frequency activity. Signal amplitudes are typically on the order of tens of microvolts (∼50–100 μV), but vary substantially with electrode size, cortical depth, referencing scheme, and local tissue properties. High-frequency activity derived from these signals is frequently used as an index of local neuronal population firing, though the relationship to single-unit spiking is indirect and task dependent.

Typical ECoG arrays include ∼70–100 electrodes arranged in grids or strips, providing broad regional coverage. The coverage is largely based on clinical indication, for seizure localization, so it can be biased and limit generalizability. However, both smaller, high-density arrays and larger clinical montages are increasingly used. The effective spatial resolution of ECoG is on the order of a few millimeters [87], reflecting the scale of local cortical populations, but it can degrade with increased inter-electrode spacing or distance from the cortical surface. Depth sEEG electrodes generally have ∼3 mm inter-contact spacing, though spacing and contact size vary across manufacturers and clinical protocols. iEEG offers relatively high spectral resolution (Figure 2), allowing reliable measurement of high-frequency and neuronal signals critical for studying human cognition.

By contrast, non-invasive scEEG offers spatial resolution typically on the centimeter scale (∼5–9 cm), more sensitivity to low-frequency activity (< 40 Hz), and smaller signal amplitudes (∼10–20 μV) [88]. Although scEEG systems can be sampled at high rates and record activity at higher frequencies, these signals are strongly attenuated and susceptible to muscle, eye-movement, and environmental artifacts due to volume conduction and skull filtering, limiting reliable interpretation of fast, localized neural dynamics. These capabilities position intracranial recordings as a unique tool for understanding rapid neural dynamics underlying cognition, while providing a translational bridge to novel interventions in neuropsychiatric and pain disorders [85].

Three major areas of progress that are particularly relevant for reward-centered models of cognition include: (1) high-density and single-unit recordings that reveal cellular-level value representations, (2) naturalistic and ambulatory recordings that capture reward processing during real behavior, and (3) emerging implantable devices that allow for long-term, closed-loop interrogation of reward circuits. Stereotactically implanted intracranial electrodes, sEEG depth electrodes, and subdural grids and strips remain a gold standard for clinical monitoring and for research on epilepsy, movement disorders, and psychiatric disease [5]. However, traditional clinical electrodes primarily capture local field potentials and lack the spatial resolution needed to resolve how reward, belief, and decision variables are encoded at the level of individual neurons and microcircuits. More recent technologies, such as Neuropixels probes and Neuropixels 2.0, have revolutionized large-scale cellular-resolution recordings in animal models by offering dramatically increased channel density, laminar sampling along the probe shank, and the ability to record from hundreds to thousands of neurons simultaneously [89]. By enabling simultaneous recordings from hundreds to thousands of neurons across cortical layers, these probes offer the potential to directly observe how reward signals are distributed across populations.

Reward learning and decision-making rarely occur in isolation. A major limitation of traditional neuroimaging, such as functional MRI (fMRI), MEG, and traditional EEG, has been the inability to study reward processing during freely moving, real-world activity. Naturalistic intracranial recording paradigms now overcome this barrier. Ambulatory and chronic intracranial systems facilitate high-resolution recordings during gait and locomotion [90], spatial navigation, and memory [83, 90]; sleep and mood fluctuations [91], and social interaction, spontaneous speech, and unconstrained decision-making. Innovation is accelerating the translation of intracranial recording technologies from short-term experimental tools into scalable platforms for long-term study of human cognition. Fully implantable, bidirectional systems now make it possible to simultaneously record and stimulate, allowing causal interrogation of reward-related circuits and adaptive modulation based on neural state [92, 93].

Neural prostheses are evolving technologies designed to restore lost or impaired function by directly interfacing with the brain’s electrical activity. These systems decode neural signals that correspond to intended movements, speech, or cognitive goals, enabling control of external devices such as robotic limbs or speech synthesizers. Early foundational studies demonstrated that neural ensembles in motor areas could be trained to control a computer cursor or robotic arm through imagined or attempted movement signals, providing proof of concept for brain-machine interfaces [94, 95]. More recently, Moses et al. (2021) [96] achieved a groundbreaking milestone by decoding attempted speech from cortical recordings, allowing real-time synthesis of speech in paralyzed individuals, a development that offers profound implications for restoring communication. In addition to decoding motor commands, advancing prosthetic function requires access to internal goal states, motivation, and learning signals that shape behavior over time. Reward-related neural processes play a critical role in this adaptation: reinforcement signals guide error correction, promote skill acquisition, and stabilize control policies [97, 98].

Over the past decade, the National Institutes of Health (NIH) has prioritized funding in areas critical to advancing human intracranial electrophysiology research (Figure 3; Table 4). A substantial portion of resources has been dedicated to understanding language comprehension and production, including speech, reflecting its foundational role in human cognition and communication [96, 99]. Parallel investments have supported research into motor function and rehabilitation, targeting not only voluntary movement but also gait and other complex motor behaviors affected by neurological disease [97]. Notably, there has been growing emphasis on mood and psychiatric disorders, particularly treatment-resistant depression and OCD. These conditions are increasingly understood as disorders of reward processing, valuation, and motivational control, making them well-suited to investigation using intracranial recordings and neuromodulation [84]. These funding trends highlight a strategic shift toward integrating cognitive neuroscience, reward-based models of behavior, and translational neurotechnology. Note, the table is not comprehensive and is intended to illustrate funding trends rather than provide an exhaustive record.

Despite advances, significant technical and clinical challenges remain. Signal stability over time is variable, with electrode encapsulation, tissue response, and hardware degradation affecting long-term performance. Decoding algorithms often require extensive calibration and may not generalize well across tasks, contexts, or individuals. Access to internal cognitive states remains methodologically complex. The precise contribution of reward-related processes to adaptive prosthetic control requires further empirical validation. Intracranial recordings are typically obtained from patients with neurological or psychiatric disorders, which may limit generalizability. Safety considerations, surgical risk, regulatory requirements, and scalability remain substantial barriers to widespread implementation. While funding trends reflect growing interest in neuromodulation and cognitive neurotechnology, these investments do not eliminate the need for careful validation, longitudinal outcome studies, and rigorous ethical oversight. The development of neurotechnology remains iterative and contingent on overcoming substantial methodological, biological, and clinical limitations.

As intracranial recording and neuroprosthetic technologies advance, ethical considerations become increasingly important. The privacy of neural data is paramount, given that brain signals can reveal deeply personal thoughts and intentions. Robust protocols must ensure informed consent, data security, and transparency regarding how neural information is used. Informed consent requires explicit disclosure of risks, including privacy breaches, cybersecurity vulnerabilities, and potential effects on identity and agency, and safeguards against therapeutic misconception, particularly in neurologically vulnerable populations [100, 101]. Neural data governance presents distinct challenges, as intracranial recordings can reveal intentions, preferences, and cognitive states. Dual-use risks are critical considerations, as neurotechnologies developed for therapy may enable cognitive manipulation or surveillance when combined with AI-driven closed-loop systems [102]. Finally, posttrial responsibilities, including continued device access, maintenance, and long-term care, represent a critical ethical obligation that should be explicitly planned and reported.

Human intracranial electrophysiology is entering an exciting phase, where advances in technology and clinical applications are beginning to reveal how reward processing emerges from distributed, temporally coordinated neural circuits. Single-cell and population recordings provide precise insights into activity patterns across distinct frequency bands. These mechanistic insights lay the groundwork for interventions such as adaptive neurostimulation and targeted therapies, while future integration with molecular- and micro-level recording techniques promises to expand our understanding of the neural computations underlying cognition and behavior.