Sensory-to-Motor Transformations: From Serial Pipelines to Dynamic, Distributed Processes

Abstract

A key organizing principle of nervous systems is the transformation from sensation into action. Accordingly, studies of sensory-to-motor transformations have yielded foundational insights into nervous system structure and function. Here, we provide a broad review of goal-directed (as opposed to reflexive) sensory-to-motor transformations. We synthesize theoretical and experimental frameworks for linking sensory-related and motor-related neuronal activity to behavior. We provide a historical overview of sensory-to-motor transformation studies, describing evidence for the propagation of neuronal activity within neocortex transforming sensory encoding into decision formation and motor execution. We then describe emerging evidence that challenges this traditional perspective, emphasizing parallel processing, distributed representation, and shifts in cortical dependence with learning. We conclude by identifying key open questions that present exciting opportunities for future discovery. Building upon decades of foundational research, ongoing and future studies of sensory-to-motor transformations are poised to reveal the dynamic, distributed neuronal mechanisms that enable flexible linkages between sensation and action.

1. Introduction

Studies of sensory-to-motor transformations (SMTs) have been central to the philosophical and experimental underpinnings of modern neuroscience. The first SMTs to be investigated were reflexes, such as blinking in response to a sudden visual stimulus or withdrawing one’s foot from the perimeter of a fire. Prior to the writings of philosopher René Descartes, it was generally believed that the movements of one’s body were orchestrated by an immaterial, immeasurable force emanating from one’s soul. Descartes reasoned that reflexes are automatic and largely independent from conscious intervention of the mind (e.g., one cannot consciously override a startle blink) (Descartes, 1649, 1664). This novel framework established the possibility that some aspects of human behavior could be understood as mechanistic processes operating within natural laws (Fig. 1A). Thus was born the scientific investigation of nervous system function.

Figure 1: Studies of sensory-to-motor transformations across centuries.

(A) An illustration of the pain pathway from Descartes’ Traité de l’homme (Descartes, 1664). In this depiction, heat applied to the foot pulls on a long filament extending to a cavity in the head, triggering the release of a fluid that causes the muscles to contract. (B) Figure 1 from Ramon y Cajal’s 1906 Nobel lecture, showing the directional flow of signaling from sensory neurons to interneurons to motor neurons of the spinal cord, revealing the anatomical substrate of the sensory-motor spinal reflex (Ramón y Cajal, 1906) (credit: © Nobel Foundation 1906). (C) Top, from Sherrington’s The Integrative Action of the Nervous System (1906), illustrating the coordination of central (within the spinal cord) excitation (+) and inhibition (−) in mediating reflex responses (Sherrington, 1906). Bottom, adapted from Eccles’s 1963 Nobel lecture, depicting the membrane currents effecting excitation and inhibition (Eccles, 1963) (credit: © Nobel Foundation 1963). (D) Panels from a 2025 study from the International Brain Laboratory, providing our most comprehensive assessment of sensory-related and choice-related signaling throughout the mammalian brain for a goal-directed, sensory-motor behavior [adapted from (International Brain et al., 2025), licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0)].

Sensory-motor reflex arcs were the setting of our first insights into the elemental units of the nervous system. Notably, the directionality of spinal reflexes was instrumental in supporting famed anatomist Santiago Ramón y Cajal’s neuron doctrine. Figure 1 in Ramón y Cajal’s 1906 Nobel Lecture shows the flow of information from afferent dorsal root ganglion cells to interneurons of the spinal cord to efferent motor neurons (Fig. 1B) (Ramón y Cajal, 1899, 1906). Such depictions provided powerful evidence for the dynamic polarization of neurons, in which conduction operates unidirectionally from dendrites to cell bodies to axons. A few decades later, the great physiologist Charles Sherrington used reflexes to provide our first insights into the coordination of excitation and inhibition. By meticulously measuring muscle responses in spinal reflexes, Sherrington demonstrated that reflexes do not merely drive muscle contractions, which could be mediated by excitation alone. In contrast, reflexes evoke a constellation of contractions and relaxations, the latter of which could only be explained by central inhibition (Fig. 1C) (Sherrington, 1906). It would be a few decades later before the first measurements of central inhibition were established within, not surprisingly, the spinal cord (Eccles, 1963).

Reflexes use the afferent and efferent machinery that bookend all SMTs. However, not all SMTs are of the automatic type described by Descartes and studied by Sherrington. Reflexes belong to a subtype of SMT that are largely ‘hardwired’ due to their near universal benefit. Other SMTs are deliberately not hardwired, to allow for behavioral flexibility as required by a changing environment and fluctuating goals. For example, determining which fruit to harvest requires integrating awareness of the shape and color of the fruit, the season, the forest, and the immediate and long-term needs of the community. Counter to the Cartesian view, between the input and the output, the afferent and efferent, there is a distinctly material substrate that is far from stereotyped. These intermediate layers, underlying the transformation, are themselves dynamic, subject to influence by incoming sensory traces and by the outcomes and contingencies of resultant actions. These types of SMTs involve higher order regions of the brain (Fig. 1D) and are likely to have been major evolutionary drivers of their development. This review focuses on this second sub-type of SMT, which we refer to as ‘goal-directed SMTs’. As we will see below, these goal-directed SMTs involve long input-output loops through the neocortex and subcortical brain regions.

Prior studies of SMTs have led to foundational insights into nervous system function, and we fully expect that future research efforts will continue to be as generative. In this review, we provide theoretical and analytical frameworks for studying SMTs, synthesize lessons learned from prior research, and identify important open questions for future investigation. This review does not cover early sensory transduction or neural control of movement kinematics. Also, this review does not cover the ‘mapping problem,’ which addresses the alignment of internal representations within different coordinate systems (Andersen et al., 1993; Filimon, 2015). Rather, here we focus on studies of neuronal recordings (typically with single unit resolution) and neuronal perturbations in subjects performing operant tasks that have provided insights into where and how sensory signals are transformed into motor commands.

2. Frameworks for understanding and modeling sensory-to-motor transformations

Goal-directed SMTs underlie the ability to use newly acquired sensory information to trigger or modulate a motor plan. As such, SMTs are required by any biological or artificial system to respond to its dynamic environment. Our review focuses on mammalian systems, heavily weighted towards the non-human primate and rodent literature (Carandini and Churchland, 2013; Parker and Newsome, 1998). The majority of these studies utilize experimenter-controlled stimulus presentation and ballistic motor outputs, such as directional saccades, lever presses, reaches, or licks. Within this structure, the boundaries of a SMT can be clearly delineated and trial-based, with each trial extending from the onset of a sensory stimulus to the initiation of a motor response. In one common task structure (Go/NoGo, Fig. 2A), the subject learns to detect a specific stimulus (e.g., deflection of the C2 whisker), and report detection with a specific motor response (e.g., lick). In another common task structure (two-alternative forced choice, Fig. 2F), the subject learns to discriminate between multiple stimuli and perform their associated motor responses (e.g., saccade left to leftward motion, saccade right to rightward motion). Substantial research efforts have focused on revealing the internal signals that provide the sensory evidence for or against triggering the associated motor response. Conceptual and statistical frameworks have been key in guiding these research efforts, and therefore we describe two of these frameworks below.

Figure 2: Signal detection theory and the drift diffusion model.

(A) Standard structure and contingency table for a Go/NoGo task design. (B) Graphical representation of a behavioral session with a d’ of 1 and criterion (c) of 0.5, reflecting a hit rate of 50% and false alarm rate of 16%. (C,D) Two different strategies for increasing hit rate. (C) A leftward (negative) shift in the criterion without changing the stimulus/noise distributions increases the hit rate to 84%; however, it also increases the false alarm rate to 50%. This may reflect increased impulsivity or task engagement. (D) A rightward shift in the stimulus distribution increases the hit rate to 84% while keeping the false alarm rate at 16%. This may reflect task acquisition or perceptual learning. Note, the gray vertical lines and curves in (C) and (D), respectively, reflect the original performance conditions. (E) Each distribution can be further decomposed into ‘no response’ and ‘response’ distributions for the calculation of choice probability (CP). Mild positive choice probability is shown for stimulus trials (left) and no stimulus trials (right). (F) Standard structure and contingency table for a two-alternative forced choice task design, in which a subject must select one of two options on each trial. (G) In tasks that require evidence accumulation, decision formation is often modeled using the drift-diffusion model. The decision variable (now plotted on the y-axis) evolves over time until it reaches a decision boundary (A or B). Shown are three simulated trials with different stimulus strengths (blue to orange), parameterized by different drift rates.

The first framework we discuss is signal detection theory (Fig. 2B-D). Signal detection theory provides a statistical and graphical framework for understanding how a subject detects a stimulus signal in the presence of noise (Britten et al., 1992; Green and Swets, 1966; Heeger, 2007). Here, noise refers to all sources that may impair the detection of a sensory stimulus, including external and internal processes. For example, detecting a faint tone may be impaired by background sounds (external) as well as fluctuations of the subject’s attention and engagement (internal). In standard Go/NoGo tasks, subjects are presented with trials containing the target stimulus (target trial) or trials that do not contain the target stimulus (catch trials). Subjects are rewarded for responding on target trials (hits) and punished for responding on catch trials (false alarms). By a simple mathematical transformation, signal detection theory converts these standard measures of hit rate and false alarm rate into perceptual ability (d’) and the tendency to respond (criterion, ‘c’, also referred to as ‘bias’ or ‘threshold’). Importantly, each d’ value implies the strength of the sensory evidence needed to achieve that performance. For example, a d’ of zero (no perceptual ability) suggests that the decision to respond on target trials and catch trials are based on overlapping distributions of internal activity (no additional sensory evidence). A d’ of 2 (moderate-high perceptual ability) suggests that the internal activity on target trials is well separated from the internal activity on catch trials (moderate-high sensory evidence). Importantly, sensory evidence can be calculated independently from neuronal recordings as the separation of activity on target trials (stimulus) and catch trials (no stimulus) in task performing subjects (Britten et al., 1992; Liu and Pack, 2017). Thus, one can compare the perceptual ability of the subject to the sensory evidence from a given brain region using the same d’ metric. While internal activity is often quantified as spike rates, other measures, such as spike synchrony or gamma LFP/EEG power, may also be used and may provide novel insights into the task-relevant neural code (Panzeri et al., 2017). Furthermore, ‘stimulus’ and ‘no stimulus’ distributions can each be further decomposed into ‘response’ and ‘no response’ distributions as a measure of behavioral choice encoding (Fig. 2E).

The second framework is the drift diffusion model (Ratcliff and McKoon, 2008; Ratcliff et al., 2016; Ratcliff et al., 1999) (Fig. 2G). In naturalistic behaviors, sensory information is often sequentially sampled and integrated over time before triggering a response (e.g., deciding whether to change lanes when driving in traffic). The ‘drift’ in drift diffusion models captures this integration and tracks the movement of the decision variable as it approaches a decision threshold (referred to as a ‘boundary’). As such, the ‘decision variable’ in drift diffusion model terminology is similar to the ‘internal activity’ in signal detection theory terminology, in that it is presumed to reflect neuronal activity encoding the task-relevant sensory evidence. In a two-alternative forced choice task, the drift diffusion model may have two decision bounds (one reflecting each stimulus category), with the decision variable initiating between these bounds on each trial. By modeling sensory stimuli and behavioral outcomes, drift diffusion models compute key features of the decision variable, such as its rates of sensory integration and decay. Importantly, as we will see below, the dynamics of the latent decision variable in drift diffusion models resemble the neural dynamics in brain regions integrating sensory evidence in task performing subjects.

Both signal detection theory and the drift diffusion model make clear predictions about the neurobiological implementations of the SMT. They predict that a singular brain region, or interacting, distributed brain regions, encode sensory evidence for or against the presence of a task-related stimulus. Once a critical level of evidence is reached (the criterion, threshold, or boundary), this triggers the initiation of the associated behavioral response. Importantly, both signal detection theory and the drift diffusion model represent sensory evidence in ways that can be quantitatively compared with recordings of neuronal activity. As such, they enable assessments of how well recorded neuronal signals match the modeled sensory evidence. These comparisons have been central to debates over which brain regions carry the task-relevant stimulus signals, and the precise coding mechanisms used within those brain regions (e.g., (Panzeri et al., 2017; Shadlen et al., 1996)). Predictions of these models regarding motor triggering, however, are less intuitive. Neurobiology is replete with examples of thresholding (that is, activating only after a certain threshold level is achieved), at both single cell and circuit levels of organization. Yet, due to the abstract nature of the threshold in signal detection theory and the drift diffusion model, these models do not make clear predictions about its neurobiological implementations. As we will see below, the neural mechanisms of threshold-evoked motor triggering have remained amongst the most elusive aspects of the SMT.

The reader may notice that both signal detection theory and the drift diffusion model are standard frameworks in decision-making studies (Gold and Shadlen, 2007). Indeed, there is substantial overlap between the fields of SMT and decision-making. However, we do not believe these to be synonymous, as SMTs can exist without deliberative decision-making, and decision-making does not necessitate a sensory-to-motor structure. Nonetheless, any task that requires the subject to transform sensory signals into categorical outcomes may benefit from the signal detection theory and/or drift diffusion model frameworks.

3. "In theory there is no difference between theory and practice - in practice there is" Yogi Berra

Armed with the appropriate statistical frameworks, we can now investigate the neurobiological implementations of SMTs. However, even the apparently simple task of identifying sensory-related (pre-transformation) vs motor-related (post-transformation) neural signals requires careful consideration. Imagine you were recording neuronal activity from a brain region. How would you determine if the neuronal signals are sensory, motor, both, or neither? An important cautionary tale is that the name of the region can be misleading and should not be used as justification. For example, motor-related signals are often found in brain regions labeled as ‘sensory’ (Matyas et al., 2010; Musall et al., 2019; Steinmetz et al., 2019; Stringer et al., 2019; Zagha et al., 2022), and sensory-related signals are prevalent in brain regions labeled as ‘motor’ (de Lafuente and Romo, 2006; Hatsopoulos and Suminski, 2011; Zagha et al., 2015; Zareian et al., 2021).

We present three criteria that have been used to distinguish sensory-related from motor-related signals: 1) temporal alignment, 2) parametric stimulus encoding, and 3) correlation with behavioral choice. 1) Given the long and variable reaction times, temporal alignments of single trial data can help distinguish between sensory-related signals (time-locked to the onset of the stimulus) and motor-related signals (time-locked and preceding the reaction time) (Fig. 3A,B) (Hanes and Schall, 1996; Roitman and Shadlen, 2002; Zareian et al., 2021). 2) A powerful yet underutilized criterion is parametric stimulus encoding. That is, correlations between stimulus encoding (as quantified by neurometric d’) and stimulus salience (e.g., intensity or contrast) (Fig. 3C). Sensory-related signals show clear parametric stimulus encoding. Importantly, the sensitivity of stimulus encoding to stimulus salience can be compared to the sensitivity of the subject performing the task (see below). In contrast, other brain regions may show categorical stimulus encoding: all-or-none encoding that is not graded along with stimulus salience (de Lafuente and Romo, 2006). It is much less likely that these all-or-none signals are being used as sensory evidence to solve the task. However, they may be part of the post-transformation, motor-related signaling. 3) The third criterion is correlations between neural activity and behavioral choice (Fig. 3D) (Britten et al., 1996; de Lafuente and Romo, 2006). Error trials (misses or false alarms) are, by definition, dissociations between stimulus and response pairings. For example, consider the hypothetical task in which stimulus A is paired with action X [A-X] and stimulus B is paired with action Y [B-Y]. Consider error trials in which the subject responds to stimulus A with action Y [A-Y]. Does the neural activity better correlate with activity on [A-X] trials or [B-Y] trials? If [A-X] (low choice probability), then the activity is likely sensory-related; if [B-Y] (high choice probability), then the activity is likely decision- or motor-related. Lastly, it is quite common for a brain region to contain both sensory-related and motor-related signals. These signals can be segregated in different layers, intermixed across neurons within the same layer, or even present within the same neurons. There are classic examples (see below) of neurons displaying sensory-related signals early in a trial and motor-related signals late in a trial, potentially reflecting the sensory-to-motor transformation occurring within individual neurons.

Figure 3: Criteria for distinguishing sensory-related vs motor-related neural activity.

(A,B) Simulation of spiking data for sensory-related (A) and motor-related (B) activity. Top panels are spike rasters per trial sorted by reaction time (shortest reaction time on top); bottom panels are average spike rates per reaction time bin (e.g., RT 200-330 ms in blue). Left panels are aligned to the stimulus onset; right panels are aligned to the response onset. Sensory-related activity will align to the stimulus onset regardless of reaction time, whereas motor-related activity will align to the response onset regardless of reaction time. Simulations are based on data presented in Zareian et al., 2021. (C) Monkeys were trained to detect a vibratory stimulus on their fingertip, with interleaved stimulus trials and no stimulus trials. Plotted are the normalized spiking responses of neurons in a given cortical region versus stimulus strength (vibration amplitude) (showing 4 of the 8 cortical regions tested in the original study). Classical sensory regions such primary somatosensory cortex (S1) and secondary somatosensory cortex (S2) display graded responses that increase with stimulus strength. In contrast, premotor cortices (PMc) and primary motor cortex (M1) display similar response amplitudes regardless of stimulus strength. See also Britten et al., 1992 for quantitative comparisons of parametric stimulus responses and behavioral performance. (D) Choice probability profiles from multiple cortical regions. Values of 0.5 reflect equal activity on response present and response absent trials (chance-level choice probability). Key task epochs include the stimulus presentation (Stim) and response window (Resp). Black bars above traces indicate timepoints of choice probability significantly different from chance. For stimulus-present trials, significant choice probability emerged in S2 during the stimulus and remained persistent throughout the delay in PMc. Note the non-significant choice probability in S1 (top row), despite robust parametric stimulus encoding [C]. Panels [C] and [D] are adapted from de Lafuente and Romo 2006, copyright (2006) National Academy of Sciences, U.S.A.

These three neural recording criteria have been applied across many decades of SMT studies. A limitation of all these criteria, however, is that they are correlational. From recording data alone, one cannot determine whether the recorded signals are indispensable to the subject conducting the SMT. For that, causal studies are needed (O'Connor et al., 2013; Salzman et al., 1992; Zatka-Haas et al., 2021). However, even causal studies require nuanced interpretation. Positive findings from stimulation studies or suppression studies may not be due to sensory/motor-related signaling in that brain region. In contrast, the causal effects may be due to the propagation of activity or the withdrawal of activity, respectively, in connected brain regions (Carrera and Tononi, 2014; Hong et al., 2018; Otchy et al., 2015; Wolff and Olveczky, 2018; Young et al., 2000). However, combined with recording studies, causal studies have provided important insights and surprises (see below) in the study of SMTs. The most stringent criterion that a specific brain region contributes to a SMT is that stable suppression (e.g., lesion) of that brain region causes irreversible impairment on that task without gross impairment in behavior in general (e.g., global suppression of movement). Interpretations of transient impairments followed by complete or partial compensation are more complicated (Hong et al., 2018; Schiller et al., 1987; Schiller et al., 1980). Beyond ‘impairment’, additional analyses can help dissociate whether the perturbation is due to changes in sensory-related vs motor-related signaling. During suppression, if the subject is able to perform the same motor response triggered by a different sensory modality, then the crucial function of that region is likely to be sensory-related. Alternatively, if the subject is able to perform a different motor response triggered by the same sensory stimuli, then the crucial function of that region is likely to be motor-related.

4. Traditional perspective

We consider the origin of modern SMT studies as when investigators began recording neural activity with single unit resolution in subjects performing quantitative operant tasks, such as the groundbreaking work of Mountcastle, Barlow, and others (e.g., (Barlow, 1972; Barlow et al., 1987; Werner and Mountcastle, 1963)). Building on this foundation, in the late 1980s and early 1990s Newsome and colleagues pioneered a series of visual motion studies that helped establish the modern framework. The task they used is a directional discrimination task, in which moving dots were displayed on a monitor and subjects reported the global direction of motion (Newsome et al., 1990; Newsome and Pare, 1988). Trial difficulty was parameterized by altering the percentage of dots contributing to the global motion. At one extreme is fully coherent motion, with 100% of the dots moving in the same direction. At the other extreme is random motion, with 0% of the dots moving in the same direction. The subjects (rhesus macaques) were required to respond (saccade) in the direction of the coherent motion, which could be in one of two directions on each trial. In Britten et al. (1992), the authors performed single unit recordings during task performance in the middle temporal (MT) visual area, a brain region which had been known to contain a high proportion of direction-selective neurons. The authors quantified neuronal performance as the separation of spiking between preferred and nonpreferred stimulus directions. Using signal detection theory to transform behavioral data and neuronal data into stimulus sensitivity functions, the authors performed key neurometric-psychometric comparisons. They found that the sensitivity of individual neurons within MT matched the sensitivity of the behaving subject (Britten et al., 1992). These data provided strong evidence that the sensory-related signals in this brain region could account for the psychometric performance of the subjects. (For extrapolation from individual neurons to neuronal pools, see the Discussion of Britten et al. 1992.) Furthermore, this study provided a roadmap for performing neurometric-psychometric comparisons that remains highly relevant today.

The Britten et al. (1992) study investigated the relationships between neuronal stimulus encoding and task performance. As such, they provide insights into the ‘sensory’ component of the SMT. Is there a comparable method to investigate the ‘decision’ component? The same group did just that four years later. On difficult trials (low or 0% coherence of the moving dots), the subjects made both correct and incorrect behavioral responses. In Britten et al., (1996), the authors quantified the extent to which neuronal activity predicted the future behavioral response, independent of stimulus strength (Fig. 2E). They termed this metric ‘choice probability’, with a value of 0.5 reflecting no relationship between neuronal activity and response choice, and values approaching 1 indicating that neuronal activity strongly predicts the subject’s choice. Across their neuronal dataset of MT neurons, the authors reported an average choice probability of 0.55 (Britten et al., 1996). The authors interpreted this finding as evidence in favor of MT spiking variability influencing behavioral choice. However, this value of 0.55 is far from the values expected from a dedicated decision signal. Yet, once again, a roadmap had been established, this time for distinguishing correlations not of stimulus encoding, but of decision/response selection.

With these two metrics (stimulus sensitivity and choice probability), we can hypothesize the arc of an SMT. Early sensory structures, such as retinal ganglion cells or dorsal root ganglion neurons, have high stimulus sensitivity, given their close proximity to the sensory receptors. However, they may have low choice probability, meaning that for the same stimulus intensity, trial-by-trial fluctuations do not substantially bias response selection. On the other end of the arc, the lateral rectus muscle of the eye is likely to have the opposite profile. That is, the lateral rectus muscle may show little variation according to stimulus strength (low stimulus sensitivity) but would undoubtedly show maximum levels of choice probability on left vs right saccade responses (high choice probability).

Applying these two metrics, we can now track throughout the brain the progression from sensory signaling (high stimulus sensitivity, low choice probability) to motor signaling (low stimulus sensitivity, high choice probability). In one influential study, de Lafuente and Romo (2006) did just that. In a somatosensory task, on half of the trials the subjects (rhesus macaques) were presented with a vibratory stimulus of varying amplitude (target trials). The other half of trials were identical, but without a vibratory stimulus (catch trials). After a post-stimulus delay, the subjects reported (by pushing one of two buttons) whether a vibratory stimulus was present or absent. During task performance, the investigators recorded spiking activities from neurons in eight distinct sensorimotor neocortical areas, including primary somatosensory, secondary somatosensory, premotor, and primary motor cortices. The authors determined stimulus sensitivity and choice probability for neurons in each area (Fig. 3C,D). The results were highly illuminating. A possible outcome would have been an abrupt transition from sensory to motor signaling. In contrast, the investigators found a gradual transition across the recorded regions that spanned the full transformation. Primary sensory cortex neurons were highly stimulus sensitive yet displayed chance-level choice probability. On the other extreme, medial premotor cortex neurons displayed low stimulus sensitivity but high choice probability. The other sensory and motor regions displayed intermediate profiles, which correlated with their stimulus response latencies (de Lafuente and Romo, 2006).

These and related studies provided a perspective of the SMT as a propagating wave across the cortical mantle, with neuronal activity gradually transforming across both space and time from sensory-related to motor-related. Here, we refer to this perspective as the ‘serial transformation hypothesis.’ With the advent of optogenetics, we received our first causal tests of this hypothesis by perturbing neural activity at specific phases within individual trials. In a pioneering study, Guo et al., (2014) trained mice in an object localization task. On each trial, the subject was presented with a probe at one of two different locations within their whisker fields. After a post-stimulus delay, the mice reported the probe location by licking a left or right lickport. The mice were genetically engineered to express the excitatory channelrhodopsin in GABAergic neurons. Channelrhodopsin activation caused robust spiking suppression of nearby pyramidal neurons, with a spatio-temporal resolution of ~1 mm and ~100 ms. With this technology, they sampled 55 cortical areas, suppressing them either during the stimulus epoch or during the post-stimulus delay. During the stimulus epoch, the largest behavioral impairments were caused by suppressing primary somatosensory cortex. However, during the post-stimulus delay, suppressing the same primary somatosensory cortex had little influence on task performance. Instead, suppressing secondary motor cortex caused the largest impairments in task performance. Notably, secondary motor cortex suppression resulted in an ipsilateral response bias, which is consistent with a motor deficit. In contrast, primary somatosensory cortex suppression resulted in reduced stimulus detection, consistent with a sensory deficit (Guo et al., 2014). These findings provided strong causal evidence for the serial transformation hypothesis, with task-related signaling propagating within each trial from sensory cortex to motor cortex.

Within the serial transformation hypothesis, it may still be possible to localize the site of transformation from sensory-related to motor-related. An early candidate region of this transformation was the posterior parietal cortex (PPC). In one of the earliest recording studies of PPC single unit activity during task performance, Vernon Mountcastle observed fascinating conditions of spiking. Mountcastle observed neurons that only fired at the conjunction of “the nature of the object, the motivational set of the animal, and the location of the object” (Mountcastle et al., 1975). This was notably different from Mountcastle’s recordings of primary somatosensory cortex neurons, which were dominantly stimulus-driven (Mountcastle et al., 1969). The ensuing decades witnessed vigorous debates as to the functions of PPC, with investigators in separate camps defending its role in ‘attention’ (i.e., sensory-related) vs ‘intention’ (i.e., motor-related) (Colby and Goldberg, 1999; Snyder et al., 1997). Is it possible that PPC’s function is at the intersection of sensory and motor, in the transformation of a sensory to a motor signal? In a seminal study, Roitman and Shadlen (2002) investigated the activities of PPC neurons during the directional discrimination task described above (Roitman and Shadlen, 2002). They found that the PPC neurons displayed high stimulus sensitivity in the early stimulus epoch, consistent with a sensory-related signal. However, instead of responding to a stimulus at a steady state, which would be expected from a canonical sensory response, these neurons ramped up their firing rates proportional to the stimulus strength (Katz et al., 2016). From the firing rates of individual PPC neurons, they appeared to be integrating sensory evidence for or against a preferred motion direction. Moreover, these neurons did not integrate throughout the stimulus epoch. The PPC neurons reached a peak firing rate ~20 ms before the response onset, which is consistent with a motor-related signal (Hanes and Schall, 1996). This peak firing rate was the same regardless of stimulus strength, which meant that the main differences between low and high stimulus strength were the firing rate slopes and time to reach peak firing.

From the above description of the Roitman and Shadlen (2002) data, the reader may be reminded of the drift diffusion model described above. Indeed, the firing dynamics of these PPC neurons appeared to match the dynamics of the drift diffusion model latent decision variable (Gold and Shadlen, 2007). From this perspective, we can interpret the PPC firing dynamics as integrating evidence for a certain decision. Furthermore, these neurons appeared to stop integrating once a decision bound had been reached, triggering the associated motor response. This isomorphism between the decision variable of the drift diffusion model and the firing rates of PPC neurons is a remarkable example of neuronal activity matching the predictions of a model. Accordingly, these data provided compelling evidence that the sensory-to-motor transformation occurs within the PPC.

While compelling, these observations alone do not reveal the mechanisms of the underlying transformation. For instance, the slowly integrating, bounded dynamics of PPC neurons recorded by Roitman and Shadlen (2002) do not match the well-studied neurophysiological properties of individual cortical neurons (Fig. 4A) (Gonzalez-Burgos et al., 2019). This discrepancy suggests the involvement of an emergent network property. Insights into the potential network mechanisms were revealed in the modeling studies of Wong and Wang (2006). These authors successfully simulated the dynamics of PPC neurons in a network of competing neuronal ensembles (Fig. 4B) (Wong and Wang, 2006). Each ensemble contained recurrently connected excitatory neurons receiving distinct sensory inputs. Through inhibitory interneurons, activation of one ensemble suppressed competing ensembles via lateral inhibition. In a two-ensemble model (reflecting the two stimulus-response associations of the task), the model simulated early stimulus sensitivity, ramping firing proportional to the stimulus strength, and convergence to a peak firing rate. Importantly, by applying dynamical systems analyses, these authors could explain the mechanisms underlying the activity dynamics. With the onset of a stimulus, the network contained two basins of attraction, reflecting the ‘winner-take-all’ firing dynamics of each ensemble (Fig. 4C). The rate of integration proportional to stimulus strength could be understood as the time it took for the network to fall into each basin of attraction. Convergence to a specific firing rate irrespective of the stimulus strength could be understood as reaching the fixed point of the attractor basin. Thus, the emergent properties of a winner-take-all network are entirely capable of converting graded, sensory stimuli into a categorical outcomes, the key transformation step in a SMT. Furthermore, the Wong and Wang (2006) computational model captures key neurophysiological features of cortical circuits, including receptive field-matched connectivity of excitatory neurons and strong lateral inhibition (Fig. 4D-F).

Figure 4: Integration-to-bound as an emergent property of neural ensembles.

(A) Firing properties of a layer 3 pyramidal neuron from the posterior parietal cortex (PPC) of a rhesus monkey, recorded in an acute slice preparation. Current injections (green traces) evoke transient spiking that terminates with the end of the current step. This transient behavior is unlike PPC activity during performance in the directional discrimination task, in which neurons display integration-like persistent activity. (B) Illustration of two ensembles of excitatory neurons (green), defined by their shared sensory inputs, and a common pool of inhibitory neurons (red). (C) State-space analyses are used to visualize the stability points of network activity. In this plot, the axes reflect the internal activity of each ensemble. The dots reflect different potential stability points depending on the configuration of the network. Lower left corner (black), low persistent activity in both ensembles. Upper right corner (red), high persistent activity in both ensembles. Upper left and lower right (blue), a state in which only one ensemble at a time may display persistent activity. (D-F) Network synaptic properties implementing the dynamics in [C]. (D) With low recurrent excitation the network lacks sufficient positive feedback to sustain activity in the absence of external input, resulting in low persistent activity. This is the condition of the PPC neuron in the acute slide preparation in [A]. (E) With high recurrent excitation, both ensembles can display high persistent activity. (F) The introduction of strong lateral/feedback inhibition evokes competition between the ensembles, such that persistent activity in one ensemble cross-suppresses the other ensemble. These competitive dynamics are compatible with the drift-diffusion model, in which the decision variable approaches one of two opposing bounds. For a quantitative description of these networks see Wong and Wang 2006 and Zagha et al., 2016. Panel [A] is adapted from Gonzales-Burgos et al. 2019, licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). Schematic elements were created with BioRender.com.

These recording and modeling experiments provided our first mechanistic answers to where and how an SMT might occur (Fig. 5). Early sensory regions encode the sensory evidence for a task-relevant stimulus feature, which is then passed on to higher order cortical regions. Some of these higher order networks organize into winner-take-all neural ensembles. The task-relevant sensory input evokes competition between these ensembles. Eventually one ensemble ‘wins’ via recurrent excitation and lateral inhibition, thereby resolving the competition into a categorical outcome. As the ensemble activity approaches its fixed point, the outcome of the competition is propagated to motor-related regions to trigger the associated motor response.

Figure 5: Serial transformation hypothesis.

Coordination of sensory, associational, and motor cortices in implementing a goal-directed, sensory-to-motor transformation.

5. Recent observations challenging the traditional perspective

The traditional perspective described above remains the starting point for understanding goal-directed SMTs. However, more recent investigations have challenged the completeness, generalizability, and validity of this perspective. Below, we identify two broad categories of challenges to the traditional perspective.

5A. Serial vs parallel processing.

The recordings and analyses described above suggest specific functions of connected brain regions in the serial processing of sensory-to-motor signals. In particular, the dynamics within the PPC appear to be uniquely suited to performing the transformation from graded sensory to categorical outcome signaling. Therefore, it came as a great surprise to the field when inactivations of PPC in rhesus macaques did not impair performance on the directional discrimination task (Katz et al., 2016), or that their impairments were rapidly compensated (Jeurissen et al., 2022). Simultaneously, similar findings were observed in a rat evidence accumulation task. While robust integration-like dynamics in PPC neurons were observed during task performance (Hanks et al., 2015), inactivating PPC yielded minimal performance impairments (Erlich et al., 2015). These dissociations of representation (assessed by recordings) and function (assessed by inactivations) complicate the serial transformation hypothesis. At a minimum, they suggest that other brain regions can rapidly compensate for PPC inactivation. Indeed, evidence integration-like dynamics have now been recorded in multiple cortical and subcortical regions (DePasquale et al., 2024). These observations suggest two possible organizations. 1) Parallel processing: SMT signaling is distributed throughout the brain in partially redundant networks (Esmaeili et al., 2020; Le Merre et al., 2018). 2) Serial processing with distributed representation: essential nodes of the SMT do exist, yet their activities propagate broadly to other brain regions that are not required for task performance. Interestingly, another study using the rat evidence accumulation task identified the dorsal striatum as essential for evidence integration (Yartsev et al., 2018). If so, this would require an entirely different neurobiological mechanism than the one proposed by Wong and Wang (2006), given the vastly different synaptic organizations of the dorsal striatum compared to the PPC.

Modern technologies are further revealing the distributed nature of sensory and motor signaling throughout the brain. One-photon and two-photon imaging with large fields of view and Neuropixels probes with >1000 recording sites now allow for the recording of multiple brain regions simultaneously with single trial (if not single unit) resolution. The consensus from these recordings has been the wide distribution of sensory, integration, and motor signals throughout the brain (Allen et al., 2017; International Brain et al., 2025; Khilkevich et al., 2024; Luo et al., 2025; Musall et al., 2019; Salkoff et al., 2020; Steinmetz et al., 2019; Stringer et al., 2019; Zagha et al., 2022). Zatka-Haas et al. (2021) demonstrated widespread sensory-related and motor-related signals throughout dorsal cortex in a visual discrimination task (Zatka-Haas et al., 2021). Interestingly, the magnitude of the behavioral effects of inactivating these cortical regions was proportional to their sensory encoding, suggesting common functions of dorsal cortex (including frontal cortices) in relaying sensory information to subcortical regions for action selection. The importance of these distributed signals, and their implications for serial vs parallel processing, are currently unknown (see section 6D below).

5B. Cortical-dependence vs cortical-independence.

Within the traditional perspective described above, neocortex is central to goal-directed SMTs. And yet, recent studies have added layers of complexity to this perspective. In a mouse joystick press task, Hwang et al. (2019) tested the representation and function of primary motor cortex (M1) across learning and expert task performance. In this task, mice learned to respond to an auditory go cue by moving a joystick into a target zone. Performance in this task stabilized after 20-25 training sessions. Interestingly, M1 inactivations caused severe impairments in task performance during the learning (sessions 4-8) and acquisition (sessions 20-25) phases. However, when tested well after expert status was achieved (sessions 61-69), M1 inactivations did not impair task performance (Hwang et al., 2019). This decrease in causality was associated with a decrease in the percentage of M1 motor-signaling neurons. This and related studies (e.g. (Kawai et al., 2015; Mizes et al., 2024)) support a framework in which the neocortex is essential for learning new sensory-motor associations, but, once a task is acquired, these associations become cortically-independent. Subsequent to ‘automatization’ or ‘habit formation’, performance may rely exclusively on subcortical structures such as the brainstem, cerebellum, and basal ganglia (Yin et al., 2009), all of which have their own sensory-to-motor signaling pathways. From this perspective, the essential role of neocortex during task learning may be to instruct subcortical networks to perform the goal-directed SMTs independently.

And yet, the nuance continues. Hong et al. (2018) tested the effects of primary somatosensory cortex (S1) chronic lesions (aspirations) on performance in a variant of the mouse whisker object localization task described above. In expertly performing mice, S1 lesions caused a transient reduction in task performance, which recovered by the second session post-lesion. This could be interpreted as task performance becoming cortically-independent after learning. However, in another cohort of mice, S1 was lesioned before task training. Remarkably, these S1 lesioned mice learned the object localization task equally well as non-lesioned controls (Hong et al., 2018). Thus, for this task, S1 may be functionally redundant both during task learning and after task acquisition. Interestingly, this study also demonstrated that expert performance was persistently impaired with additional damage to the dorsolateral striatum, suggesting the functional necessity of these subcortical circuits in stimulus detection (see also (Sippy et al., 2021; Sippy et al., 2015; Zareian et al., 2023)).

What determines whether a task is cortically-dependent? Species, sensory modality, stage of training, task complexity, and task shifting (e.g., changes in stimulus-response contingencies) are all likely to influence the extent to which neocortical networks are indispensable for task performance. Whether these variables converge to a smaller set of common factors, such as cognitive load (Stuttgen and Schwarz, 2018), remains to be determined.

6. What remains to be discovered

Decades of research have provided essential frameworks and empirical insights into the neurobiological mechanisms underlying goal-directed SMTs. And yet, as concluded by David Hubel in his 1981 Nobel lecture regarding whether the well was dry for novel discoveries of striate cortex: “it is raining, and raining hard” (Hubel, 1993). We end this review with a discussion of important SMTs topics that we believe are ripe for discovery.

6A: Dynamic linkage between sensory and motor networks.

A hallmark of goal-directed SMTs is the ability to establish any arbitrary sensory-motor association, regardless of its intrinsic or ethological validity. Licking following a left whisker deflection, moving a joystick after an auditory cue, pressing a button after a vibratory stimulus: these are some of the sensory-motor associations used in the studies described above, and a small fraction of the sensory-motor associations implemented throughout the literature. Furthermore, these sensory-motor associations are dynamic, rapidly forming and dissolving according to task context and subjective engagement (Harris and Thiele, 2011; Matteucci et al., 2022). It is currently unknown how these dynamic associations are established.

We consider separately two component processes of a dynamic sensory-motor association. The first component process is selecting the goal-relevant target stimulus. To gain insight into this process, Zagha and colleagues have been studying neuronal signaling in a whisker-based selective detection task. In this Go/NoGo task, mice learn to lick on trials with a left whisker stimulus (target) and not lick on trials with a right whisker stimulus (distractor) (with Go-NoGo assignments counterbalanced across mice). How does the brain solve the problem of responding to left whisker stimuli while ignoring right whisker stimuli? Using widefield Ca²⁺ imaging, Aruljothi et al., (2020) compared neural activity throughout dorsal neocortex on target trials vs distractor trials. Notably, both target and distractor stimuli robustly activated primary somatosensory cortex (S1), within their expected somatotopy. However, the propagation of these signals to frontal cortex was dramatically different. Target-evoked responses effectively propagated to frontal cortex, whereas distractor-evoked responses did not (Aruljothi et al., 2020). Importantly, the propagation of target stimuli to frontal cortex occurred before the onset of licking and other response-related movements (Zareian et al., 2021; Zhang and Zagha, 2023), and therefore cannot be a consequence of movement on ‘Go’ trials. These findings in mice resemble previous findings in primates, of progressive attenuation of distractor stimulus propagation along the cortical hierarchy (Moran and Desimone, 1985; Tootell et al., 1998; Treue, 2001).

These findings demonstrate that expertly performing subjects regulate the propagation of target and distractor sensory signals according to task rules. And yet, the mechanisms underlying this differential propagation are unknown and likely to be at the core of sensory selection. One possible contribution is receptive field remapping. In both primate and mouse models, neurons in sensory cortex shift their receptive fields towards the target stimulus (Connor et al., 1997; Moran and Desimone, 1985; Ramamurthy et al., 2025; Womelsdorf et al., 2006; Womelsdorf et al., 2008; Zhang and Zagha, 2023). For neurons aligned to the target stimulus, they became more selective (less responsive to distractor stimuli), whereas neurons aligned to the distractor stimulus became less selective (more responsive to target stimuli).

The ubiquity of this functional plasticity across both species and tasks suggests that it is a general adaptation for sensory selection. And yet, its function remains unknown. One possibility is that receptive field remapping establishes bidirectional gain to bias the activation of downstream networks. Furthermore, the mechanisms underlying receptive field remapping are poorly understood. Zhang and Zagha (2023) demonstrated that top-down signals from motor cortex to S1 contribute to the decreased propagation of distractor stimuli, potentially mediated by the proactive activation of inhibitory interneurons within the distractor S1 region (Zhang and Zagha, 2023). Overall, these findings support a framework of top-down mediated receptive field remapping biasing stimulus-evoked activations within sensory cortex, thereby modulating their propagation to downstream regions. Furthermore, this framework demonstrates how brain regions may contribute to sensory selection without directly encoding sensory or motor features. Frontal cortex in particular may bias sensory signaling to support cortico-striatal mechanisms of sensory selection (Rodgers and DeWeese, 2014; Schneider et al., 2014; Schneider et al., 2018; Sippy et al., 2021; Sippy et al., 2015; Znamenskiy and Zador, 2013).

Many other mechanisms likely contribute to target vs distractor sensory selection. For example, target stimuli evoke late depolarizations in sensory cortex, which contribute to response triggering (Sachidhanandam et al., 2013). This late depolarization may involve loops through higher order sensory cortices (Kwon et al., 2016; Yang et al., 2016), with potential modulations by top-down signals from frontal cortex (Liu et al., 2020; Zagha, 2020). Another potential mechanism of differential target vs distractor stimulus propagation is modulations of inter-areal coherence, thereby modulating the effective connectivity between lower order and higher order cortices (Bosman et al., 2012; Fries, 2015). Other mechanisms outside of sensory cortex likely contribute, including context-dependent integration via rotation of the decision axis in the frontal cortex (Mante et al., 2013) and selection of the current rule in the basal ganglia (Frank et al., 2001). It remains to be determined the relative contributions of each mechanism to stimulus selection across tasks and species.

The second component process is selecting the appropriate action. Once a target stimulus is detected, how does it activate the associated motor response (i.e., licking, rather than reaching or running)? The short answer is that this is currently unknown. One possibility is that motor selection is mediated by the dorsal striatum. As we have seen throughout this review, the dorsal striatum appears to play essential roles in goal-directed SMTs. Notably, there are well-developed frameworks for understanding the contributions of the dorsal striatum to ‘action selection’ (Mink, 1996; Nelson and Kreitzer, 2014; Redgrave et al., 1999). The general concept is that the dorsal striatum receives motor plans from the motor cortex and selects the motor plan that is most valued at that moment in time. Motor selection occurs by propagating the selected motor plan through the direct pathway back up to motor cortex via thalamus. Motor plans for singular actions (e.g., for Go/NoGo tasks) or multiple actions (e.g., for two-alternative forced choice tasks) may be deployed even before stimulus onset (Cisek and Kalaska, 2010). One intriguing possibility is that for goal-directed SMTs, the target sensory evidence initiates striatal-mediated action selection of the prepared motor plan. Whether this occurs, and at which stage(s) of training, remains to be investigated.

6B: From evidence integration to response triggering.

The neural processes underlying SMTs are not continuously ‘online’. For example, if a subject learns that reward is unavailable (such as the withdrawal of a lickport) they will quickly (within a few trials) withhold responding. Furthermore, the sensory evidence integrators (e.g., in a drift diffusion model framework) do not appear to integrate between trials. Modeling work predicts the involvement of a filter or gate to prevent evidence integrators from drifting before trial onset (Purcell et al., 2010). One possible solution to prevent pre-trial integration is the inclusion of a proactive ‘fixation’ ensemble (a concept related to the modeling work of (Lo et al., 2009)). In addition to the stimulus-integration ensembles described above, another ‘fixation’ ensemble may be active during the intertrial interval, inhibiting stimulus integration until a trial onset cue. In a mouse whisker detection task, Zagha et al. (2015) observed a substantial population of motor cortex neurons that decreased their firing before response onset, consistent with a ‘fixation’ ensemble. Moreover, these potential ‘fixation’ neurons were anti-correlated with intermixed motor cortex neurons carrying sensory-motor and motor signals (Zagha et al., 2015). Modeling studies have established predictions about the roles of excitatory and inhibitory cell types in this type of functional circuit (Lo et al., 2009; Wong and Wang, 2006; Zagha et al., 2015). Future studies testing these predictions have the potential to link cell type to circuit motif to behavior.

Another fascinating aspect of goal-directed SMTs is that evidence accumulation is often found in the same brain regions that are responsible for motor triggering. For example, evidence integration is robust in the frontal eye fields, a neocortical region defined by its ability to trigger directional saccades. However, these two processes can be dissociated; that is, multiple studies have demonstrated that evidence integration in the frontal eye fields does not trigger eye movements until a decision is reached. Kaufman et al. (2014) provided a potential solution: activity in a motor region will not trigger movement if that activity exists in a ‘movement-null’ subspace (Churchland and Shenoy, 2024; Kaufman et al., 2014). A recent study of visual change detection in mice provided additional support for this solution (Khilkevich et al., 2024). They found robust evidence accumulation signals in secondary motor cortex, which resided predominantly in a ‘movement-null’ subspace. At response onset, the motor cortex activity dynamics abruptly transitioned into a ‘movement’ subspace.

How do these subspaces map onto neuronal populations, and what triggers the transition between subspaces? These questions remain largely unanswered. However, we do have some key insights from studies of Svoboda and colleagues using the mouse whisker object localization task described above. In these studies, the authors investigated the transition from motor preparation during a post-stimulus delay to motor triggering during the response epoch, which was signaled by an auditory ‘Go’ cue. A study by Economo et al. (2018) demonstrated that the motor cortex neurons signaling preparatory activity were largely distinct from the motor cortex neurons signaling the response. The former population preferentially projected to the thalamus (and avoided medullary motor nuclei) whereas the latter population preferentially projected to medullary motor nuclei (and avoided the thalamus) (Economo et al., 2018). Interestingly, these projection-specific pyramidal neuron subtypes also had distinct transcriptional profiles, providing a compelling linkage between cell identity and network function. In a more recent study, Inagaki et al. (2022) addressed the mechanisms of transition from preparation to response. They identified a pathway that projects the auditory ‘Go’ cue to motor cortex to trigger the stimulus-associated response, which, interestingly, was distinct from the canonical ascending auditory pathway (Inagaki et al., 2022). Relatedly, Brody and colleagues combined large-scale neural recordings with deep-learning computational methods to reveal the neural dynamics along the transition from sensory accumulation to decision formation, without an explicit ‘Go’ cue. Early in a trial, neural activity propagates along an evidence accumulation trajectory driven by sensory inputs. At the point of decision commitment, neural activity transitions into an orthogonal trajectory driven by intrinsic activity (Luo et al., 2025). A fascinating future direction will be to determine the neurobiological mechanisms underlying these neural dynamics.

6C: Neuromodulation of SMTs.

In cognitively demanding environments, performance is highly affected by arousal levels, a phenomenon described by the Yerkes-Dodson inverted U-shaped relationship (Hebb, 1955; Yerkes and Dodson, 1908). This relationship is characterized by low motivation at one extreme and hyperarousal at the other, with optimal behavioral performance occurring at intermediate arousal levels. The brain states underlying these arousal profiles are driven in large part by neuromodulatory inputs (Lee and Dan, 2012; McCormick et al., 2020), which profoundly influence virtually all aspects of the SMT. Below, we briefly review some of the neuromodulation literature as it relates to SMTs and identify opportunities for future discovery.

Sensory selection is highly dependent on arousal and attention, with notable contributions by cholinergic and adrenergic neuromodulation (Aston-Jones and Cohen, 2005; Sarter et al., 2005). For example, acetylcholine, produced in the basal forebrain, enhances stimulus encoding and inter-areal communications (Gil et al., 1997; Howe et al., 2017; Pinto et al., 2013). Norepinephrine, produced in the locus coeruleus, is also found to enhance stimulus encoding in sensory areas (Devilbiss et al., 2006; McBurney-Lin et al., 2019). According to the Adaptive Gain Theory, tonic locus coeruleus activity modulates arousal levels, whereas stimulus-evoked phasic bursting activity supports sensory gating (Aston-Jones and Cohen, 2005). Basal forebrain and locus coeruleus receive dense projections from the prefrontal cortex (Arnsten and Goldman-Rakic, 1984; Mesulam and Mufson, 1984), which may provide the instructive signals for top-down modulation of sensory selection. One open question is the spatial resolution of basal forebrain and locus coeruleus inputs into the dorsal cortex. While traditionally believed to be broadly projecting, new research shows considerable anatomical and functional specificity (Ananth et al., 2023; Poe et al., 2020). Whether this resolution is sufficient to contribute to the receptive field modulations described above (section 6A) remains to be determined.

Neuromodulatory signals also play important roles in evidence integration and decision thresholding. Increases in phasic arousal, indexed by both pupil size and neuromodulatory activity, correlate with reduced decision bias (increased tendency to respond) (de Gee et al., 2017). Relatedly, in a visual evidence accumulation task, global increases in tonic catecholamine levels were found to enhance drift rate within a drift diffusion model framework (Nuiten et al., 2023). In a visual sustained attention task in rats, selective removal of cholinergic input significantly reduced cue-induced increases in PPC firing rates (Broussard et al., 2009). And yet, dopamine was found to reduce PPC firing rates during a visual selection attention task in macaques (van Kempen et al., 2022). Extrapolating across tasks and species, it appears that a precise balance among neuromodulatory systems is required to achieve optimal decision formation. Additionally, neuromodulatory systems are critical for switching between stimulus-response associations. Elevated norepinephrine activity shifts subjects into an exploratory state, promoting disengagement from the original task contingency (Koralek and Costa, 2021; McBurney-Lin et al., 2022). This is proposed to occur via the Network Reset Hypothesis, whereby norepinephrine disrupts ongoing network activity and functional configurations (Bouret and Sara, 2005). Relatedly, acetylcholine has been implicated in cognitive flexibility. Striatal acetylcholine activity facilitates the initiation of new behavioral responses whereas prefrontal acetylcholine activity may support the suppression of stimulus-response associations from the original task contingency (Prado et al., 2017).

Emerging research tools include cell-type-specific circuit mapping and neurotransmitter-specific biosensors (Sabatini and Tian, 2020). These tools will help address key questions regarding the substrates of neuromodulatory control (e.g., recurrent excitation, lateral inhibition, maintenance of dynamical regimes) and the coordination among neuromodulatory systems. Furthermore, the mechanisms of neuromodulatory control remains an active area of investigation. This includes a growing appreciation for neuromodulatory signaling through glial cells to influence neuronal function and behavior (Kofuji and Araque, 2021; Lefton et al., 2024; Sriram et al., 2024; Xin et al., 2025).

6D: Functions of distributed processing.

As described throughout this review, the sensory, integration, and motor signals during sensory-motor tasks are not localized to individual brain regions but emerge simultaneously in multiple brain regions. Consequently, one cannot infer that a brain region is indispensable for a task by recording data alone, and even causal studies must be interpreted with caution. What are the functions of this massively distributed signaling? While currently unknown, we propose possible functions below, which themselves are not mutually exclusive.

• Distributed signaling and functional redundancy are unavoidable consequences of highly interconnected brain regions. A benefit of this redundancy is robustness to damage of individual brain regions.

• The apparent redundancy is due to the low dimensionality of most laboratory tasks. Each brain region is adapted to make unique contributions to goal-directed SMTs. With more complex and ethological tasks, these unique functions may be revealed.

• Distributed signaling may be a substrate for learning and consolidating new tasks. For example, dynamic sensory-motor associations may at first be sparsely encoded within the neocortex, but with learning distribute throughout subcortical structures for consolidation into sensory-motor habits.

• Distributed signaling enables region-specific computations to be conducted with the appropriate context. An example of this is the cancellation of reafferent sensory inputs via corollary discharges of motor commands, thereby enabling externally-generated sensory evidence to be distinguished from self-generated sensory signals.

• Lastly, distributed signaling may reflect a limitation of the SMT framework. While sensory and motor activities can be operationally dissociated, this does not necessarily imply that they constitute fundamentally distinct processes at the level of neural implementation. Competing theories instead favor a more integrated view, in which ‘sensory’ and ‘motor’ reflect different projections of a unified system (Rao, 2024; Rigotti et al., 2013; Wolpert et al., 1995). Within these frameworks, sensory-motor transformations are not localized to discrete spatio-temporal stages but instead unfold across distributed, recurrent circuit interactions.

Future investigations, testing specific hypotheses, are poised to make fundamental discoveries about the nature of distributed neuronal signaling. Such discoveries will have implications well beyond the specific topic of goal-directed SMTs.

7: General conclusion

Over the past several decades, studies of goal-directed SMTs have uncovered fundamental insights into how the brain links sensation to action. The emerging picture is one of dynamic, distributed, and flexible networks that coordinate sensory selection, evidence integration, and motor execution across both cortical and subcortical regions. And yet, the underlying neurobiological mechanisms are far from resolved, with plenty of exciting frontiers ahead. Ongoing and future work promise to bridge insights from single-cell dynamics and circuit motifs to behavior and learning, while integrating computational frameworks that capture the brain’s adaptive nature. A central challenge moving forward will be to determine how flexible/adaptive sensory-motor associations are constructed, stabilized, and reorganized as behavioral demands change, and how these processes generalize across species, modalities, and task domains.

Highlights.

• Linking neural activity to behavior requires recording, perturbation, and computation

• Traditional models frame sensory-to-motor transformations as serial propagations across neocortex

• Emerging evidence supports parallel, distributed sensory and motor signaling

• Circuits underlying sensory-to-motor transformations shift with learning

• How neocortical and subcortical circuits coordinate flexible stimulus-response associations remains to be discovered