Orbitofrontal Cortex Encodes Decision Signals during Value-based But Not Perceptual Decisions

Abstract

A fundamental question in neuroscience is whether the brain uses specialized subsystems for different types of decisions or relies on a unified decision-making network. The OFC provides an ideal test case for this question: It has a well-established role in value-based decisions, but it remains unknown whether this reflects functional specialization or participation in a broader, general decision network. To distinguish between these possibilities, we used Neuropixels to monitor large ensembles of OFC neurons as a monkey performed both value-based and perceptual decision tasks. Consistent with prior reports, OFC robustly encoded decision-relevant information during value-based decisions. In contrast, during the perceptual task, OFC did not carry decision-relevant signals at either the single neuron or population level. This highlights the specialization of the OFC for value-based decisions and suggests that the different cognitive demands underlying different types of decisions recruit at least partially distinct neural circuits.

INTRODUCTION

Different types of decisions place fundamentally different demands on cognitive processing. Value-based decisions, such as choosing between different food options or investment strategies, require the brain to compute and compare the subjective worth of alternatives. In contrast, perceptual decisions, such as categorizing the direction of motion or identifying an object, involve extracting and reporting information from sensory input without necessarily weighing subjective preferences. Interestingly, human patients with selective ventromedial frontal damage show impaired preference judgments while their perceptual judgments were intact (Fellows & Farah, 2007). In contrast, patients with lateral prefrontal damage (as well as healthy controls) were unimpaired in both preference and perceptual judgments. This raises the possibility that different neural circuits are specialized for different types of decisions.

Value-based and perceptual decisions have been explored by largely parallel, nonoverlapping fields within neuroscience. Studies of value-based decisions grew out of behavioral economics (Kahneman, 2011; Padoa-Schioppa & Assad, 2006), whereas studies of perceptual decisions emerged from basic vision research (Salzman, Britten, & Newsome, 1990; Newsome, Britten, & Movshon, 1989). Consequently, these fields have adopted different canonical behavioral tasks and emphasized critical roles for different brain areas. The canonical task to study perceptual decisions involves discriminating the direction of a cloud of swirling dots (random dot kinematograms; Gold & Shadlen, 2007). Performance in motion discrimination tasks is well described by drift diffusion models, which posit that noisy sensory evidence accumulates over time until reaching a decision threshold. The neural correlate of this accumulation process—ramping activity that builds until the threshold is reached—was initially identified in the lateral intraparietal sulcus (Shadlen & Newsome, 2001). Similar ramping signals during perceptual decisions have since been reported across many brain regions (Goueytes et al., 2025; Peixoto et al., 2021; Gold & Shadlen, 2007; Horwitz & Newsome, 1999; Kim & Shadlen, 1999).

Value-based decisions are typically studied by asking participants to choose between pairs of options that yield different amounts of reward (Elston & Wallis, 2025a; Balewski, Elston, Knudsen, & Wallis, 2023; Kennerley, Behrens, & Wallis, 2011; Padoa-Schioppa & Assad, 2006). Typically, there is no perceptual ambiguity about what the choice options are. Thus, rather than asking participants to decide what they are looking at (“How certain am I that the dots are moving to the left?”), value-based choices depend on how participants estimate and compare the subjective value of the choice options (“Is option A better than option B?”). The OFC plays a causal role in such decisions, as its damage or inactivation impairs value-based choices in both humans and animals (Gore et al., 2023; Ballesta, Shi, Conen, & Padoa-Schioppa, 2020; Knudsen & Wallis, 2020; Rudebeck & Murray, 2008; Izquierdo, Suda, & Murray, 2005; Hornak et al., 2004; Bechara, Damasio, Damasio, & Anderson, 1994; Eslinger & Damasio, 1985). Neurophysiological studies revealed that OFC neurons encode the values of choice options during value-based decisions (Elston & Wallis, 2025a; McGinty & Lupkin, 2023; Padoa-Schioppa & Assad, 2006), and recent work has shown that targeted OFC electrical microstimulation can alter the values participants assign to options (Ballesta et al., 2020), leading them to prefer one option or another, as determined by the experimenter. Although there is strong evidence for OFC playing a causal role in value-based decisions, to our knowledge, OFC's role in perceptual decisions has yet to be explored—likely due to the lack of overlap in the traditions of value-based and perceptual decision neuroscience.

Here, we ask whether OFC is functionally specialized for value-based decisions or whether it plays a more general role across multiple decision types. If OFC is specialized for value-based decisions, then its neurons should carry decision-relevant signals only during value-based decisions but not during other types of choices. Alternatively, if OFC plays a general role in decision-making, then it should carry decision-relevant signals across a variety of decision types. To distinguish between these possibilities, we recorded neural activity from the OFC of a nonhuman primate trained to perform both value-based and perceptual decision tasks.

METHODS

Experimental Subject Details

All procedures were carried out as specified in the National Research council guidelines and approved by the Animal Care and Use Committee at the University of California, Berkeley. Both studies were performed on the same male rhesus macaque (Subject K) aged 5 years and weighing 12 kg at the time of recordings. During the experiment, the subject sat head-fixed in a primate chair (Crist Instruments) and eye movements were tracked with an infrared system (SR Research). Stimulus presentation and behavioral conditions were controlled using the NIMH MonkeyLogic MATLAB toolbox. The subject was implanted with a unilateral recording chamber that allowed access to the OFC.

State-dependent Valuation Task

Subject K was trained to perform a state-dependent value-based choice task involving eight reward-predictive pictures (Elston & Wallis, 2025a, 2025b; Figure 1A). The value mappings associated with each image depending on the cued task state. Stimuli were presented on a monitor positioned approximately 30 cm from the monkey. The subject self-initiated trials by fixating on a white dot for 500 msec. Following fixation, a state cue indicated whether the upcoming choice should be evaluated according to the value scheme associated with State A, B, or C. Two distinct cues were used for each state to dissociate neuronal responses to visual properties from cue meaning. After 500 msec, the state cue disappeared, and the subject maintained central fixation for a further 100 msec. Then, the choice options were presented. On 75% of trials, two options were shown (free choice trials), whereas on 25% of the trials, there was only one option (forced-choice trials). The presence of the forced-choice trials was to ensure that all reward options were regularly experienced. The monkey reported his decisions by shifting their gaze to the chosen target location and fixating for 300 msec. Reward values corresponded to different amounts of apple juice delivered via a peristaltic pump.

Figure 1. .

Task designs and behavioral performance. (A) Value-based choice task design. The subject initially fixated and was then cued as to which task state the current trial would take place in. Then, after a delay, the subject was presented with either one (forced choices) or two (free choices) options. In States A and B, the rewards associated with the same stimuli changed according to the cued task state; in State C, the rewards associated with the choice options remained constant. We used two cues per task state to decorrelate neural responses to the sensory particulars of the stimuli from the meaning of the task state. Seventy-five percent of the trials were free choices, where the subject chose between two options. Twenty-five percent of the trials were forced choices, where only one option was presented and the subject was required to make a saccade to it. (B) Perceptual decision task design. The subject initially fixated and was then cued to attend to either the speed or direction of the upcoming motion stimulus. After a brief delay, a motion stimulus was presented. After a final delay, the subject reported whether the dots were moving up or down (if cued to attend to direction) or whether the dots were moving fast or slow (if cued to attend to speed). Different categories of motion were associated with different reward levels such that correctly categorizing “fast” and “downward” motion paid five drops of juice whereas correctly categorizing “slow” and “upward” motion paid only three drops of juice. Incorrect categorizations yielded no reward. We used two cues per task state to decorrelate neural responses to the sensory particulars of the stimuli from the meaning of the task state. (C) Decision accuracy as a function of option value differences in the value-based choice task. Thin lines are single sessions (n = 4), and thick lines are mean values. The subject's choice was well predicted by the difference between values. (D) Decision times in the value-based choice task are influenced by the expected reward. Faster RT occurred in those trials with a greater value difference. (E) Choice performance is not affected by the expected reward in the perceptual decision task. Thin lines are single sessions (n = 4), and thick lines are mean values. (F) Decision times in the perceptual task are influenced by the expected reward. The subject responded faster when correctly reporting a high-reward category.

For States A and B, the same choice stimuli were used, but the associated rewards differed depending on the cued state. For example, during a trial cued for State A, the “postbox” target would yield one drop of juice, whereas if State B were cued, it would instead be worth four drops of juice. State cues were pseudorandomized across trials. State C used a different set of choice options and had its own unique value schema that was not in conflict with the others. This control condition allowed for us to observe the decision processing without a conflicting value scheme.

Perceptual Categorization Task

The subject was also trained on a perceptual decision-making task, which involved categorizing either the speed or direction of motion in a random-dot kinematogram (hence “motion stimulus”; Figure 1B). Analogous to the state-dependent valuation task, subjects were cued whether to categorize according to speed or motion on each trial. Subjects initiated trials by fixating on a small white dot for 500 msec. They were then shown a cue for 600 msec indicating whether to attend to the speed or direction of the upcoming motion stimulus. Like the state-dependent valuation task, two cues were used to indicate whether to direct attention to speed or direction to dissociate neural task representations from the visual features of the cues. After the cue and a 500-msec delay, the subject viewed a large motion stimulus (10° of visual angle in radius) that varied in both speed (fast vs. slow) and direction (up vs. down). We associated each category with a different reward amount such that correctly categorizing either fast motion or downward motion led to a large juice reward (five drops of apple juice), whereas correctly categorizing either slow motion or downward motion led to a smaller reward (three drops of apple juice). Incorrect categorization led to no reward. We did this to test whether OFC encoded the expected reward associated with a given perceptual category.

Decisions were reported during a dedicated choice epoch, which followed a 500-msec delay after the motion stimulus. The monkey was trained to report fast motion and downward motion by saccading to a small black circle stimulus and, conversely, was trained to report slow motion or upward motion by saccading to a pink circle stimulus. Both speed and direction were present in every trial, and we overrepresented the “conflict” trials where the animal categorized the same motion stimulus differently depending on whether it was currently categorizing speed or direction. These “conflict” trials accounted for 70% of trials in a single behavioral session.

Behavioral Analysis

We sought to compare the behavioral performance of the subject for both tasks. We began by examining the subject's ability to choose the correct option for different value combinations across each state. For the state-dependent value task, we calculated the probability of choosing the left option as a function of the value difference between the left and right options for each state. We then used a logistic regression (LR) model where the dependent variable was a binary indicator of left/right choice and the independent variables were task state (1, 2, 3) and value difference (left minus right; ± 1–3; LR 1).

$$\mathrm{L}\mathrm{R}1:p\left(\mathit{\text{left}}\right)=\mathrm{\sigma }\left({\mathrm{\beta }}_0+{\mathrm{\beta }}_1\times \text{state}+{\mathrm{\beta }}_2\times \left|\text{Δvalue}\left|+{\mathrm{\beta }}_3\times [\text{state}\times \right|\text{Δvalue}\right|]\right)$$

For the perceptual decision-making task, we assessed choice accuracy as a function of reward amount across different attentional states (attending to either motion speed or direction). We used another logistic regression, where the dependent variable was choice accuracy (−1 = correct, 1 = correct) and predictors were task state (direction and speed, coded as −1 and 1) and reward value (three or five drops of juice, coded as −1 or 1).

$$\mathrm{L}\mathrm{R}2:p\left(\mathit{\text{correct}}\right)=\mathrm{\sigma }\left({\mathrm{\beta }}_0+{\mathrm{\beta }}_1\times \text{state}+{\mathrm{\beta }}_2\times \text{reward}+{\mathrm{\beta }}_3\times \left[\text{state}\times \text{reward}\right]\right)$$

We next analyzed how choice RTs varied as a function of state and value using general linear regression models. The model (general linear model [GLM] 1) for the state-dependent valuation task regressed RTs against state (A, B, C, coded as −1, 0, 1) and the absolute difference in values present in the trial.

$$\mathrm{G}\mathrm{L}\mathrm{M}1:\mathit{\text{Reaction time}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1\times \text{state}+{\mathrm{\beta }}_2\times \mid \text{Δvalue}\left|+{\mathrm{\beta }}_3\times [\text{state}\times \right|\text{Δvalue}\mid \left]\right)$$

The model (GLM 2) for the perceptual categorization task regressed choice RT against task state (direction or speed, coded as −1 or 1) and reward amount (three or five drops of juice, coded as −1 or 1).

$$\mathrm{G}\mathrm{L}\mathrm{M}2:\mathit{\text{Reaction time}}={\mathrm{\beta }}_0+{\beta }_1\times \text{state}+{\mathrm{\beta }}_2\times \text{reward}+{\mathrm{\beta }}_3\times \left[\text{state}\times \text{reward}\right])$$

Neurophysiological Recordings

The subject was fitted with a titanium head positioner and imaged in a 3 T magnetic resonance imaging scanner. The resulting images were used to generate 3-D reconstructions of each subject's skull and brain areas of interest. We then implanted custom, radiotranslucent recording chambers made of polyether-ether-ketone (Gray Matter Research). Each session, we acutely lower one 45-mm primate Neuropixels probes (IMEC VZW) into Area 13 of OFC. Electrode trajectories were determined by incorporating the subject's magnetic resonance imaging into computer-assisted drawing software (OnShape) to design and 3-D print custom recording grids to lower the probes. We used a Form 4 3-D printer (Form Labs). To mitigate recording drift, the probes were allowed to settle for approximately 60 min after insertion, before the start of the behavioral task. We configured the probes to record from 384 active channels in a contiguous block at the tip, allowing dense sampling of neuronal activity along a 3.84-mm span. Neuronal activity was filtered and digitized for action potential bands (300-Hz high-pass filter, 30-kHz sampling frequency) and local field potential (1-kHz low-pass filter, 2.5-kHz sampling frequency). Activity was monitored during experimental sessions and saved to disk using SpikeGLX (https://billkarsh.github.io/SpikeGLX/).

Spiking in the action-potential band was identified and sorted offline using Kilosort4 (Pachitariu, Sridhar, Pennington, & Stringer, 2024). Because each Neuropixels headstage samples the data with a slightly different system clock, it was necessary to map all physiological and task-event time series into a common timeline. We accomplished this by broadcasting a 5-Volt, 1-Hz square wave to a unique reference channel on the Neuropixel probe (Channel 385) and the nonneural acquisition board. We detected and stored the times of the rising phase of the 5-V wave on each cycle and then used regression to interpolate all task events into the timeline of the probe. To be included for further analysis, neurons had to be present for >90% of the experimental session, have a mean firing rate over the course of the entire session >1 Hz, and <0.1% of spikes occurring within 1 msec of another isolated neuron's spikes (e.g., <0.1% interspike interval violations).

We also identified and merged duplicate neurons that were missplit by Kilosort. We developed a duplicate detection algorithm based on computing the coefficient of variation (CV) of the firing rate ratio between pairs of neurons. CV quantifies the consistency of the proportional relationship between two firing rate time series. The advantage of this approach over correlation-based methods is that correlation only captures whether two neurons show similar temporal patterns of increases and decreases, regardless of their absolute firing rates. In task-based recordings, many (distinct) neurons respond to the same events (e.g., cue onset, reward delivery), making it possible to observe spuriously high correlations between pairs of neurons that are nonetheless independent units. True duplicates, however, maintain a constant scaling relationship throughout the entire trial period. If two units result from spike sorting errors that missplit the same neuron, one should be a fixed multiple of the other at every time point (e.g., Unit B = 0.8 × Unit A consistently across all task phases). The CV of the firing rate ratio directly tests for this constant-scaling property, allowing us to distinguish duplicates from distinct neurons that exhibit task-driven correlation.

Our procedure was as follows: First, we calculated trial-averaged, z-scored firing rates for each neuron. Then, for each recording session, we estimated the CV of the firing rate ratio for all neuron pairs separated by at least 20 channels (200 μm). We chose this spatial separation because neurons this distant are highly unlikely to be duplicates: (1) Kilosort incorporates probe geometry during spike sorting, making cross-channel assignment errors at this distance implausible, and (2) our prior experience manually sorting v-probe recordings with 100-μm contact spacing indicated negligible spike waveform overlap across channels. However, as noted above, these distant pairs may still exhibit correlated activity if they respond to common task events, making them an appropriate reference for distinguishing task-driven correlation from true duplication. This procedure yielded a session-specific empirical null distribution of ratio CVs representing the expected variability for distinct neurons in that recording.

We then computed ratio CVs for all spatially nearby neuron pairs—those within 80 μm of one another, the spatial range where duplicate missplitting could plausibly occur. For each nearby pair, we calculated the firing rate ratio at each time bin

$$\text{ratio}\left[\mathrm{t}\right]=\mathrm{F}\mathrm{R}\text{\_}\mathrm{j}\left[\mathrm{t}\right]/\mathrm{F}\mathrm{R}\text{\_}\mathrm{i}\left[\mathrm{t}\right])$$

and then computed

$$\mathrm{C}\mathrm{V}=\mathrm{s}\mathrm{t}\mathrm{d}\left(\text{ratio}\right)/\left|\text{mean}\left(\text{ratio}\right)\right|$$

Pairs with ratio CVs falling below the 5th percentile of the session-specific null distribution were flagged as potential duplicates, indicating that their scaling relationship was unusually consistent compared with typical distinct neuron pairs in that session. Visual inspection of flagged pairs confirmed that they were indeed duplicates, which were subsequently merged.

Analysis of Single Neuron Tuning

To quantify neuronal tuning during the value task, we used a sliding window ANOVA with a linear contrast for value. We first fit a general linear regression (GLM 3) where, for each neuron, we modeled firing rate (F) during 100-msec windows stepped by 25 msec as follows:

$$\mathrm{G}\mathrm{L}\mathrm{M}3:F={\mathrm{\beta }}_1\cdot \text{state}+{\mathrm{\beta }}_2\cdot \text{value}+{\mathrm{\beta }}_3\cdot \text{state}\times \text{value}+{\mathrm{\beta }}_4\cdot \text{state}\text{\_}\mathrm{c}\mathrm{u}\mathrm{e}+{\mathrm{\beta }}_5\cdot \text{trial}\text{\_}\text{number}+{\mathrm{\beta }}_0$$

where state was a categorical factor (three levels) coded using sum contrasts, value was the number of juice drops associated with the chosen option (centered and treated as a continuous variable to test for linear value coding), and state_cue indicated which visual cue was used to indicate state for that trial. We used this to help disambiguate abstract state-encoding from the visual features of the cues used to indicate each state. Trial number was a nuisance term meant to capture linear recording drift across the session.

We then computed the ANOVA on the fitted model using Type II sums of squares to obtain F statistics and p values for each factor. To ensure that our state-encoding neurons did not reflect stimulus-driven effects related to the visual state cues, we defined state-selective neurons as those which significantly differentiated the task state but did not differentiate the visual state cues. To ensure that our results were robust, we required that factors were significant (p < .01) for at least 150 msec (1.5 times the window width).

To quantify neuronal tuning during the perceptual task, we used a sliding window regression (GLM 4). We used the same 100-msec window and 25-msec step as for the value task. We modeled each neuron's firing rate (F) in each window as follows:

$$\mathrm{G}\mathrm{L}\mathrm{M}4:F={\mathrm{\beta }}_1\cdot \text{state}+{\mathrm{\beta }}_2\cdot \text{value}+{\mathrm{\beta }}_3\cdot \text{state}\times \text{value}+{\mathrm{\beta }}_4\cdot \text{state}\text{\_}\mathrm{c}\mathrm{u}\mathrm{e}+{\mathrm{\beta }}_5\cdot \text{trial}\text{\_}\text{number}+{\mathrm{\beta }}_0$$

where state was whether the subject was cued to attend to speed or direction (coded as −1 and 1) and value was the number of juice drops associated with the chosen option (coded as −1 and 1 for three and five drops, respectively). As in the value model (GLM 3), we included a state_cue term that indicated which visual cue was used to indicate state for that trial to distinguish abstract state-encoding from the visual features of the cues themselves. Accordingly, we defined state neurons as those which significantly encoded the state main effect but not the state_cue term. Trial number was a nuisance term meant to capture linear recording drift across the session. Because the perceptual task had only two task states, we did not need to compute the ANOVA table. We used the same temporal thresholding as for the data from the value task, requiring neurons to be continuously selective (p < .01) for at least 150 msec to deem them significant.

Analysis of Task Epoch Selectivity

To identify neurons that were selectively active during specific task epochs, we performed a one-way ANOVA on firing rates across the three main task epochs (cue, motion stimulus, and choice). For each neuron, we calculated the mean firing rate during a 500-msec window beginning at the onset of each epoch. We first identified neurons that were significantly modulated by task epoch (p < .01) and then defined a neuron's preferred epoch as the one which elicited its highest mean firing rate.

Population Decoding of State-dependent Values

To assess the population dynamics within OFC during each task, we trained linear discriminant analysis decoders to classify state–value conjunctions. We used a similar decoding framework for both tasks: For the state-dependent value task, the decoder classified 12 classes (4 reward values × 3 states); for the perceptual decision-making task, the decoder classified four classes (2 reward outcomes × 2 cued features). To ensure robust single-trial resolution of our decoders, we used a bootstrapping approach where, on each bootstrap, 90% of single-saccade trials were randomly selected for training. We then balanced the training data, equating the number of trials where a given value–state combination was selected. The remaining trials were used for testing. This procedure ensured that the composition of our training data did not introduce bias into the decoder. We repeated this balanced train-test split 1000 times at each time step and then averaged over instances where the same trial was present in the testing set across bootstraps. Decoder accuracy was extracted for each time point (25 msec bin) for each trial. We empirically identified the null hypothesis of these decoders (that they detect nothing but noise) by repeating our bootstrapping procedure but shuffling the training labels on each bootstrap.

RESULTS

To examine OFC's role in value-based and perceptual decisions, we trained a monkey (Subject K) to perform both a value-based choice task (“value task”) and a perceptual categorization task (“perceptual task”) while conducting Neuropixel recordings in OFC. The recordings were made sequentially with the value task occurring first and then the perceptual task 3 months later. We designed the perceptual task so that its structure paralleled the value task's (Figure 1).

In the value task, the subject was trained to choose between different visual stimuli (emojis) to maximize a juice reward associated with different juice amounts. We previously used this task to explore how the OFC encodes state-dependent values (Elston & Wallis, 2025a, 2025b). Critically, the exact amount of juice reward depended on which of three task-states was in effect. Each task state was associated with a different reward schema (Figure 1A). Each trial began with a cue indicating which of the three reward schemes (A, B, or C) the upcoming decision should be evaluated within. After a brief delay, the monkey was presented with either one (forced choice, 25% of trials) or two options (free choice, 75% of trials). For States A and B, the same choice stimuli were used, introducing conflict by altering the reward outcome depending on the cued state. In contrast, State C used a distinct set of choice stimuli—this served as a control condition where there were no conflicting value mappings.

In the perceptual task (Figure 1B), the subject was trained to categorize either the speed or direction of a dot-motion stimulus. Following an initial fixation period, the subject was cued whether to attend to (and thus categorize) the direction or speed of the upcoming motion stimulus. Following a brief delay, the subject was then shown the motion stimulus, during which the monkey had to maintain fixation. The subject was then presented with two decision targets, which allowed the animal to report what he had seen during the motion stimulus epoch. The monkey was trained to report fast motion and downward motion by saccading to a black, speckled stimulus and, conversely, was trained to report slow motion or upward motion by saccading to a pink circle stimulus. Both speed and direction were present in every trial, and we overrepresented the “conflict” trials where the animal categorized the same motion stimulus differently depending on whether it was currently attending to speed or direction. Each category was associated with a different reward amount such that correctly categorizing either fast motion or downward motion led to a large juice reward (five drops of apple juice), whereas correctly categorizing either slow motion or upward motion led to a smaller reward (three drops of apple juice). Incorrect categorization led to no reward. This design enabled us to dissociate the perceptual basis of the decision (categorizing the motion) from reward associated with the category. This was important to test whether OFC encoded the expected reward associated with a given perceptual category independently of the stimulus itself.

The subject displayed a high proficiency in both tasks (Figure 1C–D), selecting the best option 98% of the time during free choices of the value task and correctly categorizing 97% of the time in the perceptual task. In the value task, the monkey selected the higher value option more frequently regardless of spatial position and exhibited faster RT for trials with larger option value differences. In the perceptual task, choice accuracy was nearly perfect across both attentional dimensions (direction 95%, speed 98%). Logistic regression analysis showed that choice accuracy did not differ across task states (β = .4, p > .9), expected reward amount (β = .54, p > .9), or their interaction (β = .09, p > .9). However, RTs in the perceptual task were slightly faster for trials where the subject successfully classified a high-reward feature (e.g., downward motion when direction was relevant). This indicates that the subject's response behavior was influenced by expected reward even though the decision of what to report was not determined by reward.

Both tasks involved flexibly responding to the same stimuli differently and for different reward amounts, depending on the cued task state. Next, we examined single-unit responses via Neuropixel recordings in the OFC as the subject performed both tasks. We initially recorded neural activity during the value task and then immediately retrained the animal to perform the perceptual decision task. Retraining took approximately 3 months. We used the same recording chamber, overlapping recording coordinates, and the same Neuropixel probes for both sets of recordings (Figure 2A–B).

Figure 2. .

Recording locations and single neuron encoding results. (A) Coronal magnetic resonance imaging section showing the recording location in OFC for both tasks. The black and blue lines indicate electrode trajectories for the value and perceptual tasks, respectively. (B) Mean waveforms of hundreds of simultaneously recorded individual OFC neurons localized onto the surface of the Neuropixels probe. These are two representative sessions, one for each task. (C) Percentage of neurons significantly encoding state, value, or their interaction via sliding window regression. This reflects the total proportions across all sessions. During the state cue epoch, cells began encoding the state whereas value was found encoded beginning in the choice epoch. (D) Same analysis as C but for the perceptual decision task. (E) Percentage of neurons significantly encoding each factor during each session of the value task. Each line indicates the results from a single recording. The session ID numbers correspond to the order the recordings were made. (F) Same analysis as E but for the perceptual decision task. Each line indicates the results from a single recording. The session ID numbers correspond to the order the recordings were made.

We report the results of four recording sessions per task, where we obtained large numbers of well-isolated single neurons (n = 1427 neurons in the value task; n = 1321 neurons in the perceptual decision task; see Figure 2B for spike waveforms localized onto the surface of the probe for both tasks). To assess encoding of task-relevant variables, we first binned firing rates in 100-msec bins and then assessed how firing rates varied as a function of task state, the reward of the chosen option, and the interaction of these factors over time (see the Methods section). Consistent with our previous report (Elston & Wallis, 2025a), OFC was strongly engaged in the value task with neurons encoding all decision-relevant variables (state, value, state × value) of the chosen option during the choice epoch (Figure 2C). In contrast, a nonsignificant number of neurons encoded any choice-related factor during the perceptual task (Figure 2D).

OFC neurons were not silent during the perceptual task, despite not differentiating decision-relevant features. We observed robust task-related increases in firing during the motion epoch of each individual session of the perceptual task (Figure 3A). Moreover, the magnitude of these task-evoked responses was comparable with those during the choice epoch of the value task. This suggests that the lack of decision-related signals during the perceptual task is not due to changes in our recording quality across experiments but, rather, due to differences in the cognitive demands underlying value-based and perceptual decisions.

Figure 3. .

OFC neurons encode task epoch during the perceptual categorization task. (A) Mean firing rates across all OFC neurons recorded in each session of the value and perceptual tasks. Each line and shaded region denotes the mean and bootstrapped 95% confidence interval for a single session. The session ID numbers correspond to the order the recordings were made. (B) Two OFC neurons that were driven by the onset of the motion stimulus but did not differentiate decision-relevant task features. (C) Trial-averaged firing rates for populations of neurons identified as cue, motion, or choice selective. Neurons identified as significantly modulated (p < .01) by task epoch by assessing their firing rates during each task epoch with a one-way ANOVA. A neuron's preferred epoch was defined as the one which elicited the neuron's highest firing rate. These traces show the trial-averaged grand mean across neurons identified as preferring each task epoch. Thick lines denote grand means, and the shading denotes bootstrapped 95% confidence intervals. We note that the motion-selective population showed the greatest modulation.

Although OFC neurons did not encode decision-relevant variables during the perceptual task, recent theories propose that OFC contains a “cognitive map” of task space in which the current state of the task is represented (Moneta, Garvert, Heekeren, & Schuck, 2023; Knudsen & Wallis, 2022; Basu et al., 2021; Zhou, Gardner, & Schoenbaum, 2021; Schuck, Cai, Wilson, & Niv, 2016). Under this framework, each distinct phase of a task—such as the cue, stimulus, and choice epochs—constitutes a separable “state” that OFC neurons could represent. To test whether OFC encodes these task states during the perceptual task, we examined whether individual neurons showed selective activity for specific task epochs. The majority (1087/1321, 82%) showed significant modulation across task epochs (as assessed with a one-way ANOVA, p < .01; see the Methods section). We defined a neuron's preferred epoch as the one eliciting the highest mean firing rate. For example, Figure 3B shows two neurons that were driven by the onset of the motion stimulus but did not differentiate speed, direction, or expected reward. These responses cannot be attributed to purely visual or sensory processing, as the same neurons failed to respond to other stimulus-onset events (e.g., cue onset or choice target onset). Overall, of the 1087 epoch-selective neurons, 518 (47%) preferred the cue epoch, 391 (36%) preferred the motion epoch, and 178 (16%) preferred the choice epoch. As shown in Figure 3C, OFC motion-selective units exhibited the strongest responses.

Critically, however, this task structure selectivity occurred in the absence of any encoding of decision-relevant variables. Had OFC neurons encoded any value-related factors during the perceptual task, we would have expected it to do so during the motion-stimulus period, when value-related information was made available to the monkey. Neurons carrying a context-independent value code would have fired more strongly when the monkey was presented with a high-reward motion stimulus, regardless of the cued task context, for example, showing elevated firing both when a fast stimulus was shown and speed was relevant and when downward motion was shown and direction was relevant. Had OFC neurons carried a context-dependent value code in the perceptual task, we would have expected neurons to differentially signal value depending on the cued task state, for instance, firing more for downward motion when direction was relevant but being indifferent to all stimuli when speed was relevant. Instead, as illustrated in Figure 3A, neurons that were driven by the motion stimulus responded equivalently regardless of whether the animal was categorizing a high-reward or low-reward feature and regardless of whether the animal was attending to speed or direction.

Although we could not identify decision-relevant representations at the single-neuron level during the perceptual task, it is possible that such information could be represented at the population level through distributed patterns of activity across neurons. To test this, we examined whether state-dependent values could be decoded from population activity (Figure 4). We used our established procedure of decoding values with a single-trial resolution (Elston & Wallis, 2022, 2025a; Balewski et al., 2023). Briefly, we trained linear discriminant analysis classifiers to decode state–value combinations from neural activity. For the value task, a 12-way decoder was trained (4 values × 3 states). For the perceptual task, we trained a four-way linear discriminant classifier (2 reward levels × 2 attentional states). To obtain robust single-trial estimates of decoded state-dependent values, we used a bootstrapping approach where, within each bootstrap, a random 90% of the trials were used as the training set and the remaining 10% were used for testing. We repeated this procedure 1000 times and averaged over instances where the same trial appeared in the testing set multiple times. This yielded a robust measure of state–value coding at each moment during each individual trial. We also generated null distributions by shuffling the training labels and then repeating the bootstrapping procedure.

Figure 4. .

Population decoding of state-dependent value across tasks. Decoders were trained to identify the state-dependent value for each trial based on the patterns of simultaneously recorded neuronal activity. The colored lines denote decoders trained on unshuffled data whereas the gray lines denote decoders trained on data with shuffled labels. The left column shows the value-based choice task and the right column shows the perceptual decision task. Decoder performance in the value task gradually rose upon presentation of the state cue and then sharply increased upon presentation of the choice options. Decoder performance during the perceptual decision task performed comparably to the shuffled data control. Each line and shaded region reflects the mean and bootstrapped 95% confidence intervals for each individual session. The session ID numbers correspond to the order the recordings were made.

Consistent with the single neuron results, we found robust state-dependent value decoding during the value task. In contrast, the decoder during the perceptual task never differed from the null distribution, indicating that the OFC population did not encode any task-relevant decision variables when the animal made perceptual decisions.

DISCUSSION

We found that the OFC neurons encoded decision-related variables during value-based but not perceptual decisions. Notably, recordings were conducted in the same animal, targeting the same OFC recording sites, using the same Neuropixel probes. We reliably recorded several hundred well-isolated neurons per session during both tasks. This stark difference may reflect the difference in the cognitive processes underlying value-based choices and perceptual decisions: Value-based choices require comparing and selecting between options based on their subjective worth, whereas perceptual decisions involve categorizing and reporting sensory information. Our findings complement an earlier human lesion study (Fellows & Farah, 2007) and suggest that different neural systems mediate different types of choices, rather than a single unified decision-making network.

An alternative possibility is that our results stem from the use of different types of stimuli across the tasks: In the value task, the monkey chose between emojis associated with different rewards; in the perceptual task, the monkey categorized a motion stimulus. These different stimulus types are thought to be mediated by different visual circuits—with object-like emojis being processed in the ventral visual stream (Kravitz, Saleem, Baker, Ungerleider, & Mishkin, 2013; Lehky & Sereno, 2007; Kobatake & Tanaka, 1994; Van Essen & Gallant, 1994) and motion being processed by the middle temporal area in the dorsal stream (Lehky & Sereno, 2007; Treue & Maunsell, 1999; Salzman et al., 1990; Newsome, Wurtz, Dürsteler, & Mikami, 1985; Newsome et al., 1989). OFC receives strong, direct input from the entire ventral stream and only sparse input from the dorsal stream (Petrides, Tomaiuolo, Yeterian, & Pandya, 2012; Cavada, Compañy, Tejedor, Cruz-Rizzolo, & Reinoso-Suárez, 2000; Van Essen & Gallant, 1994). Thus, one possibility is that OFC's lack of involvement in the perceptual task could reflect that the task depends on evaluating a sensory input that OFC could only receive through other, intermediate areas.

However, several lines of evidence argue against this interpretation. First, a number of studies have reported OFC engagement in tasks where subjects assigned value on the basis of visual motion and other sensory features associated with the dorsal stream. For example, Moneta et al. (2023) found robust OFC value coding when subjects chose between different motion stimuli where different directions were associated with different rewards. Similarly, Bongioanni et al. (2021) found clear OFC value coding when subjects made decisions based on the density of dots (where greater density corresponds to greater value). Moreover, a number of studies have documented zones where the dorsal and ventral streams converge (Janssen, Verhoef, & Premereur, 2018; Kristensen, Garcea, Mahon, & Almeida, 2016; Cloutman, 2013; Van Essen & Gallant, 1994), including in the parahippocampal formation (Cloutman, 2013; Whitlock, Sutherland, Witter, Moser, & Moser, 2008; Kerr, Agster, Furtak, & Burwell, 2007; Save & Poucet, 2000), which is another major input to OFC (Petrides et al., 2012; Barbas & Blatt, 1995). These findings indicate that the OFC is capable of generating value predictions based on sensory features associated with the dorsal stream, including visual motion, as used in the present experiment. This suggests that instead of differences in upstream sensory processing and routing, our results are better explained by differences in the underlying cognitive demands of perceptual and value-based decisions.

A key finding from our study was that RTs in both tasks were influenced by the amount of expected reward. We observed faster RTs for decisions that yielded higher rewards, even in the perceptual task where OFC activity was minimal. This suggests that reward-related information is still processed during perceptual decisions but outside the OFC. Our results complement Gore et al.'s (2023) recent results showing that OFC inactivation has no effect on behavior when decisions do not require value maximization. In their study, one group of rats made value-based choices between two options, and another group was conditioned to respond to individually presented stimuli that predicted different amounts of reward. As in our study, both groups responded faster to higher-reward options. Inactivating the OFC of the rats making value-based choices degraded this reward–response correlation: RTs flattened and became decorrelated from expected reward. Strikingly, OFC inactivation had no effect on RTs in the Pavlovian group—continuing to respond faster to high-reward predicting stimuli and slower to low-reward predicting stimuli. Similarly, Oyama et al. (2024) reported no effect of chemogenetically silencing the OFC of monkeys performing a task where the outcomes were binary (simply rewarded or not), but these same monkeys' behavior was impaired in a decision task where the monkeys chose between different options that paid different amounts of reward (juice). Our findings complement this growing body of evidence suggesting that OFC is engaged when subjects must simulate and compare the outcomes associated with different options (Knudsen & Wallis, 2022; Howard & Kahnt, 2021; Zhou et al., 2021), which entails a fundamentally different set of cognitive processes from reporting sensory information, as in perceptual decision tasks.

Although we explicitly designed the perceptual task to parallel the hierarchical structure of the value task, the complexity of the tasks differed. Thus, one possibility is that our results could stem from differences in task complexity rather than differences in the cognitive demands of value-based versus perceptual decisions. However, several considerations argue against this interpretation. First, both tasks shared core structural features: context-dependent decisions where a cued state determined how to evaluate the stimulus (value mappings in the value task, which motion feature to categorize in the perceptual task), and conflict trials where identical stimuli required different responses depending on context. Second, prior work has shown that OFC robustly encodes value in much simpler tasks that lack any context-dependent, hierarchical structure (Balewski et al., 2023; McGinty & Lupkin, 2023; Padoa-Schioppa & Assad, 2006). Moreover, we recently showed that OFC encodes state-dependent values even in a two-state version of the value task (Elston & Wallis, 2025a). Together, this indicates that task complexity, per se, is not a determinant of OFC value coding.

One potential limitation of our study is that the value and perceptual tasks were recorded sequentially, with the perceptual task occurring 3 months after the value task. The serial nature of our recordings raises the possibility that the absence of decision variable encoding during the perceptual task could reflect degraded recording quality rather than functional specialization. However, several lines of evidence argue against this interpretation. First, we recorded comparable numbers of well-isolated neurons across both tasks using identical spike sorting algorithms and quality control metrics. Evidence consistent with degraded recording quality would have been substantially lower numbers of neurons and noisier spike waveforms. Second, OFC neurons showed robust task-evoked responses during the perceptual task, with activity magnitudes comparable with those observed during the value task. This widespread task engagement demonstrates that OFC remained physiologically active and responsive to task events throughout the perceptual task recordings. This suggests that rather than a degradation of recording quality, the absence of decision-related signals during the perceptual task is better explained in terms of differences in the cognitive demands underlying value-based and perceptual decisions.

We found that 82% of OFC neurons showed robust task epoch selectivity during the perceptual task, with distinct ensembles changing their activity during the cue, motion, and choice epochs, respectively. This is consistent with recent cognitive map theories of OFC function (Knudsen & Wallis, 2022; Zhou et al., 2021), which emphasize that OFC's key function is to track (and predict) the structure of the environment. These theories suggest that value is a special case of a more general associative-mapping function. However, these same neurons did not carry decision-relevant representations. We note that the decision-relevant representations we observed during the value task occurred during specific task epochs, suggesting that OFC multiplexes decision-related signals with task-structure representations. Our results indicate that OFC selectively carries decision signals, in addition to tracking task structure, when choices require evaluating and comparing values.

Despite decades of research on the neural mechanisms supporting value-based and perceptual decision-making, few studies have addressed whether their underlying neural mechanisms are shared or distinct (Bakkour et al., 2019; Polanía, Krajbich, Grueschow, & Ruff, 2014). We attribute this to the different traditions associated with each field, which have evolved largely in isolation from one another. This historical siloing has led researchers to view their particular type of decision as the canonical type of choice the brain evolved to solve. Consequently, this has prevented researchers from addressing fundamental questions about how the brain makes decisions that depend on different underlying computations. Our results demonstrate that understanding how the brain makes choices requires breaking out of the silo and moving beyond single-task paradigms to directly compare neural activity across decision types. There remains much to learn about the extent to which the neural systems supporting value-based and perceptual decisions overlap and, critically, how the brain's executive control system recruits the appropriate circuits to support the computations for the decision at hand.

Corresponding author: Thomas W. Elston, The University of Texas at Austin, Neuroscience; UC Berkeley, Neuroscience, e-mail: elston@utexas.edu.

Data Availability Statement

Data are available from the corresponding author upon request. Key code can be found at github.com/t-elston.

Author Contributions

Morgan G. Moll: Data curation; Formal analysis; Writing—Original draft. Ryan K. Read: Methodology; Writing—Review & editing. Joni D. Wallis: Conceptualization; Funding acquisition; Resources; Supervision; Writing—Review & editing. Thomas W. Elston: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Project administration; Software; Supervision; Visualization; Writing—Original draft; Writing—Review & editing.

Funding Information

National Institute of Mental Health (https://dx.doi.org/10.13039/100000025), grant number: R01-MH MH131624.

Diversity in Citation Practices

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article's gender citation balance.