Orbitofrontal value dynamics and striatal population geometry during interruption of automatic behavior

Summary

A core feature of cognitive control is the ability to interrupt behaviors that conflict with current goals. Disorders impairing this ability involve abnormal interactions between orbitofrontal cortex (OFC) and caudate nucleus (CdN). However, how these regions normally interact to interrupt automatic behaviors remains unknown. To find out, we trained monkeys on a decision task that created conflict between competing value representations, evoking an automatic-then-controlled response pattern that provided a window into how initial responses are interrupted and overridden. Simultaneous OFC-CdN Neuropixels recordings revealed that during corrective responses, OFC population dynamics vacillated between competing value states. These dynamics appeared to be transmitted to CdN via enhanced local field potential synchronization, which coincided with disruption of structured CdN population activity. Critically, this occurred without changes in striatal firing rate. This suggests that interrupting an automatic response can occur through top-down disruption of striatal population geometry without suppressing firing rates.

eTOC blurb

Elston and Wallis examined how monkeys use cognitive control to override automatic behavior. Cognitive-conflict made OFC value representations volatile and increased oscillatory synchronization with the striatum, which coincided with destabilized striatal coding and halting of the automatic response. This suggests that control operates by "jamming" rather than "suppressing" the striatum.

Introduction

A central tenet of cognitive neuroscience is the distinction between controlled and automatic processes ^1-4. Controlled processing is slow, capacity-limited, and effortful but flexible enough to be reconfigured on demand. In contrast, automatic processing is fast, parallel, and efficient but is hard to override. The controlled system is typically thought to arise from the prefrontal cortex, whereas the automatic system is thought to arise in the striatum ^5-9. These ideas have been highly influential for interpreting the mechanisms underlying neuropsychiatric disorders. For example, an inability to exert control over automatic behaviors has been argued to underlie disorders involving compulsive behavior ^10,11 such as obsessive-compulsive disorder ^12,13 and addiction ^14,15. Despite the importance of these ideas, we have limited mechanistic understanding of how the brain detects the need for control and interrupts automatic behaviors.

Theories of cognitive control have frequently highlighted a role for the dorsal anterior cingulate cortex (ACC) in detecting the need for increased cognitive control ^16-19 particularly when an automatic response competes with a task-appropriate one at the level of motor output ^20-22. However, automatic responses can conflict with controlled processing at many levels, and these other forms of conflict recruit frontal areas beyond dorsal ACC ²³. For example, emotional conflict involves ventral ACC ^24,25 and conflict during perceptual categorization involves ventrolateral prefrontal cortex ²⁶. Most of this work has been done in humans using non-invasive measures, precluding characterization of the circuit mechanisms by which conflict is detected and control is implemented. Conflict in the value domain, where automatic and controlled responses are driven by competing reward associations, has received comparatively little attention — despite its central role in compulsive and addictive disorders.

The goal of this study was to determine how frontostriatal circuits interact when an automatic, value-based response must be overridden. To do this, we leveraged a behavioral paradigm we recently used to demonstrate the existence of state-dependent values in the orbitofrontal cortex (OFC) ²⁷. The key aspect of the task is that the same stimuli take on different values in different states of the world, which induces conflict between stimulus-reward mappings and produces a clear behavioral correlate of that conflict. Critically, monkeys performing the task sometimes succeed and sometimes fail to override the automatic response. This provides a window into how automatic responses are overridden — and into what happens when they are not.

We focused on interactions between the OFC and its main target in the striatum, the ventromedial caudate nucleus (CdN). OFC sends dense monosynaptic projections to ventromedial CdN ^28-30, positioning it to directly influence striatal processing. Both regions encode the value of potential choices during decision-making ^31-34 and recent work suggests that striatal value representations are sent from OFC ^30,31. However, these studies have focused primarily on straightforward choice scenarios where the optimal response was unambiguous. Here we asked how the OFC-CdN circuit operates when an automatic, value-based response must be overridden. By simultaneously recording from hundreds of neurons in both OFC and CdN using Neuropixels probes, we tracked the real-time dynamics within this circuit to reveal the frontostriatal interactions underlying the success and failure of cognitive control in the value domain.

Results

Monkeys make automatic and then controlled behavioral responses

We trained two monkeys (subjects K and D) on a state-dependent decision task designed to evoke conflict between competing value schemas. We previously used a variant of this task to demonstrate the existence of state-dependent value codes in OFC ²⁷. On each trial, the monkeys were cued to use one of three value schemas (“states”) that varied the contingencies between a set of choice options and reward amounts (Fig. 1A). After a brief delay, they were presented with either one (forced choice, 25% of trials) or two options (free choice, 75% of trials). In states A and B, the same options had different values depending on the state, while state C, the control condition, had no such conflict. Subjects reported their choices by fixating on an option for 300 ms. Both subjects performed the task at near ceiling, selecting the higher value option on >98% of trials (Fig. S1).

Figure 1. Behavioral task.

A. Subjects initially fixated and were then shown one of six state cues (two cues per state). After a delay, subjects were presented with either one or two choice options. The optimal choice depended on the state cued earlier in the trial.

B. Histograms of choice reaction times during each of the three states. The bimodal distribution in states A and B was due to the presence of corrective saccades. Inset: mean value of the option selected by the subjects’ first saccade (± 95% confidence intervals) during trials with or without a corrective saccade.

C. Likelihood of corrective saccades in all choice conditions.

D. Comparison of logistic regression models predicting the likelihood of corrective saccades during states A and B.

E. Mean likelihood of a corrective saccade (± 95% confidence intervals) as a function of cross-state conflict and within-state value difference (see also Fig S1 and Tables S1-6).

Subjects’ eye movements revealed that some trials were more difficult than others. Specifically, subjects would often respond with an initial (erroneous) saccade to the lower value option before corrective saccade to the higher value option and selecting it. This resulted in a bimodal distribution of choice reaction times (Fig. 1B). Neither subject’s initial saccade direction significantly differed from chance (binomial test; subject K: 309/572 or 54% rightward, p = 0.06; subject D: 333/668 or 50%, p = 0.97), ruling out a simple side-bias interpretation. We excluded the small number of trials (<3%) with more than two saccades from further analysis.

Conflict between competing value schemas appeared to be the primary driver of initial errors which required correction. Corrective saccades mainly occurred during state A & B (rather than C) trials (Fig. 1C; chi-squared test; subject K: 547/570 or 96%, χ² = 354, p < 0.001; subject D: 550/668 or 82%, χ² = 195, p < 0.001). Moreover, corrective saccades in states A & B (but not C) were significantly more likely to occur when the subjects initially fixated a low value option. Given that the best option (value 4) in state A was the worst option (value 1) in state B, our results suggest that initial saccades were often biased toward options that would have been best in the counterfactual state but were the worst in the current, true state (i.e. when the state-value mappings for a choice option were highly conflicting).

To test whether corrective saccades were driven by cross-state conflict, we compared logistic regression models that did or did not include a conflict term, defined as the sum of the counterfactual state values minus the sum of the actual state values. Other candidate predictors were the within-state value difference, the maximum value on offer, and the previous trial's task state. After screening all combinations for multicollinearity (see Methods), we compared the remaining models by Akaike's Information Criterion. In both subjects the best-fitting model included conflict, either alone (subject K) or together with within-state value difference (subject D; Fig. 1D, Tables S1-6), consistent with conflict being the primary driver of corrective saccades in states A and B.

To visualize this result, we plotted p(corrective saccade) as a function of conflict and within-state value difference (Fig. 1E). Conflict was a significant, positive predictor for both subjects (subject K: β = 1.9, p < 0.001; subject D: β = 1.5, p < 0.001). Within-state value difference reached significance for subject K (β = −0.29, p < 0.001) but not subject D (β = −0.03, p = 0.66). Comparison of odds ratios in subject K indicated that conflict was a significantly larger driver of corrective saccades than within-state value difference (cross-state conflict: odds ratio = 6.8; within-state value difference: odds ratio = 0.75; Wald test on the difference in standardized coefficients: z = 14, p < 0.001). We obtained similar results when assessing choice reaction times (Fig. S1B-C). Corrective saccades could not be predicted as a function of trial or reward history (Fig. S1D). Together, these results indicate that corrective saccades were strongly associated with the conflicting reward contingencies across states that required the recruitment of cognitive control to resolve.

State-dependent and general value coding by single OFC and CdN neurons

We used Neuropixels probes to simultaneously record from OFC and CdN as subjects performed the task. Across four recordings per subject, we obtained large numbers of well isolated neurons from both regions (Fig. S2). We quantified neuronal tuning via a sliding window ANOVA with a linear contrast for value (see Methods). During the choice epoch, neurons in both brain areas encoded task state, chosen value, and their interaction (Fig. 2A, B). We show three example neurons that encode a state x value interaction (Fig. 2C) or a main effect of value (Fig. 2D). We performed additional analysis to confirm that visual responses to the state cues could not explain state encoding (Fig. S3).

Figure 2. Neurophysiological recordings and single neuron encoding.

A. Percentage of neurons significantly encoding different experimental parameters.

B. Percentage of neurons encoding the different experimental parameters during the choice epoch (shaded region of A).

C. Three example state-dependent value coding neurons in OFC.

D. Three example state-independent value coding neurons in OFC (see also Figs S2 and S3).

In addition to state and value variables, our regression model also quantified encoding of saccade direction (left vs. right). Consistent with prior reports ^34-38, relatively few OFC neurons encoded the choice direction (subject K: 128/1528, 8%; subject D: 48/1293, 4%). We also found relatively few CdN neurons encoded the choice direction (subject K: 150 / 1085, 14%; subject D: 83 / 1149, 7%). This sparse direction encoding is consistent with the topographic organization of corticostriatal circuits, in which ventromedial CdN receives projections from ventromedial prefrontal regions like OFC while progressively more dorsolateral regions receive input from prefrontal and premotor areas associated with action selection ^28,29,39. Indeed, our prior recordings from dorsolateral CdN revealed strong response encoding in nearly 50% of neurons ⁴⁰.

State-dependent value coding at the population level

Having established that individual neurons encode state-dependent values, we next asked how state-dependent values were represented at the population level. We trained Linear Discriminant Analysis (LDA) classifiers to decode state-dependent values (12 classes: 4 values × 3 states) from simultaneously recorded neuronal populations in each brain area with single-trial resolution (see Methods).

To evaluate the robustness of population-level decoding and how it scaled with the number of recorded neurons, we systematically subsampled neurons within each recording session. During the choice epoch of single saccade trials, we could decode state-dependent values well above chance in both OFC and CdN in each individual recording session (Fig. 3A). We also tested whether we could decode value from neurons which did not significantly encode any task factor during the choice epoch. These decoders performed above chance, indicating that state-dependent value information extended beyond neurons identified by our single neuron analyses. However, these decoders performed worse than those that included task-tuned neurons.

Figure 3. State-dependent value decoding and transmission.

A. Relationship between the number of simultaneously recorded neurons and state-dependent value decoding accuracy during the choice epoch of single saccade trials. The chance level was 1/12. Each trace shows results from a single brain area and single session.

B. State-dependent value decoding during single saccade trials with matched conditions. Left: The y-axis denotes the mean posterior probabilities (± 95% confidence intervals) from our 12-class state-dependent value decoder on state A and B trials where the choice options and their spatial location were held constant. Right: Confusion matrices from the decoder. The decoder performed with high accuracy across all conditions.

C. OFC transmits state-dependent values to CdN. Top: difference between models assessing OFC→CdN and CdN→OFC transmission at each lag and time point. Net directionality was computed as the difference in F-statistics between OFC→CdN and CdN→OFC transmission. Positive values indicate stronger top-down transmission from OFC to CdN. Bottom: time course of the peak lag for each subject. Shading indicates the 5-95th percentiles of null distributions.

To ensure that our results reflected decoding of state-dependent values rather than sensory factors, we examined the posterior probabilities from our 12-class state-dependent value decoder on state A and B trials when the choice options and their spatial position were held constant. This allowed us to test whether the population could discriminate the value of a choice option in the true state from the value of that same option in the counterfactual state. In both brain areas and both subjects, we observed significantly higher posterior probabilities associated with the option's value in the true state than in the counterfactual state (Fig. 3B, left). This indicates that OFC and CdN encode state-dependent value representations that are dissociable from visual responses to choice options.

Since we established that both OFC and CdN contain state-dependent value population codes, we next asked whether these representations were coordinated across the frontostriatal circuit. Does state-dependent value information emerge independently in each area, or does one region transmit this information to the other? To address this, we performed a transmission analysis ⁴¹ that tested whether millisecond-scale fluctuations in decoded value representations in one area could predict value fluctuations in the other. We focused on single saccade trials to isolate state-dependent value transmission from processes related to cognitive control. For each trial, we extracted the time series of posterior probabilities for the chosen, contextually relevant value from both OFC and CdN decoders. We then fit autoregressive models to each time series to remove autocorrelation and then used linear regression to quantify the strength of the predictive relationships between the two decoders at different time lags.

We observed strong OFC→CdN transmission during the choice epoch (Fig. 3C). This directional influence was strongest at lags of 50-75 ms, and was significantly stronger than shuffled data generated by shuffling trial labels. This indicates that fluctuations in OFC value representations predicted corresponding fluctuations in CdN approximately 50-75 ms later, consistent with prior reports that OFC is a primary source of value information across frontal and striatal circuits ^31,42.

Gaze-dependent and counterfactual value coding in OFC and CdN

Several recent studies have shown that OFC value coding is affected by gaze ^43,44. To confirm this, we compared posterior probabilities for the chosen value aligned to either the onset of the choice options or the onset of a saccade (Fig. 4A). Posteriors aligned to saccades were significantly larger than posteriors aligned to choice option onset in both OFC (paired t-test; subject K: t₃ = 24, p < 0.001; subject D: t₃ = 5.5, p < 0.01) and CdN (subject K: t₃ = 12, p < 0.001; subject D: t₃ = 5.0, p < 0.01). We next examined how corrective saccades affected value coding. For each trial, we calculated the posterior probabilities from the state-dependent value decoder for the values of the chosen and unchosen options in both the true state and the counterfactual state (Fig. 4B, S4). During single saccade trials, where the subject immediately made a saccade to the correct option, both OFC and CdN represented the value of the chosen option in the true state. During trials with a corrective saccade, OFC vacillated between the values associated with what the initially fixated (but ultimately rejected) option would have been worth in the alternative state and the ultimately-chosen option’s value in the true state. We also examined error trials, where the subject failed to correct an initial, incorrect saccade (Fig. 4C). These error trials resembled correct responses in that there was a single, stable representation – but of the counterfactual value of the fixated option. In other words, during error trials, OFC’s value representation reflected the wrong state.

Figure 4. Decoding aligned to saccades.

A. State-dependent value decoding in OFC and CdN was significantly stronger when aligned to saccades rather than choice picture onset. Each data point denotes a single recording session.

B. Mean posteriors (± 95% confidence intervals) from the state-dependent value decoder aligned to the saccade onset. Decoding was restricted to correct trials and grouped according to whether a corrective saccade occurred. Data is from subject D. Data from subject K is shown in Supplementary Figure 4.

C. Mean posteriors (± 95% confidence intervals) from the state-dependent value decoder aligned to saccade onset during error trials where subjects failed to correct an initial, incorrect response.

D. Mean posteriors (± 95% confidence intervals) from the state-dependent value decoder during the 250 ms following the onset of the first saccade (see also Figs S4 and S5).

We quantified these effects by computing the mean posterior probability in each category during the 250 ms following the onset of the first saccade, a period when gaze was stationary on the initially fixated option (Fig. 4D). We then performed a three-way ANOVA on the posterior probabilities using factors of value (high or low), state (true or counterfactual), and trial type (single saccade, double saccade, or error). There was a significant three-way interaction in both OFC and CdN in both subjects (subject K OFC: F_2,6672 = 1440, p < 0.001; subject K CdN: F_2,6672 = 240, p < 0.001; subject D OFC: F_2,5612 = 781, p < 0.001; subject D CdN: F_2,5612 = 381, p < 0.001).

We then performed post-hoc comparisons between trials with a corrective saccade and error trials. In OFC, the posterior for the high value option in the counterfactual state was significantly higher on error trials than on correct double saccade trials (subject K: t₅₈₄ = 7.4, p < 0.001; subject D: t₅₉₇ = 3.9, p < 0.001). In addition, the high value option in the true state was significantly lower on errors than during correct double saccade trials (subject K: t₅₈₄ = −2.5, p = 0.01; subject D: t₅₉₇ = −3.9, p < 0.001). A similar pattern occurred in CdN. The posterior for the high value option in the counterfactual state was significantly higher on error trials than correct double saccade trials (subject K: t₅₈₄ = 2.1, p = 0.03; subject D: t₅₉₇ = 4.4, p < 0.001). The high value option in the true state was significantly lower on errors than during correct double saccade trials for subject D (t₅₉₇ = −3.9, p < 0.001) but not subject K (t₅₈₄ = −0.40, p = 0.69). A floor effect related to the low levels of decoding of any state in CdN during corrective trials may have contributed to this inconsistency.

To directly link counterfactual value representations to the degree of cross-state conflict, we computed an option-specific conflict measure defined as the difference between the value of the chosen option in the true state and the value of the unchosen option in the counterfactual state (Fig. S5). This quantified the amount of cross-state conflict that the subject needed to overcome to choose the optimal option. We then examined how the strength of the representation of the unchosen option in the counterfactual state varied with option-specific conflict. We again focused on the 250 ms when gaze was stationary on the initially fixated option and we included both single saccade and double saccade trials. We regressed the posterior of the unchosen, counterfactual value against the option-specific conflict measure, brain area, and their interaction. The strength of the counterfactual representation significantly increased with option-specific conflict in both subjects (subject K: β = 0.05, p < 0.001; subject D: β = 0.09, p < 0.001). We also observed a significant main effect of brain area (subject K: β = −0.02, p < 0.001; subject D: β = −0.05, p < 0.001), indicating that OFC counterfactual value representations were significantly stronger than those in CdN. Finally, we observed a significant conflict-brain area interaction (subject K: β = −0.04, p < 0.001; subject D: β = −0.05, p < 0.001), indicating that OFC's counterfactual value representations scaled with the degree of conflict between competing value schemas significantly more than CdN.

In sum, the dynamics of decoded state-dependent value representations in OFC revealed a sequence whereby cross-state conflict involves a representation of the counterfactual value of the initially fixated option that switches to a representation of the value of the ultimately chosen option in the true state. This indicates that OFC actively evaluates options under competing contextual interpretations in a manner that is both constrained by but dissociable from overt spatial attention.

Single trial neural dynamics in OFC and CdN

An advantage of population level decoding is that it provides us with near real-time resolution of neural dynamics. Here we asked whether frontostriatal dynamics differed during state A and B trials in cases where an automatic response does or does not need to be interrupted.

Figure 5A shows single-trial value dynamics during four example trials - two with and two without a corrective response. The line colors denote each of the four possible values that were present in a given trial: the values of the chosen and unchosen options in the true and counterfactual states. Each pair of panels shows the simultaneous dynamics in OFC and CdN during the same trial. During single saccade trials, both brain areas exhibited a rapid, stable representation of the chosen option in the true state (Fig. 5A, left). However, during trials requiring a corrective saccade, OFC value representations became volatile and vacillated between the representing the value of the chosen option in the true state and the value of the unchosen option in the counterfactual state (Fig. 5A, right). To illustrate, the top right panel of Figure 5A shows a state A trial where subject K chose between a postbox and a blue swirl. The monkey first fixated the postbox (value 1 in state A, but value 4 in state B) before a corrective saccade to the blue swirl (value 3 in state A). OFC activity fluctuated between the postbox's counterfactual value and the blue swirl's true value. In contrast, these dynamics were absent in CdN, which showed weak value decoding during corrective saccades.

Figure 5. Value dynamics during single and double saccade trials.

A. Decoding from example trials with either a single saccade (left) or a double saccade (right). The triangular ticks indicate saccades.

B. Mean probability of a corrective saccade (± 95% confidence intervals) as a function of the number discrete decoded value states. No significant relationship was observed in CdN.

C. Mean number of discrete values (± 95% confidence intervals) decoded during the first 250 ms following the first saccade when gaze was stationary.

To establish the behavioral relevance of these dynamics, we quantified the number of discrete decoded values that occurred within the first 500 ms from picture onset on each trial (see Methods). We used logistic regression to test whether the number of decoded values from either brain area predicted whether a corrective saccade would occur, while also including cross-state conflict and within-state value difference as covariates (Fig. 5B). We observed a significant, positive relationship between the number of discrete values decoded and the likelihood of a corrective saccade in OFC (subject K: β = 0.50, p < 0.001; subject D: β = 0.21, p < 0.001) but not in CdN (subject K: β = −0.05, p = 0.2; subject D: β = −0.08, p = 0.07). In other words, volatility in OFC (but not CdN) value dynamics could predict corrective saccades over and above the experimental parameters identified in our behavioral analysis.

As noted above, OFC value coding is influenced by gaze and, since subjects were free to shift their gaze during the choice epoch, an alternative possibility is that the vacillating dynamics could reflect eye movements rather than an internal deliberative process. To distinguish between these possibilities, we used a 2-way ANOVA with factors of trial type (single or double saccade trials) and brain area (OFC or CdN) to compare value dynamics in OFC and CdN during the first 250 ms following fixation on the initially fixated option prior to making a corrective saccade. We compared these value dynamics to those that occurred on single saccade trials, using the same first 250 ms of the subject’s fixation. In both subjects, the number of decoded values during trials with a corrective saccade increased in OFC and decreased in CdN relative to trials without such saccades (Fig. 5C), resulting in a significant interaction (subject K: F_1,4775 = 62, p < 0.001; subject D: F_1,3886 = 6.8, p < 0.01). In sum, we could decode more discrete values from OFC on trials with a corrective saccade, even while gaze was stationary. In contrast, value decoding in CdN was reduced.

OFC value dynamics and disrupted CdN population geometry

Next we examined candidate mechanisms underlying the decrease in CdN value coding during corrective saccades. One possibility is widespread inhibition of CdN neurons. However, this was not the case: population-averaged firing rates in CdN were similar across trials with and without corrective saccades (Fig. 6). We next asked whether pooling across cell types obscured effects in specific subpopulations (Fig. S6). Using unsupervised clustering on the mean waveforms of each CdN neuron, we identified four putative cell types: two broad-spiking and two narrow-spiking. There was no evidence that any of these populations fired differently on trials with or without a corrective saccade. This means that rate-based inhibition cannot account for the suppressed CdN value coding during corrective saccades.

Figure 6. Firing rates during matched trials with and without corrective saccades.

Mean firing rates (± 95% confidence intervals) of all state or state-dependent value selective OFC and CdN neurons during single and double saccade trials (see also Fig S6).

An alternative possibility is that suppressed CdN value decoding could reflect disrupted population geometry. We tested this by examining population trajectories in the state space defined by the first two discriminant components of our LDA decoders. We illustrate this approach by examining trajectories during single saccade, correct trials (Fig. 7A), which exhibited trajectories that progressed directly toward the subspace corresponding to the chosen option's state-dependent value. To test the hypothesis that suppressed CdN value decoding arises from disrupted representational geometry, we compared neural trajectories during carefully matched choice scenarios. Specifically, we held constant the choice options, their values, the task state, and the subjects’ decision, and compared trials where a corrective saccade did or did not occur. We restricted the analysis to trials from states A and B where the monkey chose between the value 1 option and either the value 2 or value 3 option, as these were the conditions where most corrective saccades occurred (Fig. 1C).

Figure 7. Neural trajectories during trials with and without corrective saccades.

A. Example population trajectories for each value in states A and B during single saccade trials. The state space is defined by the first two linear discriminant components of our decoder. Colored numbers indicate the location of each state-dependent value class in the LDA space (blue = state A manifold, red = state B manifold). Thin lines show trajectories from 10 individual example trials, thick lines indicate the mean trajectory, and circles denote the final target subspace associated with the chosen value. These trajectories come from a representative session from subject K. In both areas, each trajectory travels directly to the correct value in the correct state subspace.

B. Example population trajectories from a matched condition that sometimes did and sometimes did not require a corrective saccade (state A, value 2 vs. value 1). Gray, single saccade trials; purple, double saccade trials. The arrow indicates the counterfactual value of the initially fixated option (state B, value 4). Conventions as in (A).

C. Mean posterior probabilities (± 95% confidence intervals) for decoding state-dependent values in OFC and CdN aligned to saccade onset during the matched condition trials highlighted in B. The pattern shown in (B) is seen in both subjects. The shaded region beginning at saccade onset was used for the statistical tests described in the main text (see also Fig S7).

Figure 7B shows population trajectories from subject K in state A, choosing between the value 1 and value 2 options, with and without a corrective saccade. During single saccade trials, both OFC and CdN followed smooth trajectories to the state A, value 2 subspace (similar to all other single-saccade conditions, shown in Fig 7A). During corrective saccade trials, OFC trajectories initially went to the counterfactual value of the fixated option (the value it would have in the alternative state) before shifting to the chosen value in the correct state. In contrast, CdN trajectories wandered through intermediate “null” regions unassociated with a particular value. Trajectories in principal component space showed the same pattern, meaning that these results do not depend on the LDA decoder (Fig. S7).

We quantified these effects using saccade-aligned decoded posterior probabilities (Fig. 7C). The posteriors from our LDA decoder index how close the population state was to each value subspace, with higher values indicating proximity to the region of the manifold representing a given state-dependent value. We ran a three-way ANOVA with factors of brain area (OFC vs. CdN), posterior type (chosen value in the true state vs. unchosen value in the counterfactual state), and trial condition (single saccade, first saccade of a double saccade trial, second saccade of a double saccade trial). In both subjects there was a significant three-way interaction (subject D: F_2,4228 = 43, p < 0.001; subject K: F_2,4936 = 235, p < 0.001), confirming that OFC and CdN differed in how their value representations changed across trial conditions.

Post-hoc pairwise comparisons (FDR corrected), run separately for OFC and CdN, showed that both areas had reduced value coding at the first and second saccades of corrective trials relative to single saccade trials (all p < 0.001). The two areas diverged between the first and second saccades: in OFC, encoding of the chosen option’s value in the true state significantly increased between the first and second saccades (subject D: t₈₆₄ = −3.0, p = 0.002; subject K: t₈₈₀ = −4.3, p < 0.001) while the encoding of the unchosen option’s value in the counterfactual state significantly decreased (subject D: t₈₆₄ = 3.1, p = 0.001; subject K: t₈₈₀ = 4.5, p < 0.001). In contrast, encoding of the chosen option’s value in the true state did not increase in CdN. In subject K, the encoding of the chosen option’s value in CdN was statistically indistinguishable between the first and second saccades (t₈₈₀ = −0.62, p = 0.54), and in subject D it in fact significantly decreased during the second saccade (t₈₆₄ = −3.3, p < 0.001).

Elevated OFC-CdN synchronization during corrective saccades

Our population decoding results suggest that OFC transmits value information to CdN during decision-making. However, during corrective saccades, OFC vacillated between competing value representations. These volatile dynamics coincided with disrupted CdN population geometry, as if CdN could not maintain coherent representations when its OFC input shifted too rapidly. To further investigate the circuit mechanisms underlying how OFC and CdN communicate, we examined their local field potential (LFP) coherence during trials with and without a corrective saccade.

Across both subjects, we observed elevated coherence in the high alpha, low beta band (~8 - 16 Hz, hence ‘α/β’) during decisions which required a corrective saccade as compared to single saccade trials (Fig. 8A). We observed an additional peak in the high beta band (~ 20 - 28 Hz) in subject K, but not subject D. We focused on the α/β band since that was consistent across subjects. Coherence in the α/β band peaked around the time of each saccade (Fig. 8B).

Figure 8. Frontostriatal coherence during corrective saccades.

A. OFC-CdN spectral coherence during single and double saccade trials for subjects K (top) and D (bottom). The boxed region highlights that α/β coherence is elevated during double saccade trials compared to single saccade trials (p < 0.001 for both subjects, unpaired t-tests).

B. Data as in (A) but aligned to onset of saccades on the double saccade trials.

C. Transmission analysis shows that OFC drives α/β coherence and that this directionality is stronger during trials with a corrective saccade. Conventions as in Figure 3C; shading indicates bootstrapped 95% confidence intervals.

D. Mean OFC-CdN α/β coherence (± 95% confidence intervals) as a function of the volatility of the value representation.

We next asked whether OFC or CdN drives this α/β synchronization. We performed our transmission analysis on the amplitude envelopes of α/β band LFPs to assess directional interactions between the two regions. Across both subjects, we found that OFC→CdN α/β coherence was stronger during trials with a corrective saccade as compared to those without (Fig. 8C). We also observed that α/β coherence increased as a function of the number of discrete value states decoded from OFC but not CdN (Fig. 8D). We quantified this with a 2-way ANOVA on the mean z-scored α/β coherence with factors of brain area and the number of decoded values in a trial. In both subjects, there was a significant interaction (subject K: F_{1, 8788} = 11, p = 0.001; subject D: F_1,7924 = 7.3, p = 0.006). Post-hoc linear contrasts showed that the relationship between the number of decoded states and α/β coherence was positive and significant in OFC (subject K: β = 0.061, p = 0.001; subject D: β = 0.057, p = 0.02,) but not CdN (subject K: β = 0.0008, p = 0.09; subject D: β = −0.01, p = 0.33). This suggests that as OFC shifted between value representations, α/β coherence with CdN increased, and CdN was displaced from its value coding manifold.

We also examined whether this relationship was specific to the α/β band. We fit regressions predicting OFC-CdN coherence from the number of decoded values, frequency band, and their interaction (see Methods). α/β was the only band significant in both subjects (subject K: β = 0.044, p = 0.005; subject D: β = 0.042, p = 0.01). Theta was significant in subject D (β = 0.046, p = 0.008) but not K (β = −0.004, p = 0.79), consistent with the elevated theta coherence seen only in that subject. High beta was not significant in either (subject K: β = 0.026, p = 0.10; subject D: β = 0.022, p > 0.1).

Discussion

An OFC mechanism for cognitive control

Most theories of cognitive control posit that control is recruited to resolve conflict. Typical conflict paradigms focus on conflict that is elicited by sensorimotor contingencies. For example, the Stroop task uses stimulus features to simultaneously activate competing motor responses. These tasks have highlighted a role for the ACC in detecting response conflict and signaling the need for increased cognitive control, with ACC activity increasing as the degree of conflict increases ^16,19,48,49. Our task generated a different type of conflict than the Stroop task: competing value assignments to a choice option rather than competition between motor responses. Consequently, we saw competition between distinct OFC neural ensembles responsible for encoding value in different task states. Our results suggest that conflict detection and resolution could be a common function of prefrontal cortex with different areas responsible for cognitive control elicited by different types of conflict ²³.

An alternative possibility is that intra-OFC ensemble competition could be the source of cognitive conflict that ACC has been proposed to detect and resolve. There are strong, reciprocal connections between OFC and ACC ^50,51 and we have previously shown that the temporal dynamics of action related ramping activity in ACC are driven by fluctuating OFC value signals ⁴⁰. In this model, ACC would detect value related conflict in OFC and then disrupt value coding in CdN. However, several lines of evidence argue against this. First, we observed that OFC value dynamics were correlated and led those in CdN during trials that did not require conflict resolution. This suggests that OFC and CdN can interact directly, without requiring ACC to act as an intermediary. Second, the anatomical organization of frontostriatal circuits makes ACC mediated control of ventromedial CdN unlikely: it is typically dorsal regions of ACC that are implicated in conflict detection ^16,19,48,49 and these areas project to the dorsolateral caudate and putamen ⁵² rather than the ventromedial caudate where we observed α/β coherence with OFC. Consistent with this view, OFC damage does not produce behavioral deficits in classic motor related tasks of cognitive control ^53,54, but does produce clear behavioral deficits in tasks involving value and rule driven behavioral adjustment ^55-57. Taken together, this suggests that the focus on the role of ACC in cognitive control may be due to the use of conflicting sensorimotor contingencies to elicit cognitive control and that other forms of conflict detection may rely on an expanded set of brain areas.

Our findings also inform theories of how cognitive control is implemented once the need for control has been detected. Several models propose that cognitive control is implemented by prefrontal-driven, top-down inhibition over automatic responses driven by the striatum ^5,6,58-60. This framework predicts reduced striatal firing rates when automatic responses are overridden. However, we found that CdN firing rates were similar during trials with and without a corrective response. Instead, we observed that population-level striatal activity became unstructured. These ideas have parallels with theories of motor control, in particular the notion that patterns of population-level neural activity can shift from an output-potent subspace that produces a movement, and an output-null subspace that does not ⁶¹. One interpretation of our results is that, although striatal neurons remained active, their population activity no longer occupied a subspace from which value-based action commands could be read out. Future experiments could examine whether the frontostriatal dynamics that we uncovered apply to other scenarios thought to require cognitive control, such as decision-making under uncertainty ^62,6.

OFC and CdN value coding

Our finding that OFC transmits state-dependent value signals to CdN complements prior reports that OFC guides striatal value representations during decision-making, in both rodents ³¹ and non-human primates ^30,64. Our results extend these findings by showing that OFC specifically transmits state-dependent value representations.

Consistent with prior reports ^43,44, we found that value coding in OFC was affected by gaze location. However, our results go beyond overt spatial attention: during the first (erroneous) saccade of trials requiring correction, OFC encoded the counterfactual value – what the fixated option would be worth in the alternative state – rather than its actual value in the true state. This shows that OFC was actively evaluating competing hypotheses about the state of the world. When an option had different values depending on the context, this serial evaluation process naturally produced the shifting value dynamics we observed during trials with conflicting state-value mappings. Given the strong monosynaptic projections from OFC to CdN ^24,25,27 and our finding that OFC transmits state-dependent value signals to CdN, these shifting dynamics could conceivably be relayed to CdN. This suggests that the same neural dynamics that signal conflict – unstable OFC value representations arising from conflicting value schemas – could also interrupt the initial, automatic CdN response by disrupting striatal population geometry, thereby preventing commitment to a response driven by contextually inappropriate information. Experiments utilizing causal perturbations of the OFC-CdN circuit are necessary to confirm this.

Frontostriatal coherence as a biomarker of control

Across both subjects we observed elevated OFC-CdN coherence in the high alpha / low beta band (8 – 16 Hz) during the interruption of automatic behaviors. Similar effects have been reported in top-down attentional control ^65-70. Understanding LFP dynamics has potential clinical significance, since LFP signals form the basis of closed-loop deep brain stimulation devices that are increasingly used to treat neuropsychiatric disorders ^45-47,71. For example, recent studies have demonstrated that people with obsessive-compulsive disorder can be effectively treated with devices that use LFP signals recorded in OFC to control striatal stimulation ⁴⁶. Our findings place these interventions on a more mechanistic footing and provide a computational framework that could inform the next iteration of deep brain stimulation treatments.

Disrupted OFC-CdN connectivity is a key pathophysiological feature in several neuropsychiatric conditions characterized by a loss of cognitive control over automatic behaviors, including obsessive-compulsive disorder ^12,72,73 and addiction ^10,14,15. While the specific mechanisms underlying these disorders remain to be determined, our results reveal how normal OFC-CdN interactions could support the interruption of an automatic response. More broadly, our findings warn against simple accounts of control that rely on global changes in firing rates and highlight the need for understanding population-level dynamics in the implementation of executive processes.

STAR Methods

Experimental model and 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. Two male rhesus macaques (subjects D and K, respectively) aged 9 and 5 years, both weighing 12 kg at the time of recording were used in the current study. Their fluid intake was monitored and regulated to ensure they were sufficiently motivated to work for juice reward. Subjects sat in a primate chair (Crist Instrument, Hagerstown, MD) and their eye movements were tracked with an infrared system (Eyelink 1000, SR Research, Ottawa, Ontario, CN). Stimulus presentation and behavioral conditions were controlled using the MonkeyLogic toolbox ⁶⁶. Subjects had unilateral recording chambers (Gray Matter Research, Bozeman, MT) implanted that allowed access to both OFC and CdN.

Behavioral task

Stimuli were presented on a computer monitor positioned at a viewing distance of approximately 30 cm. The subjects were trained to perform a state-dependent decision task. This required them to choose between eight reward-predictive pictures, where the reward amounts associated with the pictures depended on a cued task state (Fig. 1A). Subjects self-initiated trials by fixating on a white dot for 500 ms. After initial fixation, the subjects were shown a state cue that indicated whether the upcoming choice should be evaluated according to the value scheme associated with state A, B, or C. To unconfound neuronal activity related to the physical properties of the stimuli from the meaning they signified, we used two distinct cues for each state (so six total state cues). After 500 ms, the state cue disappeared, and the subjects had to maintain central fixation for a further 100 ms. The choice options were then presented. On 75% of the trials two options were shown (free choice trials), while on the remaining 25% of the trials there was only one option (forced choice trials). The inclusion of forced choice trials ensured that subjects regularly experienced all reward contingencies. Subjects reported their choice response by shifting their gaze to their selection and fixating for 300 ms. The values of the choice options corresponded to different amounts of apple juice reward. The juice amounts for values levels 1 through 4 were 0.15, 0.30, 0.45, and 0.60 mL, respectively. There was an inter-trial interval of 1 s. The animals completed between 700 and 1200 trials per session.

To illustrate how state modulates value and choice, consider a trial where the state cue was a tentacle-like fractal (indicating state A), and the choice options were the blue swirl and postbox emojis. In state A, selecting the blue swirl yielded three drops of juice whereas selecting the postbox yielded only one drop of juice. However, had a state B cue been shown earlier (e.g., the rainbow-colored fractal), selecting the postbox would have yielded four drops of juice and the blue swirl only two drops. State cues were pseudorandomized across trials. State C functioned as a control to assess neural responses during choice where cognitive control was not required. It had a different set of choice options to states A and B and there was only one value scheme, so the picture-value contingencies were constant across trials.

Neurophysiological recordings

Subjects were fitted with titanium head positioners and imaged in a 3T magnetic resonance imaging scanner. The resulting images were used to generate 3D reconstructions of each subject’s skull and brain areas of interest. We then implanted custom, radio-translucent recording chambers made of polyether-ether-ketone (Gray Matter Research, Bozeman, MT). Each session we acutely lowered two 45 mm primate Neuropixels probes (IMEC VZW, Leuven, Belgium) into area 13 of OFC and the ventromedial aspect of CdN, which is a major striatal projection target of OFC ^28,30. Electrode trajectories were determined by incorporating the subject’s MRI into computer-assisted drawing software (OnShape, Cambridge, MA). They were then used to design and 3D print custom recording grids that allowed us to accurately lower the probes to the appropriate brain areas (Form 4 3D printer, Form Labs, Somerville, MA). After insertion, the probes were allowed to settle for around 60 min before the start of the experiment, to mitigate neuronal drift. We configured 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 signals were 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 ⁷⁴. 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 onto a common timeline. We accomplished this by broadcasting a 5 V, 1 Hz square wave to a unique reference channel on the Neuropixels probes (channel 385) and the non-neural 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 events onto the timeline of the first Neuropixels probe. To be included for analysis, neurons had to be present for >90% of the session, have a mean firing rate >1 Hz, and have <0.1% of spikes occurring within 1 ms of another isolated neuron’s spikes (that is, <0.1% interspike interval violations).

Single neuron tuning

To quantify neuronal tuning, we used a sliding window ANOVA with a linear contrast for value. We restricted the analysis to single saccade trials to isolate state-dependent value representations from control related processes that may accompany corrective saccades. We first fit a general linear regression where, for each neuron, we modeled firing rate ($F$) during 100 ms windows stepped by 25 ms as:

$$\begin{gathered} F={\beta }_1\cdot \text{state}+{\beta }_2\cdot \text{value}+{\beta }_3\cdot \text{state}\times \text{value}+{\beta }_4\cdot \text{direction}+{\beta }_5\cdot \text{state}\_\text{cue}+ \\ {\beta }_6\cdot \text{state}\_\text{cue}\times \text{state}+{\beta }_7\cdot \text{trial}\_\text{number}+{\beta }_0 \end{gathered}$$

where state was a categorical factor (3 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), direction was coded as −1 for left choices and +1 for right choices, and state cue indicated which visual cue was used to indicate state for that trial. We used this to help disambiguate abstract stateen-coding from the visual features of the cues used to indicate each state. Trial number was a nuisance term meant to capture any 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 < 0.05) for at least 150 ms (1.5 times the window width).

Population-level decoding of value information

We trained Linear Discriminant Analysis (LDA) decoders to predict state-dependent values from OFC and CdN neural activity. For each brain area, the decoder classified 12 distinct value classes (4 possible reward values × 3 task states) using z-scored firing rates from hundreds of simultaneously recorded neurons. In each of 1000 bootstrap iterations, we randomly selected 90% of single saccade trials for training and used all remaining trials for testing. The decoder computed posterior probabilities for each value class in the test trials. We manually equated the priors for each of the 12 value classes to ensure our decoder was not affected by sampling bias. This procedure was repeated across time using 100 ms bins with a 25 ms step between bins. Across bootstraps, all trials were used in the testing set and we averaged across bootstrap iterations where the same trial appeared in the test set. Thus, we obtained trial-by-trial estimates of the relative strength of each state-dependent value at each timestep for both brain areas. For each trial, we could then recategorize decoded values based on their behavioral relevance. For example, whether they corresponded to the chosen or unchosen option, or whether they represented the value in the true state or the counterfactual state.

To evaluate how decoder performance scaled with population size and to assess the consistency of value representations across recording sessions, we systematically varied the number of neurons included in the decoder. We performed this neuron dropping analysis separately for each recording session using only neurons recorded in that session – in other words no pseudo-populations were constructed. For each population size (ranging from 10 to 500 neurons), we randomly subsampled the target number of neurons and trained an LDA decoder on this subpopulation using the procedure described above. We then repeated the process for 1000 bootstrap iterations and computed the mean decoding accuracy across iterations.

To quantify the dynamics of the value representations, we identified sustained periods of value decoding, defined as the posterior probability for a given value exceeding twice the chance level (8.3% x 2 = 16.6%) for at least two consecutive time bins (50 ms). These thresholds were selected because they yielded a false positive rate during the fixation epoch below 5% across all 12 value classes. We then quantified the number and duration of discrete value representations decoded during the choice epoch.

Transmission of state-dependent value signals

To determine whether state-dependent value signals were transmitted between OFC and CdN, we used a transmission analysis that tested whether fluctuations in decoded value from one area predicted fluctuations in the other area, after accounting for autocorrelation within each region ⁴¹. For each single saccade trial, we extracted the time series of posterior probabilities for the chosen, contextually relevant value from both OFC and CdN decoders. We then fitted autoregressive models to each trial's posterior probability time series to remove autocorrelation, yielding whitened residual time series. We used ARIMA models (autoregressive integrated moving average, order 10, 2, 2) implemented in Python using the statsmodels library. These models predicted each value in a time series as a weighted sum of prior values and prior prediction errors, after differencing the time series twice to improve stationarity.

We used a 500 ms sliding window advanced in 25 ms steps to test whether fluctuations in one area's residuals predicted fluctuations in the other at different temporal lags (0 ms to 200 ms in 25 ms increments). At each window position and lag, we aggregated residual values over trials and ensembles while maintaining the pairing of simultaneously recorded data between areas and then performed linear regression. To evaluate OFC→CdN transmission, we regressed CdN residuals at time t onto OFC residuals at time t - lag. To evaluate CdN→OFC transmission, we performed the reverse regression. We quantified transmission strength using the F-statistic associated with each regression, producing time-varying estimates of directional signal transmission throughout the trial. We computed net directionality as the difference between F-statistics for the two directions (F_OFC→CdN - F_CdN→OFC) at each time point and lag. Positive values indicated stronger OFC→CdN transmission whereas negative values indicated stronger CdN→OFC transmission.

To establish statistical significance, we performed a bootstrap procedure (1000 iterations) in which we shuffled trials to break the simultaneity of neural activity between areas while preserving the residual posterior probability time series structure within each area. This shuffling procedure estimated the degree of transmission attributable to factors such as parallel entrainment to task events, rather than genuine inter-area signal transmission. We determined significance by assessing when the F-statistic from actual data exceeded the 95th percentile of the shuffled data distribution.

Population trajectories in neural state space

To understand the nature of suppressed CdN decodability during corrective saccades, we examined trajectories in state-value space formed by the activity of state-dependent value neurons. We visualized population dynamics by projecting neural activity into the space defined by the linear discriminants from our state-dependent value decoder. For each brain area, we trained a Linear Discriminant Analysis (LDA) classifier to decode the 12 state-dependent value classes (4 values × 3 states) using mean firing rates during the choice epoch (500 ms from the choice option onset) from single saccade trials. We then projected neural activity from every time point into this LDA space, creating continuous population trajectories throughout each trial. We also computed the mean location in LDA space for each state-dependent value, which defined the value encoding manifold for each brain area.

We used the posterior probabilities from the LDA decoder to quantify observations drawn from these observations of neural state trajectories. The posterior probabilities from the LDA decoder provided a normalized measure of how close the neural population state was to each value subspace, considering all LDA dimensions, not just the ones we used for visualization. Larger posteriors for a given state-value class indicated that the trajectory was nearer that region of the value manifold.

Local field potential analysis

We computed spectral coherence between simultaneously recorded OFC and CdN LFPs using the scipy.signal.coherence method in Python. Spectral coherence quantifies the frequency dependent correlation between two signals, measuring how consistently the phase relationships between frequency components are maintained over time. We defined theta as 4-8 Hz, alpha/low beta as 8-16 Hz, and high beta as 17-30 Hz. Our parameters for the coherence computation were a sampling frequency of 1000 Hz, a 1 s reading window with 95% overlap between windows, and a Hamming window as the tapering function. We computed the OFC - CdN coherence at each timestep of each trial. For statistical comparisons between single and double saccade trials, we focused on the 500 ms period following choice option onset, comparing coherence values at each frequency using unpaired t-tests.

To assess directional interactions, we performed our transmission analysis on the analytic amplitude envelopes of LFP signals. LFP data from each region were first z-scored per channel across all trials within a session. After z-scoring, we averaged signals within each brain area to obtain a single representative LFP trace per region per trial. These averaged LFP signals were then bandpass filtered into the α/β band using a 4th-order Butterworth filter applied bidirectionally (filtfilt) to isolate a prominent oscillatory signal observed consistently across both subjects during task performance. Following filtering, signals were downsampled from 1000 Hz to 40 Hz (25 ms time bins) using the decimate method with anti-aliasing. Analytic amplitude was extracted via the Hilbert transform, and trial-averaged amplitude envelopes were subtracted from each trial to remove evoked components and isolate trial-specific amplitude fluctuations. All other aspects of the analysis were the same as described above.

To assess whether the relationship between OFC value volatility and coherence was frequency-specific, we fit a regression predicting trialwise OFC-CdN coherence from the number of decoded values, frequency band (theta, α/β, high beta), and their interaction. We obtained each band's slope by refitting the model three times, using a different band as the reference level each time, so that every slope was estimated while accounting for the other two bands.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
Original code	This paper	https://github.com/t-elston/FS_self_control
NIMH MonkeyLogic	National Institute on Mental Health	https://monkeylogic.nimh.nih.gov/
MATLAB	MathWorks	https://www.mathworks.com/products/matlab.html
Python	Python Software Foundation	https://www.python.org/

Highlight bullets.

• OFC transmits state-dependent values to striatal circuits

• Distinct neural dynamics in OFC and CdN during inhibition of automatic responses

• Disrupted CdN population geometry but not firing rates during response inhibition

• OFC→CdN LFP synchronization associated with overriding automatic responses