{"paper_id":"3ec4a129-0309-4c19-a215-f962fcb2b73e","body_text":"Neural dynamics in ventrolateral prefrontal cortex underlie learning from feedback | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return\"[object Function]\"==o.call(a)}function e(a){return\"string\"==typeof a}function f(){}function g(a){return!a||\"loaded\"==a||\"complete\"==a||\"uninitialized\"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){(\"c\"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){\"img\"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),\"object\"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height=\"0\",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),\"img\"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||\"j\",e(a)?i(\"c\"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName(\"script\")[0],o={}.toString,p=[],q=0,r=\"MozAppearance\"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&\"[object Opera]\"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?\"object\":l?\"script\":\"img\",v=l?\"script\":u,w=Array.isArray||function(a){return\"[object Array]\"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split(\"!\"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split(\"=\"),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(\".\").pop().split(\"?\").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split(\"/\").pop().split(\"?\")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&\"css\"==i.url.split(\".\").pop().split(\"?\").shift()?\"c\":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Neural dynamics in ventrolateral prefrontal cortex underlie learning from feedback View ORCID Profile Runhao Lu , Mikiko Kadohisa , Makoto Kusunoki , Daniel J. Mitchell , Alexandra Woolgar , Mark J. Buckley , John Duncan doi: https://doi.org/10.1101/2025.11.16.688684 Runhao Lu 1 MRC Cognition and Brain Sciences Unit, University of Cambridge , Cambridge CB2 7EF, UK 2 Montreal Neurological Institute, Department of Neurology and Neurosurgery, McGill University , Montreal, Quebec H3A 2B4, Canada Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Runhao Lu Mikiko Kadohisa 1 MRC Cognition and Brain Sciences Unit, University of Cambridge , Cambridge CB2 7EF, UK 3 Department of Experimental Psychology, University of Oxford , Oxford OX2 6GG, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Makoto Kusunoki 1 MRC Cognition and Brain Sciences Unit, University of Cambridge , Cambridge CB2 7EF, UK 3 Department of Experimental Psychology, University of Oxford , Oxford OX2 6GG, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Daniel J. Mitchell 1 MRC Cognition and Brain Sciences Unit, University of Cambridge , Cambridge CB2 7EF, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Alexandra Woolgar 1 MRC Cognition and Brain Sciences Unit, University of Cambridge , Cambridge CB2 7EF, UK 4 Department of Psychology, University of Cambridge , Cambridge CB2 3EB, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Mark J. Buckley 1 MRC Cognition and Brain Sciences Unit, University of Cambridge , Cambridge CB2 7EF, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site John Duncan 1 MRC Cognition and Brain Sciences Unit, University of Cambridge , Cambridge CB2 7EF, UK 3 Department of Experimental Psychology, University of Oxford , Oxford OX2 6GG, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: John.Duncan{at}mrc-cbu.cam.ac.uk Abstract Full Text Info/History Metrics Preview PDF Abstract Learning often depends on feedback, yet how positive and negative outcomes reorganize target representations to support later memory retrieval remains poorly understood. Accumulating evidence suggests that the ventrolateral prefrontal cortex (vlPFC) acts as a central hub linking learning and retrieval, raising the possibility that it plays a critical role in this process. Here we analysed spiking activity and local field potentials (LFPs) recorded from vlPFC while monkeys performed a multi-cycle object-learning task. During the initial learning cycle, correct and incorrect feedback elicited distinct neural responses in both spiking and LFPs. In particular, positive feedback produced elevated theta power and enhanced theta-gamma phase-amplitude coupling (TG-PAC), associated with sustained suppression of neural spiking. Incorrect feedback induced stronger beta power. Despite comparable levels of object information under both feedback conditions, decoders trained and tested within the same feedback state outperformed those tested across states, revealing feedback-dependent coding formats. State-space analyses further showed that object representations following positive feedback were geometrically closer to those reinstated during later retrieval, indicating that feedback reshapes neural geometry toward retrieval-compatible states. Moreover, these geometric effects were selectively expressed on electrodes showing stronger TG-PAC or beta power, suggesting that oscillatory coordination may regulate how feedback signals are transformed into stable target codes. Together, our results reveal how vlPFC serve as a critical bridge between learning and memory retrieval, with feedback-driven dynamics reorganizing population geometry through rhythmic coordination and bringing successful outcome states closer to future retrieval representations. Introduction Learning from feedback is fundamental to adaptive, goal-directed behaviour ( 1 – 6 ). Positive and negative outcomes not only modulate learning of current goals but also shape subsequent selection and memory retrieval ( 7 – 10 ). While the neural mechanisms of reinforcement have been extensively studied ( 11 – 13 ), how feedback reorganizes cortical representations of learning targets into neural codes that support later retrieval and goal-directed behaviour remains poorly understood. In the primate brain, feedback engages interconnected prefrontal circuits including the ventrolateral and dorsolateral prefrontal cortex (vlPFC and dlPFC), which play central roles in learning, memory, and cognitive control ( 14 – 18 ). Classical accounts proposed distinct functional specializations for these regions, with vlPFC linked to object processing and dlPFC to spatial control ( 19 , 20 ). However, accumulating evidence indicates that vlPFC may serve a more domain-general role ( 16 , 17 , 21 , 22 ), coordinating multiple cognitive operations through its widespread frontotemporal and frontoparietal connections and strong oscillatory synchronization ( 23 – 26 ). Recent large-scale recordings across vlPFC, dlPFC, and temporal cortex during a multi-step object-learning task showed that vlPFC encoded all task-relevant features (cue, object, and location) earlier and more strongly than other regions and was the only area maintaining target information across both learning and retrieval ( 22 ). These findings position vlPFC as a key integrative node bridging learning and memory retrieval. Despite this, the role of feedback in shaping target representations within prefrontal cortex remains unresolved. During learning, positive and negative feedback elicit widespread spiking and local field potential (LFP) responses across prefrontal cortices ( 4 , 15 – 18 , 27 , 28 ). Such responses have often been interpreted as reflecting reward valuation or outcome monitoring, but they may also transiently reconfigure cortical population geometry, creating distinct neural states that bias subsequent processing toward or away from the learning target. This idea aligns with dynamic-coding frameworks in prefrontal cortex, which propose that information is flexibly represented through time-varying rather than static activity patterns ( 29 – 33 ), implying that positive and negative feedback may establish distinct representational formats within the same cortical network. A related but unresolved question concerns the mechanisms that stabilize these reconfigured representations. Neural oscillations have long been implicated in coordinating cortical computations during learning and plasticity ( 27 , 34 – 38 ). In prefrontal cortex, theta rhythms (∼4-8 Hz) typically emerge after feedback or errors and are associated with cognitive control and updating internal representations ( 27 , 39 – 42 ), whereas beta activity (∼20-30 Hz) has often been linked to transient suppression of neural ensembles that are no longer relevant ( 24 , 38 , 43 ). Cross-frequency coupling between theta phase and gamma amplitude (theta-gamma phase-amplitude coupling, TG-PAC) may provide a mechanism for integrating slow feedback-related control signals with fast ensemble activity supporting associative learning and memory updating ( 44 – 49 ). These oscillatory and cross-frequency dynamics may thus represent candidate mechanisms through which feedback transforms transient neural responses into stable representational formats that can be reinstated during later retrieval. To investigate how feedback reshapes target representations to support later retrieval, we analysed neural activity from macaque monkeys performing a multi-cycle object-learning task encompassing both learning and retrieval phases. We focused on the vlPFC, as previous work showed that it was the only region maintaining target information across these phases ( 22 ), whereas dlPFC and temporal cortex did not. We examined how positive and negative feedback differentially modulate spiking and LFP activity, and how these feedback-induced states reorganize target representations during learning and bias their geometry during subsequent retrieval. Furthermore, we asked whether feedback-related oscillatory dynamics, particularly theta, beta, and TG-PAC, modulate the stabilization of representations that guide future behaviour. Together, these analyses allow us to test how feedback-driven dynamics in vlPFC serves as a mechanism for reformatting neural codes to support learning and goal-directed control. Results Experimental paradigm and recordings As described in detail previously ( 22 ), two monkeys performed a multi-cycle object-learning task that required discovering targets through trial-and-error feedback in the first learning cycle (Cycle 1), and subsequently retrieving the same targets based on learned cues in later cycles (Cycles 2-4) ( Figure 1A ). In each recording session, the animal completed a series of problems, each composed of four cycles of trials. On every trial, six objects were presented in a circular array, and the animal awaited a go cue before making a saccade to one object and maintaining fixation until visual feedback appeared. Positive feedback was indicated by a coloured square (green or yellow) that replaced the chosen object and was followed by reward delivery, whereas incorrect choices produced a red square with no reward. Download figure Open in new tab Figure 1. Task design and recording sites (A) Schematic of the multi-cycle object-learning task. A fixed set of eight objects was used throughout the study (inset), four each associated with green and yellow cues. Each problem consisted of four cycles of trials. In the first cycle (Cycle 1, acquisition), through trial-and-error feedback, monkeys discovered two rewarded target objects, one from each four-object set (marked here with green and yellow circles, not present in real array). White arrows indicate the monkey’s saccade to a selected object. Positive feedback was indicated by a square of the appropriate cue colour (green or yellow) replacing the chosen object and followed by reward delivery, whereas incorrect selections produced a red square with no reward. In subsequent cycles (Cycles 2-4, retrieval), a coloured cue presented at fixation instructed which of the two previously rewarded objects should be selected. Each of these cycles consisted of two correct trials, one for each cue-target pair. After cycle 4, the problem terminated and a new pair of targets was introduced. (B) Recording locations in two monkeys. Semi-chronic 96-channel microelectrode arrays targeted the lateral prefrontal cortex, encompassing ventrolateral (vlPFC) and dorsolateral (dlPFC) prefrontal regions on the cortical convexity. The present study focused on vlPFC, which previous analyses identified as the only region maintaining target object representations across both learning and retrieval. The location of each electrode is shown by a red dot. sAS, superior arcuate sulcus; iAS, inferior arcuate sulcus; PS, principal sulcus. Across the experiment, a fixed set of eight objects was divided into two colour-defined sets (four green and four yellow). Each problem used six of these objects, three from each colour set, arranged randomly within the array so that selections had to be based on object identity rather than spatial position. During the first learning cycle (Cycle 1), monkeys had no prior cue information and discovered the two rewarded targets (one from each colour set) through trial and error. Each target was rewarded once; after both targets had been found, the cycle ended. In subsequent retrieval cycles (Cycles 2-4), a coloured cue (green or yellow) presented at the start of each trial indicated which of the two previously rewarded objects should be selected. The cue did not reappear within the same cycle once its associated target had been chosen, so that each cycle consisted of two correct trials, one for each cue-target pair. After completion of the fourth cycle, the problem terminated, and a new pair of targets replaced the previous ones to begin a new learning episode. Neural activity was recorded using semi-chronic microelectrode arrays implanted in the lateral prefrontal cortex ( Figure 1B ). Electrodes within each array were independently movable between sessions, allowing recordings from different electrode sites across days. The original study ( 22 ) included recordings from the vlPFC, dlPFC, temporal cortex, and regions within the principal sulcus. Previous analyses ( 22 ) demonstrated that, among these regions, only the vlPFC maintained robust target object representations across both learning and retrieval phases. Thus, the present study focused exclusively on vlPFC to examine how feedback-related neural dynamics reshape target representations that bridge learning and subsequent retrieval. A 96-channel semi-chronic array targeted extensive regions of the inferior frontal convexity, corresponding to areas 45a, 12l, and 12r as confirmed by MRI and post-mortem histology. Across 122 sessions from animal A and 125 sessions from animal B, we recorded 1,252 neurons from vlPFC (933 from animal A; 319 from animal B). After preprocessing and exclusion of noisy or unstable channels, LFP analyses were conducted on 946 vlPFC electrode locations (729 from animal A; 217 from animal B), hereafter simply referred to as “electrodes”. Distinct spiking and LFP responses to positive and negative feedback in vlPFC We first examined how positive and negative feedback modulated neural activity during the feedback period of the initial learning cycle (Cycle 1). As shown in Figure 2 , feedback onset produced pronounced changes in both spiking and LFP signals in vlPFC. Population firing rates increased transiently after both correct and incorrect feedback, peaking around 140 ms after feedback onset and returning to around baseline levels by about 250ms. Thereafter, responses diverged: activity following incorrect feedback remained around baseline, whereas activity after correct feedback continued to decline and remained significantly lower during the late feedback window (250-400 ms; Figure 2A ). Download figure Open in new tab Figure 2. Spiking and LFP responses to positive and negative feedback in vlPFC (A) Population firing rates averaged across vlPFC neurons for correct and incorrect feedback. (B) High-frequency power (HFP, 100-250 Hz) averaged across vlPFC electrodes for correct and incorrect feedback. (C) Event-related potentials (ERPs) averaged across vlPFC electrodes for correct and incorrect feedback. (D) Time-frequency decomposition results of oscillatory power (correct minus incorrect feedback conditions) after removing the aperiodic 1/f component using FOOOF. (E-F) Time courses of theta (4-8 Hz) and beta (20-30 Hz) power for correct and incorrect feedback. (G) Aperiodic-corrected power spectra computed using the 200-450 ms post-feedback window. (H) Cross-frequency phase-amplitude coupling (PAC) between low (4-28 Hz) and high (40-250 Hz) frequency ranges, shown as difference between positive and negative feedback. (I) Theta-gamma PAC (TG-PAC; 4-8 Hz phase, 100-250 Hz amplitude) during 200-450 ms window for correct and incorrect feedback. All time 0 values correspond to feedback onset in Cycle 1 (the learning cycle). Gray shading regions indicate time windows with significant differences between correct and incorrect trials (corrected p < 0.05, cluster-based permutation test). Significance was assessed by constructing empirical null distributions from permutation samples while accounting for serial-position effects (see Methods). Error shading or error bars denote ± SEM across neurons or electrodes. *** p < 0.001. In part, high-frequency power (HFP, 100-250 Hz) paralleled the spiking response. Following feedback onset, HFP rose sharply for both correct and incorrect trials, peaking at ∼140 ms and aligning well with the population firing-rate peak ( Figure 2A-B ). After this initial rise, HFP quickly plateaued in incorrect trials, whereas after correct feedback it displayed a rhythmic modulation suggestive of coupling with slower oscillations in the theta range. LFP voltage (i.e., event-related potentials, ERP) also differentiated feedback valence ( Figure 2C ). Both correct and incorrect feedback elicited a clear negative-going deflection beginning shortly after feedback onset. This deflection was significantly larger for correct trials and extended over ∼100-380 ms. Around 400 ms, correct-feedback trials showed a pronounced polarity reversal, producing a late positive component that was absent following incorrect feedback. To characterize the oscillatory activity of the LFP more precisely, we applied time-resolved fitting oscillations & one over f (FOOOF) analysis ( 50 , 51 ) to separate periodic oscillations from the aperiodic 1/ f background ( Figure 2D ). After removing the aperiodic component, we observed two distinct oscillatory differences during the feedback period: a relative increase in theta power and a relative decrease in beta power for correct compared with incorrect feedback. Time-resolved analyses of each condition separately showed that theta power increased markedly after correct feedback but changed little after incorrect feedback, yielding a significant difference between conditions from ∼160 to 450 ms ( Figure 2E ). In contrast, beta power decreased after correct feedback but rose transiently after incorrect feedback, peaking around 290 ms before declining, with significant differences persisting throughout ∼100-450 ms ( Figure 2F ). To confirm the frequency specificity of these effects, we examined the aperiodic-corrected power spectra averaged over the 200-450 ms window of the feedback period ( Figure 2G ). The spectra revealed clear peaks confined to the theta and beta ranges, with no evidence for an alpha-band peak, indicating that feedback-related oscillations in vlPFC were dominated by these two frequencies. Subsequent analyses therefore focused on theta and beta activity. We next examined the strength of phase-amplitude coupling during the 200-450 ms feedback window across a low-frequency range of 4-28 Hz and a high-frequency range of 40-250 Hz using modulation index (MI) ( 52 ). The difference map between positive and negative feedback in Figure 2H revealed a selective increase in theta-gamma sector, with no reliable effects in other cross-frequency combinations. To quantify this effect, we computed TG-PAC using theta phases (4-8 Hz) and gamma amplitudes (100-250 Hz). We found that TG-PAC was significantly higher for correct than incorrect feedback, consistent with the oscillatory modulation of HFP on correct trials ( Figure 2B ), and confirming enhanced theta-gamma coordination associated with successful outcomes ( Figure 2I ). Together, these results show that positive and negative feedback engage distinct spiking and LFP responses in vlPFC. Correct feedback was accompanied by enhanced theta, gamma, and TG-PAC activity, signals typically associated with learning and representational updating ( 27 , 42 ), alongside a sustained suppression of spiking. Incorrect feedback elicited stronger beta activity consistent with inhibitory control ( 24 , 43 ). These complementary feedback-related patterns suggest coordinated changes in local firing and oscillatory dynamics that may set the stage for the reorganization of object representations in subsequent learning and retrieval cycles. Feedback-dependent object representations in vlPFC We next examined how feedback valence influenced representations of the target object in vlPFC ( Figure 3 ). Object information (the identity of each object within each colour set) was decoded separately from spiking activity and LFP voltage during the feedback period of Cycle 1. For both signals, target decoding accuracy was substantially above chance under both correct and incorrect feedback conditions ( Figure 3A, B , left), suggesting that object-specific information was represented before and after feedback regardless of outcome. However, despite similar overall decoding performance, representational formats differed between the two feedback states. Classifiers trained and tested within the same feedback condition (within-feedback decoding) yielded higher accuracy than classifiers trained in one condition and tested in the other (between-feedback decoding; Figure 3A, B , right), although decoding accuracies for both within- and between-feedback classifiers remained significantly above chance. This effect was consistent across spikes and LFPs and was seen from ∼160 to 400 ms after feedback onset. These results suggest that feedback modulates the underlying population representation of target objects, such that positive and negative outcomes bias the representational format in different ways. Download figure Open in new tab Figure 3. Feedback-dependent object representations in vlPFC (A) Firing-rate-based decoding of target objects during the feedback period in Cycle 1. Left, decoding accuracy for correct (blue) and incorrect (red) feedback trials, with the lower panel showing their difference. Right, decoding accuracy when classifiers were trained and tested within the same feedback condition (within-FB) or across conditions (between-FB), with the lower panel showing within-minus-between differences. (B) Same analyses as in (A) but based on LFP voltage signals. All analyses were aligned to feedback onset (time 0). Horizontal dotted line indicates theoretical chance level (0.25). Coloured horizontal bars indicate time windows in which decoding accuracy was significantly above chance (corrected p < 0.05, cluster-based permutation test). Shaded regions denote time windows with significant differences between conditions (corrected p < 0.05, cluster-based permutation test). In the difference panels, the grey dashed line shows the permutation-derived baseline (chance-level). Together, these results indicate that although target information persists in vlPFC following both positive and negative feedback, it is embedded within distinct representational states shaped by the preceding feedback. These state-dependent formats suggest that feedback reorganizes object codes in ways that constrain their cross-state generalization, potentially biasing how information is stabilized or accessed in memory. Positive feedback reshapes object representations in vlPFC toward retrieval states To test whether feedback reorganizes target representations to support later memory retrieval, we compared population activity patterns during feedback (Cycle 1) with those during the retrieval phase of later cycles (Cycles 2-4, array period; Figure 4 ). Neural activity during the retrieval period (−200 to 100 ms from array onset, Cycles 2-4) was used to define an object space by principal component analysis (PCA) applied to either spiking rates ( Figure 4A ) or LFP voltages ( Figure 4C ). Object representations from three time periods—Cycle 1 feedback after correct trials (200-450 ms from feedback onset; C1FB_corr), Cycle 1 feedback after incorrect trials (200-450 ms from feedback onset; C1FB_incorr), and Cycles 2-4 array presentations (−200 to 100 ms from array onset; C234A)—were then projected into this retrieval-based object space (i.e., C234A space). Because the eight objects were divided into two colour-defined cue sets, the analysis was performed separately for each set. Download figure Open in new tab Figure 4. Positive feedback reshapes object representations toward retrieval states (A, C) Object spaces constructed by principal component analysis (PCA) from population activity during the retrieval period of later cycles (Cycles 2-4 peri-array period, −200 to 100 ms from array onset). Projections are shown separately for Object Set 1 and Set 2 based on (A) spiking activity and (C) LFP voltage. For each set, object representations from three periods [Cycle 1 feedback after correct trials (C1FB_corr), Cycle 1 feedback after incorrect trials (C1FB_incorr), and Cycles 2-4 peri-array presentations (C234A)] were projected into this retrieval-based space. In each 3-D plot, different colours correspond to activity from the three task periods, and each colour’s four points represent the four objects within that object set. (B, D) Mean centroid distances between C234A and C1FB_corr or C1FB_incorr representations for (B) spike-based and (D) LFP-based object spaces. Error bars denote ± 95% confidence interval across bootstrapped samples. Statistical significance was assessed using bootstrapping analyses to generate empirical null distributions (see Methods). * p < 0.05, *** p < 0.001. In this object space, the geometry of representations differed according to feedback valence. Across both spiking rates ( Figure 4A ) and LFP voltages ( Figure 4C ), C234A object representations were consistently closer to those from C1FB_corr than to those from C1FB_incorr. This pattern suggests that positive feedback reshaped target codes in vlPFC so that their geometry became more aligned with the states later reinstated during memory retrieval. To quantify this relationship, we computed centroid distances between population activity patterns from each period based on bootstrapping procedures that estimated empirical distributions (see Methods). For each object set, we first averaged activity coordinates across the four objects within a period and then calculated the Euclidean distance between centroids across periods, averaging across the two sets. Both spike-based ( Figure 4B ) and LFP-based analyses ( Figure 4D ) showed significantly shorter distances between C234A and C1FB_corr than between C234A and C1FB_incorr ( M = −0.25, 95% CI[Δ] = [−0.48, −0.01] for spike-based analysis; M = −0.91, 95% CI[Δ] = [−1.29, - 0.52] for LFP-based analysis). These findings indicate that positive feedback drives representational convergence toward retrieval states, whereas negative feedback shifts object codes farther away, consistent with a feedback-dependent reformatting of target representations in vlPFC. Beta power and TG-PAC modulate feedback-driven reorganization of object space in vlPFC Given that theta and TG-PAC increased after correct feedback, whereas beta power was stronger after incorrect feedback ( Figure 2 ), we next examined whether these oscillatory features modulated the feedback-dependent reshaping of object representations observed in Figure 4 . Because TG-PAC was most enhanced following correct feedback, we first ranked all vlPFC electrodes by their TG-PAC magnitude during 200-450 ms after positive feedback. The top one-third were defined as PAC+ electrodes and the bottom one-third as PAC-electrodes. When object spaces were constructed using only PAC+ electrodes, the strong convergence of C1FB_corr toward retrieval representations (C234A) was reproduced ( M = −0.84, 95% CI[Δ] = [−1.17, −0.50]; Figure 5A ). In contrast, object spaces derived from PAC-electrodes showed no significant difference between C1FB_corr and C1FB_incorr ( M = 0.09, 95% CI[Δ] = [−0.20, 0.37]; Figure 5B ), indicating that enhanced TG-PAC is selectively associated with representational convergence toward retrieval states. Download figure Open in new tab Figure 5. Oscillatory and cross-frequency dynamics modulate feedback-dependent reorganization of object representations (A, B) Object spaces constructed from vlPFC electrodes showing the strongest (PAC+) or weakest (PAC-) theta-gamma phase-amplitude coupling (TG-PAC) during 200-450 ms after correct feedback. (C, D) Object spaces constructed from electrodes with the highest (Theta+) or lowest (Theta-) theta power during 200-450 ms after correct feedback. (E, F) Object spaces constructed from electrodes with the highest (Beta+) or lowest (Beta-) beta power during 200-450 ms after incorrect feedback. Each representation geometry plot shows PCA-based projections of object representations from Cycle 1 feedback after correct (C1FB_corr) and incorrect (C1FB_incorr) trials and from retrieval periods in later cycles (C234A). Different colours correspond to activity from different task periods. Within each colour, the four points represent the four objects belonging to a given object set. Bar plots display corresponding centroid distances between feedback- and retrieval-period representations. Error bars denote ± 95% confidence interval across bootstrapped samples. Statistical significance was assessed using bootstrapping procedures to generate empirical null distributions. * p < 0.05, *** p < 0.001. We next repeated this analysis based on theta power. Electrodes were sorted by theta power during 200-450 ms after correct feedback, with the highest one-third designated Theta+ and the lowest one-third Theta-. Because animal A exhibited a subset of electrodes with unusually strong theta activity (see Figure S1 ), these electrodes were excluded from ranking to ensure reliability. As shown in Figure 5C and 5D , object spaces constructed from Theta+ and Theta-electrodes both showed the convergence effect ( M = −1.28, 95% CI[Δ] = [−1.61, −0.96] for Theta+ electrodes; M = −0.66, 95% CI[Δ] = [−1.00, −0.30] for Theta-electrodes), suggesting that theta power itself may not directly govern this reorganization. Finally, we tested whether beta power modulated the same phenomenon. Electrodes were ranked by beta power following negative feedback, and the upper and lower one-third defined as Beta+ and Beta-electrodes, respectively. Object spaces based on Beta+ electrodes alone reproduced the convergence effect ( M = −0.37, 95% CI[Δ] = [−0.67, −0.07]), whereas those based on Beta-electrodes did not ( M = −0.16, 95% CI[Δ] = [−0.46, 0.13]) ( Figure 5E , F), indicating that stronger beta activity also influenced how feedback reshaped representational geometry. Together, these results suggest that oscillatory and cross-frequency dynamics in vlPFC contribute to how feedback shapes target representations. Electrodes showing stronger TG-PAC after correct feedback, as well as those showing stronger beta power after incorrect feedback, both exhibited clearer convergence of feedback-period representations toward retrieval states. These findings indicate that multiple oscillatory processes within vlPFC may jointly support the reorganization of object codes that underlies feedback-guided learning and subsequent memory retrieval. Discussion Learning from feedback requires transforming brief outcome signals into stable neural formats that can be reinstated during subsequent memory-guided behaviour. Our results suggest that the vlPFC implements this transformation by reorganizing population geometry in a feedback-dependent manner ( Figure 6 ). Feedback onset elicited robust spiking and LFP responses, with HFP tracking the population firing peak and ERPs differentiating outcome valence. After correct feedback, aperiodic-corrected time-frequency analyses revealed increased theta and a selective rise in TG-PAC; after incorrect feedback, beta power was stronger. Despite comparable object information being represented under both kinds of feedback, multivariate decoders trained within one feedback state consistently outperformed between-feedback generalization, indicating distinct, state-specific coding formats. Critically, population representations sampled after correct feedback were geometrically closer to those reinstated during later retrieval, whereas representations after incorrect feedback diverged from the retrieval geometry. Finally, sites with stronger TG-PAC (post-positive) or stronger beta (post-negative) showed larger modulation of these representational distances, suggesting that oscillatory coordination regulates how feedback reshapes object codes toward (or away from) retrieval-relevant states. Download figure Open in new tab Figure 6. Possible mechanisms underlying learning from feedback in vlPFC In vlPFC, positive (Pos FB) and negative (Neg FB) feedback elicited distinct neural activity including enhanced TG-PAC after positive feedback and elevated beta power after negative feedback. Although object information was represented under both feedback conditions, classifiers trained and tested within the same feedback state (within-FB) outperformed those trained across states (between-FB), indicating state-dependent representational formats. In population object space, representations following positive feedback were geometrically closer to those reinstated during later retrieval, whereas negative-feedback states diverged from the retrieval geometry. This feedback-dependent convergence and divergence of representational geometry is further modulated by post-feedback TG-PAC and beta activity. Together, these results suggest that feedback reorganizes object representations in vlPFC through oscillatory coordination, aligning correct-feedback states with future memory retrieval. These findings situate feedback as a driver of dynamic recoding in vlPFC: different feedback signals reconfigure the coding format rather than merely scaling a static code. This view resonates with dynamic coding models that prefrontal neurons continuously adjust their selectivity as behavioural demands evolve ( 29 – 33 ), but adds a geometric link to memory: positive feedback not only updates neural activity in the moment but also reorganizes the representational landscape in preparation for future retrieval. In geometric terms, positive feedback appears to reshape population trajectories so that the manifold occupied during learning approaches the subspace later engaged during memory-guided selection. Such alignment could facilitate the reinstatement of relevant target codes while minimizing interference from previously active states. This view parallels recent proposals that learning involves a gradual reorganisation of representational geometry that supports successful behaviour ( 31 , 53 – 55 ). In this framework, feedback may serve as a teaching signal that projects future retrieval demands onto the current neural space, thereby linking outcome evaluation to the formation of stable and task-relevant representations. The emergence of this alignment was closely tied to distinct oscillatory signatures, suggesting that rhythmic coordination provides a mechanism for stabilizing the reconfigured representational geometry ( 34 , 35 ). Positive feedback selectively enhanced theta power and TG-PAC, whereas negative feedback increased beta activity. These frequency-specific patterns may suggest complementary computations through which feedback regulates population dynamics. Theta oscillations have been widely linked to adaptive control and updating of internal models ( 27 , 39 – 42 ), while beta rhythms are thought to stabilize the currently active representational state and suppress competing ensembles ( 24 , 38 , 43 ). The concurrent increase in TG-PAC after positive outcomes implies a hierarchical coordination in which slow control rhythms gate fast local assemblies that encode object information ( 44 – 49 ). Such cross-frequency coupling may serve as a temporal bridge that binds transient outcome signals to longer-lasting representational formats, supporting credit assignment across successive task epochs. In this framework, oscillations do not merely mark feedback processing but implement a dynamic infrastructure that reconciles flexibility and stability, allowing prefrontal circuits to reorganize representations while preserving continuity in the learned code ( 24 , 54 ). Together, our findings reveal how feedback simultaneously shapes local neural dynamics and large-scale representational geometry in the primate prefrontal cortex. In vlPFC, feedback converts transient outcome signals into stable, retrieval-compatible population codes through coordinated oscillatory mechanisms. Positive feedback promotes convergence of object representations toward future retrieval states, supported by enhanced theta-gamma coupling, whereas negative feedback induces divergence accompanied by elevated beta activity. These rhythmic processes provide a dynamic infrastructure that balances flexibility and stability, allowing neural populations to update representations while preserving continuity in the learned code. Conceptually, this study outlines a unified framework in which feedback, oscillatory coordination, and representational geometry jointly implement a mechanism for adaptive cognition. Such principles may generalize beyond feedback learning, describing how prefrontal circuits continually restructure information to sustain flexible, goal-directed behaviour. Methods Subjects Two adult male rhesus monkeys ( Macaca mulatta ) participated in the experiments (∼14 kg each). All experimental procedures complied with the Animals (Scientific Procedures) Act 1986 of the UK and were approved by the Home Office under a Project License following review by the University of Oxford Animal Care and Ethical Review Committee. Procedures conformed to the European Community guidelines for the care and use of laboratory animals (EUVD, European Union directive 86/609/EEC). Experimental Task Monkeys performed a multi-cycle object-learning task in which they learned and selected visual targets for soft food rewards. The task required animals to discover target objects through trial-and-error feedback in the first learning cycle and subsequently retrieve the same targets based on colour cues in later cycles ( Figure 1A ). Each recording session contained a series of problems; each composed of four cycles of trials. On average, monkeys completed approximately 57 problems per session, all using a fixed set of eight objects. Four of these objects were associated with a green cue ( Figure 1A , inset), and four with a yellow cue. Six objects were used in each problem, excluding the two target objects from the preceding problem. Before each trial began, a central white fixation point (FP) appeared together with six surrounding black squares (6 × 6° visual angle, cantered 14° from fixation). To initiate the trial, the animal was required to fixate on the FP within a window of 2.6 × 2.6° (animal A) or 3 × 3° (animal B). Upon fixation, the FP turned red, initiating a short preparatory delay (0.3-0.6 s). A cue stimulus (6 × 6° square) then appeared at the fixation point. In the initial learning cycle (Cycle 1), the cue was grey, providing no information about which target to select; in subsequent retrieval cycles (Cycles 2-4), the cue was either green or yellow, indicating which of the two previously rewarded targets should be chosen. After a fixed delay of 0.5 or 1 s (fixed for each session), the circular array of black placeholders was replaced by six choice objects. Following a variable delay of 1 to 2.5 s, the FP changed to cyan, serving as the “go” signal. The cue stimulus simultaneously disappeared, and the animal was required to make a saccade to one of the objects within 1 s. After the chosen object was fixated and held for 0.35-0.45 s, it was replaced by a visual feedback signal (FB), presented as a 6 × 6° square. Positive feedback consisted of the appropriate cue colour (green or yellow) and was followed by a drop of soft food reward 0.05-0.15 s after feedback offset. Negative feedback was indicated by a red square and was not rewarded. The feedback stimulus remained visible for 0.4 s followed by an inter-trial display (see below). Trials were aborted without reward if fixation was broken before the go cue or if gaze was not maintained until feedback onset; such trials were excluded from analysis. Distinct inter-trial displays indicated transitions within and between task phases. For trials within a cycle, the inter-trial display consisted of the white FP with surrounding black squares, maintained for 0.7-0.9 s before the next trial began. To signal the end of a cycle, this display was preceded by a brief period showing only the grey FP (3.2-3.5 s). At the end of a problem, the screen blanked for 3.3-3.6 s before a new problem began, with a new pair of targets replacing the previous ones. All task events, including visual presentation, timing control, and reward delivery, were managed using the REX real-time data acquisition and control software (Laboratory of Sensorimotor Research, NIH). Stimuli were presented on a 17.5-inch LED screen positioned in front of the animal. Eye position was continuously monitored at 120 Hz using an infrared eye-tracking system (Applied Science Laboratories) and synchronized with neural and behavioural event markers. Neural Recordings Each monkey was implanted with a titanium head holder and recording chambers (form-fitting chamber system, Gray Matter Research) under general anesthesia using aseptic surgical procedures. Frontal chambers were positioned over the right lateral prefrontal cortex in both animals (monkey A: anterior-posterior (AP) = 34.9 mm, medio-lateral (ML) = 14.7 mm; monkey B: AP = 36.4 mm, ML = 19.1 mm), and a second chamber was placed over the right temporal cortex (monkey A: AP = 4.3 mm, ML = 11.3 mm; monkey B: AP = 3.0 mm, ML = 18.3 mm). Craniotomies were made under each chamber to permit electrophysiological recordings. Neural signals were collected using semi-chronic microdrive arrays (SC-96 for frontal cortex, SC-32 for temporal cortex; Gray Matter Research) with 1.5 mm inter-electrode spacing. Electrodes were independently movable between sessions, enabling recordings from largely non-overlapping neural populations across days. The frontal array targeted extensive regions of the lateral prefrontal cortex encompassing vlPFC and dlPFC areas on the cortical convexity, while the temporal array covered ventral subdivisions of temporal cortex. Neural activity was recorded at 30 kHz, amplified, and band-pass filtered (300 Hz to 10 kHz) using a multichannel recording system (Cerebus, Blackrock Microsystems) and stored for offline spike sorting (Offline Sorter, Plexon). Local field potentials (LFPs) were simultaneously acquired from the same electrodes, low-pass filtered at 250 Hz, and down-sampled to 1 kHz for analysis. Between sessions, electrodes were advanced by at least 62.5 µm to ensure sampling of new units. Recordings were conducted over a total of 247 daily sessions (122 with monkey A and 125 with monkey B). No pre-selection of neurons based on task responsiveness was performed; recordings proceeded once well-isolated single-unit activity was obtained. At the completion of experiments, animals were deeply anesthetized with barbiturate and perfused transcardially with heparinized saline followed by 10% formalin. Brains were removed for histological verification of recording sites, which confirmed electrode tracks within the intended regions. Univariate Analyses on Spiking Activity and LFPs To characterize feedback-related neural responses in vlPFC, we analysed both spiking activity and LFPs aligned to feedback onset during the learning cycle (Cycle 1). Population firing rate Spiking activity was smoothed with a Gaussian kernel of 50 ms full-width-at-half-maximum (FWHM) and converted to firing-rate time series aligned to feedback onset. For each neuron, firing rates were z-scored across all trials and time points and averaged separately for correct and incorrect feedback. Population activity was then obtained by averaging normalized firing rates across neurons, producing mean time courses that captured the temporal profile of feedback-locked population responses. High-frequency power (HFP) To estimate high-frequency activity associated with local spiking, LFPs were filtered in the 100-250 Hz range and transformed using the analytic amplitude of the Hilbert signal. Instantaneous power (squared amplitude) was then smoothed with a Gaussian kernel (FWHM = 50 ms) and then downsampled to 100 Hz for analysis. All signals were aligned to feedback onset and z-scored within each electrode across all trials and time points. As in other univariate analyses, for each electrode, HFP time courses were averaged across trials within each feedback condition (correct vs. incorrect) and then averaged across electrodes. LFP voltage analyses Time-domain LFP activity (i.e., ERP) were low-pass filtered below 20 Hz to extract slow potential fluctuations, down-sampled to 100 Hz, and aligned to feedback onset. Data were z-scored within each session and averaged across electrodes to obtain regional ERPs for correct and incorrect feedback. FOOOF and oscillatory power analyses To obtain aperiodic-adjusted oscillatory power time series, we combined time-frequency decomposition with spectral parameterization using the FOOOF algorithm ( 50 ). Broadband LFPs were first band-filtered to 2-30 Hz and transformed using Morlet wavelets with frequency-dependent cycles ( n_cycles = frequency/2). Prior to decomposition, phase-locked ERP activity for each object condition was subtracted from each trial to remove non-oscillatory slow components. For each electrode, trial, and time point, the resulting power spectra were parameterized using FOOOF in a fixed aperiodic mode (frequency range 2-30 Hz, peak width limits [2, 8] Hz, peak threshold 2.0, and maximum of 4 peaks). The aperiodic component (1/f background) was subtracted from the total spectrum to yield periodic, oscillatory power estimates. This procedure produced time-resolved, aperiodic-adjusted oscillatory power matrices for each recording site. Power values were z-scored within each electrode across trials and time points and down-sampled to 100 Hz for analysis. Oscillatory power in the theta (4-8 Hz) and beta (20-30 Hz) bands was then extracted from the FOOOF-corrected spectra and compared between correct and incorrect feedback conditions. TG-PAC analyses To quantify theta-gamma phase-amplitude coupling (TG-PAC) following feedback, we filtered LFPs into two frequency ranges and then computed the modulation index (MI) ( 52 ). Specifically, we defined a low-frequency range of 2-30 Hz sampled every 2 Hz and a high-frequency range of 40-250 Hz sampled every 8 Hz. For each low/high pair, we band-pass filtered the signal, extracted phase (low band) and amplitude envelope (high band) with the Hilbert transform, and estimated the MI using the Tort method ( 52 ), as implemented in the toolbox pactools ( 56 ). We aligned data to feedback onset, subtracted the condition-specific ERP to suppress slow potentials, and restricted PAC estimation to the 200-450 ms post-feedback window. To normalize MI and control for spurious coupling, we built a surrogate distribution by circularly time-shifting each single-trial time series (200 iterations), which preserves spectral content while disrupting cross-frequency dependencies. We then converted observed MI to z-scores at every frequency pair using: These surrogate-corrected normalised MI maps were averaged across electrodes and compared between correct and incorrect feedback conditions. Multivariate Decoding Analyses To decode object-specific neural representations, we used data from all recorded neurons (for spike-based decoding) or electrodes (for LFP-based decoding) across sessions and applied time-resolved multivariate pattern analysis. Data were aligned to feedback onset, down-sampled to 100 Hz, and normalized within each neuron/electrode by z-scoring across all trials and time points. We implemented a bootstrapped pseudotrial approach to improve signal-to-noise while maintaining trial-level variability. For each object condition, all trials were randomly split into two independent halves for training and testing. Within each half, we performed 30 bootstrap resampling iterations; in each iteration, trials were resampled with replacement and averaged to generate one pseudotrial. Consequently, each half contained 30 pseudotrials per object condition, yielding 120 pseudotrials in total for each object set (4 objects × 30 pseudotrials). The training set therefore consisted of 120 pseudotrials from one half, and the test set of 120 pseudotrials from the other. Decoding was conducted separately for the two object sets (green and yellow), each comprising four objects (chance level = 0.25). A linear support vector machine (SVM) classifier was trained to discriminate among the four objects within each set. Input features consisted of concatenated firing rates (or LFP amplitudes) across all neurons/electrodes at each time point. To increase the signal-to-noise ratio for each decoding sample, neighbouring timepoints (±2 samples) were concatenated as additional features. Prior to classification, features were whitened and reduced using PCA (variance retained = 95%). The classifier was trained and tested on independent pseudotrial sets (training set vs. testing set) for within-feedback decoding (e.g., correct → correct, incorrect → incorrect) and between-feedback generalization (e.g., correct → incorrect, incorrect → correct). Decoding accuracy was computed at each time point, producing time-resolved accuracy curves that were averaged across repetitions. The entire procedure was repeated 100 times to obtain the final population-level decoding curves. Representational Geometry Analyses To examine the structure of object representations across learning and retrieval, we constructed a population activity space using PCA based on neural population activity from the peri-array period of Cycles 2-4 (−200 to 100 ms from array onset; C234A). We then examined projections into this space from three task periods: Cycle 1 feedback after correct trials (200-450 ms from feedback onset; C1FB_corr), Cycle 1 feedback after incorrect trials (200-450 ms from feedback onset; C1FB_incorr), and C234A itself. Analyses were performed separately for the two object sets and for both spiking activity and LFP voltage. For each neuron or electrode, activity values were averaged within the relevant time windows, down-sampled to 100 Hz, and normalized by z-scoring across all trials and time points within a session. Correct and incorrect trials were further balanced by serial position (see Statistical analyses for details) and object identity to ensure matched sampling across conditions. To construct the population vectors, neural responses for each of the eight objects were averaged across trials and randomly split into two independent halves (Split A/B) to allow cross-validation. For each neuron, Split A and B means formed two independent representations per object. We concatenated these across all neurons (or electrodes) to form population activity matrices of dimension neurons × objects × splits . We then applied PCA separately for each object set (four objects per set) to reduce dimensionality to the first three components, capturing the majority of population variance. We trained the PCA model on neural population activity from the retrieval period (C234A) using one independent data split (Split A), and projected activity from all periods (C1FB correct, C1FB incorrect, and C234A) from the held-out split (Split B) into this retrieval-based space. This cross-validated approach ensured that the PCA projection and subsequent geometric analyses were performed on independent data. The resulting 3-D object constellations for each condition were visualized and compared within this common representational space. To quantify representational relationships, we generated bootstrapped samples (1000 iterations) by resampling neurons (or electrodes) with replacement within one split (Split B) to obtain variability estimates. Each bootstrap sample was projected into a fixed PCA space (trained on C234A), and distances were computed between the feedback-related object constellations (C1FB correct or incorrect) and the retrieval constellation (C234A). Centroid distances were computed as the Euclidean distance between the mean positions of the four-object constellations across periods. The resulting bootstrap distributions were used to compute 95% confidence intervals and to statistically compare the correct- and incorrect-feedback conditions using paired t -tests. Finally, to test how oscillatory activity modulated these geometric effects, we repeated the same analysis using subsets of electrodes grouped by their theta, beta, or TG-PAC strength. Electrode selection was based on the previously described FOOOF-based time-frequency and TG-PAC analyses. For theta power and TG-PAC, we averaged activity during the 200-450 ms post-feedback window after correct feedback, ranked all electrodes by their average strength, and defined the upper and lower 1/3 electrodes as Theta+/Theta- and PAC+/PAC-groups, respectively. For the Theta+/Theta-grouping, because animal A showed a secondary cluster of electrodes with unusually high theta power during correct feedback (see Figure S1 ), we excluded electrodes exceeding 0.45 power units (corresponding to the local minimum between the two peaks of the distribution) prior to defining Theta+/Theta-groups to ensure consistent ranking across animals. For beta power, electrodes were ranked according to post-feedback activity after incorrect trials, and the upper and lower 1/3 electrodes were taken as Beta+/Beta-groups. We then performed geometry analyses independently within each of these electrode subsets to assess how oscillatory dynamics influenced the stability of object-space organization. Statistical Analyses We used nonparametric, resampling-based statistics throughout. For both univariate and multivariate decoding analyses, we used permutation methods to construct empirical null distributions while explicitly controlling the potential confound from serial-position effects. In each learning cycle, monkeys viewed three objects per colour set, resulting in a structured sequence of possible outcomes: the first object (serial position 1) had a 33.3% chance of being correct, the second (serial position 2) 50%, and the third (serial position 3) 100%, as each object could only be selected once and revisits were excluded from analysis. Consequently, animals were likely to anticipate negative feedback early and positive feedback late within a cycle. To remove this confound, permutation of feedback labels was performed within each serial-position bin, preserving the empirical distribution of serial positions while disrupting the mapping between feedback type and neural response. For each comparison (e.g., correct vs. incorrect feedback), we generated 1000 permutations. At each time point, we compared the observed difference between conditions against the null distribution obtained from the permuted data to compute pointwise p values. We used cluster-based permutation tests ( 57 ) to control for multiple comparisons across time, and clusters with corrected p < 0.05 were considered significant. All tests were two-tailed unless otherwise specified. For the within-between decoding comparison, we used a one-tailed test based on the a priori hypothesis that within-feedback decoding accuracy would exceed cross-feedback decoding. For the representational geometry analyses, significance of distance metrics was determined from bootstrap distributions (1000 iterations). Confidence intervals were computed directly from the bootstrap samples; effects were considered significant if the 95%, 99%, or 99.9% confidence intervals did not include zero (corresponding approximately to p < 0.05, p < 0.01, and p < 0.001, respectively). Author contributions M. Kadohisa, M. Kusunoki, M.J.B., and J.D. designed the research. R.L. and J.D. conceived the study. M. Kadohisa and M. Kusunoki collected data. R.L., D.J.M, and J.D. analysed data. R.L. wrote the first draft of the paper. All authors contributed to the final version of the paper. J.D. supervised the work. Declaration of interests The authors declare no competing interests. Supplementary Download figure Open in new tab Figure S1. Distributions of theta, beta, and TG-PAC values across electrodes and animals. Histograms show the distribution of (A) theta power, (B) beta power, and (C) theta-gamma phase-amplitude coupling (TG-PAC) across all 946 vlPFC electrodes (729 from animal A, 217 from animal B). Each panel plots correct trials (left), incorrect trials (middle), and the difference between them (right), based on mean activity within the 200-450 ms window following feedback onset in Cycle 1. Black lines indicate kernel-density estimates computed across both animals. Only theta power during correct feedback in animal A showed a clear bimodal distribution, with a small secondary cluster of electrodes exhibiting unusually high power. Beta power and TG-PAC showed unimodal, approximately normal distributions across both animals. Acknowledgements This project was supported by UKRI MRC intramural funding MC_UU_00030/7 to D.J.M and J.D, and MC_UU_00030/15 to A.W. R.L. was supported by a Gates Cambridge Scholarship (OPP1144) and a postdoctoral fellowship from the Canadian Institutes for Health Research (200883). For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising from this submission. Funder Information Declared UKRI MRC intramural funding , MC_UU_00030/7 , MC_UU_00030/15 Gates Cambridge Trust , OPP1144 Canadian Institutes of Health Research , 200883 References 1. ↵ C. D. Luft , Learning from feedback: the neural mechanisms of feedback processing facilitating better performance . Behav Brain Res 261 , 356 – 368 ( 2014 ). OpenUrl CrossRef PubMed 2. Z. Fu , A. Sajad , S. P. Errington , J. D. Schall , U. Rutishauser , Neurophysiological mechanisms of error monitoring in human and non-human primates . Nat Rev Neurosci 24 , 153 – 172 ( 2023 ). OpenUrl CrossRef PubMed 3. D. Badre , Cognitive Control . Annu Rev Psychol 76 , 167 – 195 ( 2025 ). OpenUrl PubMed 4. ↵ M. X. Cohen , K. Wilmes , I. Vijver , Cortical electrophysiological network dynamics of feedback learning . Trends Cogn Sci 15 , 558 – 566 ( 2011 ). OpenUrl CrossRef PubMed Web of Science 5. M. Ullsperger , C. Danielmeier , G. Jocham , Neurophysiology of performance monitoring and adaptive behavior . Physiol Rev 94 , 35 – 79 ( 2014 ). OpenUrl CrossRef PubMed 6. ↵ P. Domenech , S. Rheims , E. Koechlin , Neural mechanisms resolving exploitation-exploration dilemmas in the medial prefrontal cortex . Science 369 , ( 2020 ). 7. ↵ C. B. Holroyd , M. G. H. Coles , The neural basis of human error processing: reinforcement learning, dopamine, and the error-related negativity . Psychol Rev 109 , 679 – 709 ( 2002 ). OpenUrl CrossRef PubMed Web of Science 8. S. Nieuwenhuis , C. B. Holroyd , N. Mol , M. G. Coles , Reinforcement-related brain potentials from medial frontal cortex: origins and functional significance . Neurosci Biobehav Rev 28 , 441 – 448 ( 2004 ). OpenUrl CrossRef PubMed Web of Science 9. G. Hajcak , What We’ve Learned From Mistakes . Current Directions in Psychological Science 21 , 101 – 106 ( 2012 ). OpenUrl CrossRef 10. ↵ M. J. Frank , B. S. Woroch , T. Curran , Error-related negativity predicts reinforcement learning and conflict biases . Neuron 47 , 495 – 501 ( 2005 ). OpenUrl CrossRef PubMed Web of Science 11. ↵ W. Schultz , P. Dayan , P. R. Montague , A neural substrate of prediction and reward . Science 275 , 1593 – 1599 ( 1997 ). OpenUrl Abstract / FREE Full Text 12. M. F. Rushworth , M. P. Noonan , E. D. Boorman , M. E. Walton , T. E. Behrens , Frontal cortex and reward-guided learning and decision-making . Neuron 70 , 1054 – 1069 ( 2011 ). OpenUrl CrossRef PubMed Web of Science 13. ↵ P. R. Montague , S. E. Hyman , J. D. Cohen , Computational roles for dopamine in behavioural control . Nature 431 , 760 – 767 ( 2004 ). OpenUrl CrossRef PubMed Web of Science 14. ↵ H. Tomita , M. Ohbayashi , K. Nakahara , I. Hasegawa , Y. Miyashita , Top-down signal from prefrontal cortex in executive control of memory retrieval . Nature 401 , 699 – 703 ( 1999 ). OpenUrl CrossRef PubMed Web of Science 15. ↵ Z. Fu et al. , Single-Neuron Correlates of Error Monitoring and Post-Error Adjustments in Human Medial Frontal Cortex . Neuron 101 , 165 – 177 e165 ( 2019 ). OpenUrl CrossRef PubMed 16. ↵ S. W. Kennerley , J. D. Wallis , Reward-dependent modulation of working memory in lateral prefrontal cortex . J Neurosci 29 , 3259 – 3270 ( 2009 ). OpenUrl Abstract / FREE Full Text 17. ↵ S. E. Cavanagh , J. P. Towers , J. D. Wallis , L. T. Hunt , S. W. Kennerley , Reconciling persistent and dynamic hypotheses of working memory coding in prefrontal cortex . Nat Commun 9 , 3498 ( 2018 ). OpenUrl CrossRef PubMed 18. ↵ J. Vezoli , E. Procyk , Frontal feedback-related potentials in nonhuman primates: modulation during learning and under haloperidol . J Neurosci 29 , 15675 – 15683 ( 2009 ). OpenUrl Abstract / FREE Full Text 19. ↵ S. P. Ó. Scalaidhe , F. A. Wilson , P. S. Goldman-Rakic , Areal segregation of face-processing neurons in prefrontal cortex . Science 278 , 1135 – 1138 ( 1997 ). OpenUrl Abstract / FREE Full Text 20. ↵ R. E. Passingham , S. P. Wise , The Neurobiology of the Prefrontal Cortex: Anatomy, Evolution, and the Origin of Insight . ( Oxford University Press , 2012 ). 21. ↵ F. X. Neubert , R. B. Mars , A. G. Thomas , J. Sallet , M. F. Rushworth , Comparison of human ventral frontal cortex areas for cognitive control and language with areas in monkey frontal cortex . Neuron 81 , 700 – 713 ( 2014 ). OpenUrl CrossRef PubMed Web of Science 22. ↵ M. Kadohisa et al. , Frontal and temporal coding dynamics in successive steps of complex behavior . Neuron , ( 2022 ). 23. ↵ L. R. Trambaiolli et al. , Anatomical and functional connectivity support the existence of a salience network node within the caudal ventrolateral prefrontal cortex . Elife 11 , ( 2022 ). 24. ↵ M. Lundqvist et al. , Gamma and Beta Bursts Underlie Working Memory . Neuron 90 , 152 – 164 ( 2016 ). OpenUrl CrossRef PubMed 25. M. B. Broschard , J. E. Roy , S. L. Brincat , M. K. Mahnke , E. K. Miller , Evidence for an active handoff between hemispheres during target tracking . J Neurosci , ( 2025 ). 26. ↵ J. Riddle , J. M. Scimeca , D. Cellier , S. Dhanani , M. D’Esposito , Causal Evidence for a Role of Theta and Alpha Oscillations in the Control of Working Memory . Curr Biol 30 , 1748 – 1754 e1744 ( 2020 ). OpenUrl CrossRef PubMed 27. ↵ J. F. Cavanagh , M. J. Frank , Frontal theta as a mechanism for cognitive control . Trends Cogn Sci 18 , 414 – 421 ( 2014 ). OpenUrl CrossRef PubMed Web of Science 28. ↵ F. Kirsch , H. Kirschner , A. G. Fischer , T. A. Klein , M. Ullsperger , Disentangling performance-monitoring signals encoded in feedback-related EEG dynamics . Neuroimage 257 , 119322 ( 2022 ). OpenUrl PubMed 29. ↵ M. G. Stokes et al. , Dynamic coding for cognitive control in prefrontal cortex . Neuron 78 , 364 – 375 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 30. T. J. Buschman , E. K. Miller , Working Memory Is Complex and Dynamic, Like Your Thoughts . J Cogn Neurosci , 1 – 7 ( 2022 ). 31. ↵ T. J. Buschman , Balancing Flexibility and Interference in Working Memory . Annu Rev Vis Sci , ( 2021 ). 32. Y. Zheng , R. Lu , A. Woolgar , Radical flexibility of neural representation in frontoparietal cortex and the challenge of linking it to behaviour . Current Opinion in Behavioral Sciences 57 , 101392 ( 2024 ). OpenUrl 33. ↵ J. Duncan , An adaptive coding model of neural function in prefrontal cortex . Nature reviews neuroscience 2 , 820 – 829 ( 2001 ). OpenUrl CrossRef PubMed Web of Science 34. ↵ P. Fries , A mechanism for cognitive dynamics: neuronal communication through neuronal coherence . Trends Cogn Sci 9 , 474 – 480 ( 2005 ). OpenUrl CrossRef PubMed Web of Science 35. ↵ G. Buzsáki , A. Draguhn , Neuronal oscillations in cortical networks . Science 304 , 1926 – 1929 ( 2004 ). OpenUrl Abstract / FREE Full Text 36. W. Fang , X. Jiang , J. Chen , C. Zhang , L. Wang , Oscillatory control over representational geometry of sequence working memory in macaque frontal cortex . Curr Biol 35 , 1495 – 1507 e1495 ( 2025 ). OpenUrl PubMed 37. N. A. Herweg , E. A. Solomon , M. J. Kahana , Theta Oscillations in Human Memory . Trends Cogn Sci 24 , 208 – 227 ( 2020 ). OpenUrl CrossRef PubMed 38. ↵ R. Schmidt et al. , Beta Oscillations in Working Memory, Executive Control of Movement and Thought, and Sensorimotor Function . J Neurosci 39 , 8231 – 8238 ( 2019 ). OpenUrl Abstract / FREE Full Text 39. ↵ J. F. Cavanagh , M. J. Frank , T. J. Klein , J. J. Allen , Frontal theta links prediction errors to behavioral adaptation in reinforcement learning . Neuroimage 49 , 3198 – 3209 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 40. L. T. Hsieh , C. Ranganath , Frontal midline theta oscillations during working memory maintenance and episodic encoding and retrieval . Neuroimage 85 Pt 2 , 721 – 729 ( 2014 ). OpenUrl CrossRef PubMed 41. E. L. Johnson et al. , A rapid theta network mechanism for flexible information encoding . Nat Commun 14 , 2872 ( 2023 ). OpenUrl CrossRef PubMed 42. ↵ C. Bradley , E. McCann , A. S. Nydam , P. Dux , J. B. Mattingley , Causal evidence for increased theta and gamma phase consistency in a parieto-frontal network during the maintenance of visual attention . ( 2024 ). 43. ↵ E. K. Miller , M. Lundqvist , A. M. Bastos , Working Memory 2.0 . Neuron 100 , 463 – 475 ( 2018 ). OpenUrl CrossRef PubMed 44. ↵ R. T. Canolty , R. T. Knight , The functional role of cross-frequency coupling . Trends Cogn Sci 14 , 506 – 515 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 45. J. E. Lisman , O. Jensen , The theta-gamma neural code . Neuron 77 , 1002 – 1016 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 46. J. Daume et al. , Control of working memory by phase-amplitude coupling of human hippocampal neurons . Nature 629 , 393 – 401 ( 2024 ). OpenUrl CrossRef PubMed 47. M. Tamura , T. J. Spellman , A. M. Rosen , J. A. Gogos , J. A. Gordon , Hippocampal-prefrontal theta-gamma coupling during performance of a spatial working memory task . Nat Commun 8 , 2182 ( 2017 ). OpenUrl CrossRef PubMed 48. A. B. Tort , R. Komorowski , J. Manns , N. Kopell , H. Eichenbaum , Theta–gamma coupling increases during the learning of item–context associations . Proc Natl Acad Sci U S A 106 , 20942 – 20947 ( 2009 ). OpenUrl Abstract / FREE Full Text 49. ↵ M. Ursino , G. Pirazzini , Theta–gamma coupling as a ubiquitous brain mechanism: implications for memory, attention, dreaming, imagination, and consciousness . Current Opinion in Behavioral Sciences 59 , 101433 ( 2024 ). OpenUrl 50. ↵ T. Donoghue et al. , Parameterizing neural power spectra into periodic and aperiodic components . Nat Neurosci 23 , 1655 – 1665 ( 2020 ). OpenUrl CrossRef PubMed 51. ↵ A. Bender , C. Zhao , E. Vogel , E. Awh , B. Voytek , Differential representations of spatial location by aperiodic and alpha oscillatory activity in working memory . Proc Natl Acad Sci U S A 122 , e2506418122 ( 2025 ). OpenUrl PubMed 52. ↵ A. B. Tort , R. Komorowski , H. Eichenbaum , N. Kopell , Measuring phase-amplitude coupling between neuronal oscillations of different frequencies . J Neurophysiol 104 , 1195 – 1210 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 53. ↵ N. Kriegeskorte , J. Diedrichsen , Peeling the Onion of Brain Representations . Annu Rev Neurosci 42 , 407 – 432 ( 2019 ). OpenUrl CrossRef PubMed 54. ↵ A. Libby , T. J. Buschman , Rotational dynamics reduce interference between sensory and memory representations . Nat Neurosci 24 , 715 – 726 ( 2021 ). OpenUrl CrossRef PubMed 55. ↵ S. Bernardi et al. , The Geometry of Abstraction in the Hippocampus and Prefrontal Cortex . Cell 183 , 954 – 967 e921 ( 2020 ). OpenUrl CrossRef PubMed 56. ↵ T. Dupre la Tour et al. , Non-linear auto-regressive models for cross-frequency coupling in neural time series . PLoS Comput Biol 13 , e1005893 ( 2017 ). OpenUrl CrossRef PubMed 57. ↵ E. Maris , R. Oostenveld , Nonparametric statistical testing of EEG- and MEG-data . J Neurosci Methods 164 , 177 – 190 ( 2007 ). OpenUrl CrossRef PubMed Web of Science View the discussion thread. Back to top Previous Next Posted November 17, 2025. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Neural dynamics in ventrolateral prefrontal cortex underlie learning from feedback Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Neural dynamics in ventrolateral prefrontal cortex underlie learning from feedback Runhao Lu , Mikiko Kadohisa , Makoto Kusunoki , Daniel J. Mitchell , Alexandra Woolgar , Mark J. Buckley , John Duncan bioRxiv 2025.11.16.688684; doi: https://doi.org/10.1101/2025.11.16.688684 Share This Article: Copy Citation Tools Neural dynamics in ventrolateral prefrontal cortex underlie learning from feedback Runhao Lu , Mikiko Kadohisa , Makoto Kusunoki , Daniel J. Mitchell , Alexandra Woolgar , Mark J. Buckley , John Duncan bioRxiv 2025.11.16.688684; doi: https://doi.org/10.1101/2025.11.16.688684 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7635) Biochemistry (17697) Bioengineering (13895) Bioinformatics (41951) Biophysics (21456) Cancer Biology (18594) Cell Biology (25520) Clinical Trials (138) Developmental Biology (13381) Ecology (19903) Epidemiology (2067) Evolutionary Biology (24323) Genetics (15612) Genomics (22510) Immunology (17737) Microbiology (40401) Molecular Biology (17183) Neuroscience (88622) Paleontology (667) Pathology (2833) Pharmacology and Toxicology (4825) Physiology (7644) Plant Biology (15158) Scientific Communication and Education (2046) Synthetic Biology (4296) Systems Biology (9825) Zoology (2271)","source_license":"CC-BY-4.0","license_restricted":false}