Tracking dynamic adjustments to decision making and performance monitoring processes in conflict tasks

How we exert control over our decision-making has been investigated using conflict tasks, which involve stimuli containing elements that are either congruent or incongruent. In these tasks, participants adapt their decision-making strategies following exposure to incongruent stimuli. According to conflict monitoring accounts, conflicting stimulus features are detected in medial frontal cortex, and the extent of experienced conflict scales with response time (RT) and frontal theta-band activity in the electroencephalogram (EEG). However, the consequent adjustments to decision processes following response conflict are not well-specified. To characterise these adjustments and their neural implementation we recorded EEG during a modified Flanker task. We traced the time-courses of performance monitoring processes (frontal theta) and multiple processes related to perceptual decision-making. In each trial participants judged which of two overlaid gratings forming a plaid stimulus (termed the S1 target) was of higher contrast. The stimulus was divided into two sections, which each contained higher contrast gratings in either congruent or incongruent directions. Shortly after responding to the S1 target, an additional S2 target was presented, which was always congruent. Our EEG results suggest enhanced sensory evidence representations in visual cortex and reduced evidence accumulation rates for S2 targets following incongruent S1 stimuli. Results of a follow-up behavioural experiment indicated that the accumulation of sensory evidence from the incongruent (i.e. distracting) stimulus element was adjusted following response conflict. Frontal theta amplitudes positively correlated with RT following S1 targets (in line with conflict monitoring accounts). Following S2 targets there was no such correlation, and theta amplitude profiles instead resembled decision evidence accumulation trajectories. Our findings provide novel insights into how cognitive control is implemented following exposure to conflicting information, which is critical for extending conflict monitoring accounts.


Introduction 70 71
In everyday life we are constantly adapting our decision--making strategies to fit the 72 demands of our environment. The cognitive systems that allow us to do this, and their 73 implementation in the brain, have been the focus of extensive investigation. One core 74 area of this research concerns how we deal with conflicting information, specifically in 75 situations where different features of a stimulus are each associated with different motor 76 actions. For example, in the Eriksen Flanker task (Eriksen & Eriksen, 1974) participants 77 must respond based on the direction of a central target arrow, which is flanked by performing the task at hand (Botvinick et al., 2001;Gratton et al., 2017). For example, in 107 trials where distractors cue the incorrect choice, attention to these distractors is reduced 108 in the following trial, which in turn also reduces the detrimental effects of 109 target/distractor incongruence on accuracy and response times (RTs). A second 110 consequence of this adjustment is that participants do not as effectively take advantage of 111 the information conveyed by distractors in trials where they are congruent with the target 112 and cue the correct response (Gratton et al., 1992;Shenhav et al., 2013). 113 How this theorised attention shifting mechanism operates, and how it influences 114 those processes which are active during perceptual decision--making, are not well 115 specified in these models. It is unclear whether attention refers to spatial or feature--based 116 attention, corresponding to response gain changes in stimulus--selective sensory neurons 117 (e.g., Reynolds & Heeger, 2009), or changes to how information provided by sensory 118 cortex is used to make a decision, which may instead occur across a network of parietal 119 and prefrontal areas (e.g., Afacan--Seref et al., 2018). It is also unclear whether other 120 adjustments associated with cognitive control are implemented following response 121 conflict, such as changes in the amount of sensory evidence required to make a decision 122 (e.g., Forstmann et al., 2008). Without extending conflict monitoring models to describe 123 these adjustments it is difficult to meaningfully test them against competing accounts, 124 such as those which explain congruency sequence effects as driven by associative learning 125 of stimulus--response associations (Hommel et al., 2004;Abrahamse et al., 2016;Schmidt, 126 2019). Accordingly, the primary aim of the current study was to develop a framework for 127 tracking rapid adjustments to decision--making processes that occur following response 128 conflict, in order to develop more specific and testable versions of conflict monitoring 129 models. 130 To characterise the adjustments that occur following response conflict, we recorded 131 electroencephalography (EEG) while participants completed a novel variant of the 132 Flanker task (here termed a modified Flanker task). In each trial participants responded 133 to two target stimuli, termed the S1 and S2 targets. The first target could be either a 134 congruent or incongruent stimulus. Shortly after responding to this S1 target, the S2 135 target was presented, which was always congruent. Importantly, the correct responses 136 corresponding to the S1 and S2 targets were not dependent on each other (i.e. participants 137 made two independent perceptual decisions in each trial). Here, we assessed whether 138 patterns of behavioural and neural responses for S2 targets differed by S1 congruency 139 (note that 'congruency' in our study refers to the target/distractor elements within S1 and 140 S2, respectively, and not the relation between S1 and S2). We adopted an EEG analysis to simultaneously track the neural correlates of multiple processes that are critical for 143 perceptual decision--making. This approach is directly inspired by evidence accumulation 144 models of decision--making, such as the diffusion model (Ratcliff, 1978;Ratcliff & Smith, 145 2004; Ratcliff et al., 2016), which formalise perceptual decision--making as a process 146 whereby sensory evidence in favour of each decision outcome is accumulated over time.

147
When this evidence reaches a set threshold, or 'decision bound' , associated with a decision 148 outcome, the motor action corresponding to that decision is initiated. Using our EEG 149 analysis framework, we could trace the time--courses of multiple processes described in 150 these models, including the representation of sensory evidence in visual cortex, the build--151 up of decision evidence, and preparatory motor activity corresponding to different decision 152 alternatives. Critically, the adjustments to decision--making processes that occur following 153 response conflict should be indexed by effects on these EEG measures. Our approach also 154 facilitates identification of distinct adjustments at different stages of the decision--making 155 process, each of which may have opposing effects on accuracy and response speed, which 156 are difficult to identify based on behavioural data alone (see Kelly 157 et al., 2020). 158 To track the representation of decision--relevant sensory evidence we presented 159 stimuli that consisted of overlaid left and right tilted gratings and asked participants to 160 judge which grating was of dominant (i.e. of higher contrast, as done by Steinemann et 161 al., 2018). In our modified Flanker task we presented S1 target and distractor stimuli that 162 could have dominant gratings in congruent or incongruent directions. The distractor 163 consisted of a central annulus that encircled a fixation cross, and the target consisted of a 164 larger annulus that encircled the distractor. This resulted in a difficult Flanker task that 165 required participants to ignore the centrally--presented distractor in incongruent trials 166 and make decisions based on the more peripheral target. In contrast to more typical 167 Flanker tasks, the distracting information was presented near fixation, and the target 168 stimulus was presented in the periphery. al., 2020). In these contexts, adjustments to pre--target levels of motor activity can produce 200 decision--making strategies that favour either fast or accurate responses.

201
By simultaneously tracking each of these neural markers, we could identify the 202 adjustments to decision--making strategies that correspond to attention shifts as described 203 in conflict monitoring models. If these attention shifts are associated with spatial or 204 feature--based attention in visual cortex, then we would expect to observe smaller SSVEP 205 sensory evidence measures evoked by the (always congruent) S2 targets when the previous 206 stimulus was incongruent, accompanied by a shallower build--up rate of the CPP. If the 207 attention shift relates to how sensory information is utilised after it is transmitted from 208 visual cortex, then we would expect to observe slower CPP build--up rates following 209 incongruent stimuli, but without co--occurring reductions in SSVEP amplitudes. If there 210 are also shifts in pre--target motor response preparation following response conflict, then 211 we would expect to see shifts in pre--target Mu/Beta amplitudes. Based on the notion that 212 attention shifts in conflict monitoring accounts reflect changes in spatial attention (e.g., 213 Janssens et al., 2017), we expected to observe slower RTs to S2 targets following incongruent 214 S1 targets (a post--conflict slowing effect for congruent stimuli; Gratton  . We included this measure to assess whether similar 223 markers of conflict signalling could be found when using our modified Flanker task, and 224 also to see whether frontal theta amplitudes for S2 targets would be modulated by conflict 225 at S1 (as reported by Jiang et al., 2018). 226 This also allowed us to assess the temporal profiles of frontal theta amplitudes in 227 relation to other neural correlates of decision--making processes. Notably, frontal theta 228 activity has been proposed as a neural correlate of decision evidence accumulation, as it 229 shows similar accumulation--to--bound profiles to the CPP ( markers of decision processes, we observed two potential contributions to frontal theta 236 power in addition to effects that are specifically associated with the detection of conflict. 237 To foreshadow our results, we identified reductions in the rate of evidence 238 accumulation following response conflict, indexed by build--up rates of the CPP following 239 S2 targets. Further evidence congruent with changes in evidence accumulation rates was 240 observed in a follow--up behavioural experiment. These findings provide initial evidence of 241 how cognitive control is implemented across decision--making circuits following response 242 conflict corrected--to--normal vision. 6 participants were excluded from analyses because they 258 achieved less than 60% accuracy in at least one experimental condition. An additional 2 259 participants were excluded from the EEG dataset: one due to excessively noisy data, and 260 one due to data loss following a technical error. This resulted in a sample of 29 participants 261 for behavioural data analyses, aged between 18--36 (M = 23.9), and 27 for EEG analyses. 262 Participants were compensated 25 AUD for their time. This study was approved by the 263 Human Ethics Committee of the Melbourne School of Psychological Sciences (ID 1750871). 264 265

Stimuli 266
Stimuli were presented using a gamma--corrected 24" LCD monitor with a refresh rate 267 of 60 Hz. Stimuli were presented using functions from MATLAB (Mathworks) and 268 PsychToolbox (Brainard, 1997;Kleiner et al., 2007). Code used for stimulus presentation 269 will be available at https://osf.io/eucqf/ at the time of publication. 270 The critical stimuli consisted of two overlaid gratings within a circular aperture, 271 presented against a grey background, similar to stimuli in . Each of 272 the two gratings were oriented 45° to the left and right of vertical, respectively ( Figure  1B).

273
The left--tilted grating contrast--reversed at a rate of 15 Hz; the right--tilted grating contrast--274 reversed at a rate of 20 Hz. The circular aperture was divided into two concentric circles; an 275 inner circle and an outer circle with radii of 1.32° and 4.70° of visual angle, separated by a 276 grey spacer ring. Hz, respectively, and were each presented at 50% contrast (Neutral Stimulus). The fixation 290 cross then changed to red, and 400 ms later the relative contrast levels of the two gratings 291 were changed to create the S1 target. C) The outer circle was the target stimulus, and the 292 inner circle was the distractor. Participants indicated which set of stripes in the target 293 stimulus was dominant (i.e. of higher contrast). In trials with congruent distractors the 294 dominant grating orientations were the same for the inner and outer circles. In trials with 295 incongruent distractors they were of opposite orientations. Participants were required to 296 respond within 800 ms following S1 target onset, after which a neutral stimulus was 297 presented for a further 600--800 ms. D) Following this the S2 target appeared, which always 298 consisted of a target and congruent distractor, and was presented until the response 299 deadline at the end of the trial. The direction of the dominant grating in the S2 target was 300 not dependent on that of the S1 target. The contrast level of the dominant grating could be 301 either weak (65%) medium (75%) or strong (85%) for both target and distractor elements. 302 After each trial participants received feedback on their response to the S2 target, or 303 feedback indicating they had responded too early or missed the S1 response deadline.

Procedure 306
Participants sat 100 cm from the monitor in a darkened room and were asked to fixate 307 on a central cross throughout all trials. Each trial included two task phases; each task 308 phase comprised a single target that required a response. The trial structure is depicted in 309 Figure  1A. In each trial, a white fixation cross appeared for 800 ms. Following this, both 310 gratings gradually increased in contrast from 0% to 50% over a 400 ms period. Both 311 gratings remained at 50% contrast for a further 1000 ms, during which the contrast levels 312 of both gratings were identical (i.e. the stimulus was "neutral", see Figure  1B). 400 ms 313 before the end of this period the central fixation cross changed colour to red, signifying 314 that the first target (here termed the S1 target) would soon appear. Immediately after this 315 neutral stimulus period, one of the gratings within each circle increased to 80% contrast 316 and the other decreased to 20% contrast. This contrast difference persisted for 400 ms, 317 after which the neutral stimulus was presented. In congruent trials, the stripes of higher 318 contrast were the same orientation for inner and outer circles; in incongruent trials they 319 were of opposite orientations ( Figure  1C). Congruent and incongruent trials were each 320 presented with 50% probability. Participants indicated which grating was dominant (i.e. of 321 higher contrast) in the outer circle (while ignoring the inner circle) by pressing keys on a 322 TESORO Tizona Numpad (1000 Hz polling rate) using their left and right index fingers. 323 Participants were required to respond within 800 ms of S1 target onset. Following the 324 response to S1 the neutral stimulus was presented for a further 600 ms.

325
Following this neutral stimulus period, the second (S2) target appeared at the time of 326 the next screen refresh when the phase--reversals of both gratings were in synchrony, which 327 occurred every 12 frames [200 ms]. The S2 target always consisted of targets and congruent 328 distractors, however the contrast difference between the higher and lower contrast gratings 329 varied across 3 levels: weak (one grating 65% contrast, the other 35%), medium (75/25%) 330 and strong (85/15%; for examples see Figure  1D). Participants indicated via keypress which 331 grating was dominant, as done following the S1 target. The S2 target was presented for 332 durations ranging from 1800--2400 ms depending on the timing of the S1 response, so that 333 the total presentation duration of the gratings (including both S1/S2 target and neutral 334 stimulus periods) was equated across trials. Importantly, the direction of the dominant 335 grating in the S2 target was independent of the direction of the S1 target within the same 336 trial, and participants were explicitly notified of this.

337
A feedback screen was then displayed for 1000 ms. Participants received "Correct" or 338 "Error" feedback depending on their response to the S2 target. If responses were made 339 prior to the S1 target or after the 800 ms S1 response deadline, then "Too Early" of "Too 340 Slow" feedback appeared instead. If no response was given following the S2 target, then 341 "No Response" feedback was presented. Correct/error feedback relating to the S1 target was 342 not provided, so that participants would not keep their response to the S1 target in memory 343 and match this to the feedback presented, which may have interfered with subsequent 344 decisions related to the S2 targets. 345 Participants completed 480 trials, split into 10 blocks of 48 trials each. Participants 346 were allowed self--paced breaks between blocks (minimum break duration 15 seconds). 347 Prior to the experimental blocks, participants completed a practice block of 12 trials, which 348 was repeated until participants demonstrated adequate performance. During this block, 349 participants received feedback on their responses to both S1 and S2 targets in each trial. 350 Trial order was randomised and proportions of trials with each S1 and S2 target type 351 combination were balanced within each block. 352 353

Analyses of Accuracy and RT Data 354
Trials with responses that were too slow or earlier than stimulus onset were removed 355 from the dataset. Only trials with correct responses and RTs of >100 ms were included for 356 analyses of RTs. We modelled proportions of correct responses using generalised linear 357 mixed effects logistic regressions (binomial family) as implemented in the R package lme4 358 (Bates et al., 2015). We modelled RTs using generalised linear mixed effects regressions 359 using a Gamma family and an identity link function, as recommended by Lo & Andrews 360 (2015). Given that error RTs are also diagnostic of decision processes within evidence 361 accumulation model frameworks (e.g., Ratcliff & Smith, 2004), we have also plotted mean 362 RTs for errors in Supplementary Figure  S1. 363 To test for effects of each factor of interest on accuracy and RT measures, we compared 364 models with and without that fixed effect of interest using likelihood ratio tests. For each 365 comparison, both models included all other fixed effects that would conceivably influence 366 the data, as well as identical random effects structures. Fixed effects of interest for S1 367 targets included S1 congruency (congruent, incongruent). In comparison models we also 368 included effects of S1 target orientation (left, right). Fixed effects of interest for S2 targets 369 included S1 congruency (congruent, incongruent) and S2 evidence strength (strong, 370 medium, weak). Additional effects included in both models with and without each fixed 371 effect of interest included additive and interactive effects of S1 and S2 target orientation 372 (left, right). The structure of each model and the coefficients of each fitted model are 373 detailed in the Supplementary Material. 374 375

EEG Data Acquisition and Processing 376
We recorded EEG at a sampling rate of 512 Hz from 64 active electrodes using a Biosemi 377 Active Two system (Biosemi). Recordings were grounded using common mode sense and 378 driven right leg electrodes (http:// www.biosemi.com/faq/cms&drl.htm). We added 6 379 additional channels: two electrodes placed 1 cm from the outer canthi of each eye, and 380 electrodes placed above and below the center of each eye. 381 We processed EEG data using EEGLab v13.4.4b (Delorme & Makeig, 2004). All data 382 processing and analysis code and corresponding data will be available at 383 https://osf.io/eucqf/ at the time of publication. First, we identified excessively noisy 384 channels by visual inspection (median number of bad channels = 1, range 0--7) and 385 excluded these from average reference calculations and Independent Components Analysis 386 (ICA). Sections with large artefacts were also manually identified and removed. We re--387 referenced the data to the average of all channels, low--pass filtered the data at 30 Hz 388 (EEGLab Basic Finite Impulse Response Filter New, default settings), and removed one 389 extra channel (AFz) to correct for the rank deficiency caused by the average reference. We 390 processed a copy of this dataset in the same way and additionally applied a 1 Hz high--pass 391 filter (EEGLab Basic FIR Filter New, default settings) to improve stationarity for the ICA. . After ICA we interpolated any excessively noisy channels and AFz using the 397 cleaned data (spherical spline interpolation). EEG data were then high--pass filtered at 0.1 398 Hz (EEGLab Basic Finite Impulse Response Filter New, default settings).

399
The resulting data were segmented from --2200 ms to 3800 ms relative to the S1 target 400 onset, and baseline--corrected using the interval of 400--600 ms prior to the S1 target. This 401 baseline period was used to ensure that ERPs evoked by the fixation cross colour change 402 did not influence baseline estimates. Epochs containing amplitudes exceeding ±150 μV at 403 any scalp channels were rejected (mean trials retained = 462 out of 480, range 413--479). 404 Numbers of retained epochs by condition are displayed in Supplementary  Table  S1. Data 405 were then converted to current source density (CSD) estimates using the CSD Toolbox 406 (Kayser & Tenke, 2006; m--constant = 4, λ = 0.00001). The resulting S1 target--locked epochs 407 were used for all subsequent EEG analyses. Only data from trials with correct responses to 408 S1 targets were used for analyses of S1 neural responses, as is typical in EEG analyses of 409 conflict task data (e.g., Cohen & Donner, 2013). Similarly, only trials with correct responses 410 to both S1 and S2 targets were included in analyses of S2 neural responses, to exclude 411 effects of errors and post--error adaptations (Wessel, 2017). 412 We then segmented the EEG data according to four time windows of interest: from --413 500 ms to 1000 ms relative to S1 target onset, from --1000 ms to 500 ms relative to S1 414 responses, from --200 ms to 1000 ms relative to S2 target onset, and from --700 ms to 300 ms 415 relative to S2 responses. Data were transformed into frequency domain representations 416 (using data from the entire trial) before these epochs were derived for SSVEP, Theta and 417 Mu/Beta analyses. 418 An overview of the EEG data analysis approach is depicted in Figure  2. Each EEG 419 measure (SSVEPs, CPP, mu/beta and frontal theta--band activity) corresponded to 420 successive stages of the hypothesised decision--making process (shown in Figure  2A). These 421 measured were compared across S1 congruency and S1 RT quantile conditions following S1 422 targets, and across S1 congruency and S2 evidence strength conditions following S2 targets 423 ( Figure  2B). and performance monitoring or conflict detection (indexed by frontal theta--band activity). 432 Grey text denotes whether each measure was measured relative to stimulus onset, the time 433 of the response, or relative to both of these time points. B) Different fixed effects of 434 interest were tested in the time windows following S1 and S2 targets. For the time window 435 following S1 targets effects of S1 congruency and S1 RT quantile were of interest. Then, a measure of evidence in favour of the dominant target grating orientation was 446 derived by subtracting the SSVEP signal corresponding to the higher contrast grating from 447 that of the lower contrast grating. Importantly, both the target and distractor elements of 448 the stimulus contributed to the SSVEPs corresponding to each grating orientation, and so 449 these measures index the sum of evidence across target and distractor elements. To avoid 450 biases due to differences in trial numbers across targets with different flicker frequencies 451 and S1/S2 orientation conditions, we averaged signals across trials of each S1 and S2 target 452 orientation combination separately, and then averaged the resulting signals within each 453 condition of interest. 454 For S1 target--evoked SSVEPs we compared congruent and incongruent conditions. For 455 S2 target--evoked SSVEPs we tested for differences by S1 congruency and S2 evidence 456 strength (weak/medium/strong). Differences in the magnitude of sensory evidence 457 favouring the correct response across conditions were tested for using mass--univariate 458 analyses. Paired--samples t tests were performed at all time steps between --500 ms and 1000 459 ms relative to the S1 target, and between --200 ms and 1000 ms relative to the S2 target. To . We analysed CPP pre--response amplitudes 477 using linear mixed effects regression models and tested for effects using likelihood ratio 478 tests as described above. Fixed effects of interest for S1 targets included S1 congruency and 479 S1 RT. Fixed effects of interest for S2 targets included S1 congruency and S2 evidence 480 strength. Additional effects in both models with and without each fixed effect of interest 481 included additive and interactive effects of S1 and S2 target orientation. The structures of 482 all models used in these analyses are detailed in the Supplementary Material. amplitudes across all trials with correct S1 responses for each participant, averaged across a 490 baseline period from --900 ms to --700 ms from S1 target onset. This baseline period was 491 used to minimise contributions of the fixation cross colour change to baseline estimates.

492
S2 stimulus--locked epochs and S1 and S2 response--locked epochs were created using this 493 decibel--transformed data. Frontal theta and Mu/Beta amplitudes were averaged across trials for each S1 and S2 502 target orientation combination separately, and then averaged across all S1/S2 combinations 503 within each condition of interest. We performed mass--univariate paired--samples t--tests 504 and corrections for multiple comparisons using cluster--based permutation tests as 505 described above. We did not have specific a priori hypotheses about how Frontal theta or 506 Mu/Beta responses would vary by S2 target evidence strength, and instead performed post--507 hoc mass--univariate comparisons to test for differences between strong and weak evidence 508 conditions. 509 510 3. Results 511

512
We observed typical effects of response conflict on behaviour. Participants were slower 513 and less accurate in responding to S1 targets with incongruent distractors (likelihood ratio 514 test p's < 0.001, Figures  3A,  3B). The delta plot in Figure  3C displays larger effects of 515 congruency on RTs for slower RT quantiles. This pattern is typical of both Flanker and 516 Stroop tasks (see Ulrich et al., 2015), but differs from delta functions observed in some 517 variants of Simon tasks (e.g., Ridderinkhof, 2002), Consistent with task conditions involving strict response deadlines, RTs for error trials 520 appeared to be slightly faster than correct responses, particularly in trials with congruent 521 stimuli (depicted in Supplementary Figure  S1). 522 Following S2 targets, participants were faster and more accurate in trials with targets of 523 higher evidence strength (p's < 0.001, Figures  3D,  3E). However, the addition of S1 524 congruency did not significantly improve model fits for accuracy (p = 0.484) or RT data (p 525 = 0.099). The estimate of the S1 distractor congruency effect indicated a tendency toward 526 post--conflict speeding rather than slowing (fixed effect of congruency = --6.2 ms). 527 528 529 530 531 532 533 534 535 536 537 Figure  3. Accuracy and mean RTs following S1 and S2 targets. A) Proportions of correct 538 responses following S1 targets with congruent and incongruent distractors. B) Mean RTs 539 for correct responses to S1 targets with congruent and incongruent distractors. C) Delta 540 plot displaying the [incongruent - congruent] differences in RTs for 10--90% quantiles in 541 steps of 5% on the Y axis and the average of each congruent and incongruent condition RT 542 quantile on the X axis. D) Proportions of correct responses to S2 targets, split by S2 543 evidence strength and S1 distractor congruency. E) Mean RTs for correct responses to S2 544 targets. 545 546 3.2 Neural Responses Following S1 Targets 547 Effects of S1 congruency were clearly visible in all EEG measures, except for motor smaller following S1 targets with incongruent distractors from ~50--500 ms after target 552 onset ( Figure  4A). This is to be expected, given that the distractor contained higher 553 contrast gratings of the opposite orientation to the target, and both the target and 554 distractor contribute to the SSVEP measures. Estimates of sensory evidence magnitudes 555 were also smaller in trials with slower RTs, as visible for both congruent and incongruent 556 conditions ( Figure  4A). To verify this, we tested whether SSVEP measures in trials with RTs 557 in the fastest tertiles were larger than those in the slowest tertiles, using a one--tailed 558 We found a significant cluster ranging from ~0--400 ms following S1 target onset. Note that 562 this test was done after observing the data, and such tests are circular and have an inflated 563 false positive rate (see Kriegeskorte et al., 2009), but that our results are similar to 564 previously--observed patterns in . 565 As expected based on the patterns of SSVEP results, CPP slopes preceding responses to 566 incongruent stimuli were shallower, t(26) = 6.81, p < 0.001. Additionally, pre--response CPP 567 amplitudes were smaller in trials with incongruent stimuli (likelihood ratio test p < 0.001). 568 Pre--response CPP amplitudes were also smaller in trials with longer RTs (p < 0.001). We Frontal theta--band (4--8 Hz) amplitudes were larger following incongruent stimuli from 573 400--800 ms relative to target onset, and from --450 to 50 ms relative to the time of the 574 keypress response ( Figure  4C; for time--frequency plots see Supplementary Figure  S2), 575 similar to that reported in previous studies using conflict tasks (e.g., Cohen & Donner,576 2013). 577 Plotting frontal theta amplitudes by S1 distractor congruency and fast/medium/slow 578 RT quantiles (displayed in Figure  4E) allowed us to better characterise the temporal 579 dynamics of this response (as done by Murphy et al., 2015). Theta amplitudes gradually 580 increased at a fixed rate over the course of the trial until ~100 ms before the time of the 581 response, after which frontal theta amplitudes rapidly decreased (see also  . This build--up was terminated 583 earlier in trials with faster RTs, leading to lower pre--response theta following congruent S1 584 targets, as participants responded more quickly in this condition (see also   Figure  S3). However, there were no significant effects of S1 595 congruency on motor preparation activity until 50--250 ms after the keypress response 596 ( Figure  4D). In addition, motor activity at electrodes contralateral to the response hand averaged across S1 congruent and incongruent conditions. 612 613

Neural Responses Following S2 Targets 614
3.3.1 Effects of S1 congruency 615 Although we did not find group--level differences in accuracy or RTs to S2 targets by S1 616 congruency, there was evidence of two opposing adjustments at different stages of the 617 decision process. Sensory evidence magnitudes in favour of the correct response (as 618 indexed by SSVEPs) were larger following S2 targets that appeared after incongruent S1 619 stimuli between ~200--400 ms post S2 target onset ( Figure  5A). However, CPP build--up 620 rates were slightly slower following incongruent S1 stimuli, t(26) = 2.25, p = 0.033 ( Figure  621 5B) indicative of a slower rate of decision evidence accumulation (Twomey et al., 2015). 622 Pre--response CPP amplitudes did not appear to differ between S1 congruent and 623 incongruent trials (likelihood ratio test p = 0.214). We did not observe effects of S1 624 congruency on measures of frontal theta or motor preparation responses ( Figures  5C,  5D) 625 and Mu/Beta amplitudes were very similar across conditions at the time of S2 target onset. 626 627

Effects of S2 evidence strength 628
Effects of S2 evidence strength were clearly visible across all neural measures. Sensory 629 evidence magnitudes scaled with the contrast differences between higher and lower 630 contrast gratings (i.e. evidence strength) from ~0--800 ms from S2 target onset ( Figure  6A). 631 The build--up rate of the CPP was also faster in trials with higher evidence strength, F(2, 52) 632 = 12.02, p < 0.001 ( Figure  6B) and pre--response CPP amplitudes were also smaller in weaker 633 evidence strength conditions (likelihood ratio test p = 0.020). Mu/Beta motor preparation 634 amplitude profiles also followed a similar pattern to the CPP, with steeper (negative--going) 635 build--up rates for stronger evidence strengths leading up to the response at contralateral 636 electrodes, but similar amplitudes around the time of response ( Figure  6D). 637 Similar to motor preparation profiles following S1 targets, there appeared to be a 638 gradual build--up of Mu/Beta amplitudes at both contralateral and ipsilateral electrodes 639 from 0--300 ms following S2 target onset. After this point, motor preparation trajectories 640 were determined by the evidence strength of the S2 target. However, motor preparation 641 amplitudes preceding S2 target onset were much less negative--going than those preceding 642 the S1 target (approx. --0.5 dB compared to --1.2 dB, relative to a --2 dB motor execution 643 threshold), indicating relatively less preparatory motor activity preceding the onset of the 644 S2 target. 645 646

Frontal theta amplitude profiles 647
Frontal theta amplitude profiles for S2 targets differed markedly from those profiles 648 observed following S1 targets. Theta amplitudes for S2 targets did not exhibit the rise--649 until--response pattern that was seen for S1 targets. Instead, theta amplitudes following S2 650 targets more closely resembled the trajectories of CPP and Mu/Beta amplitudes ( Figure  6C  651 left panel; for time--frequency plots see Supplementary Figure  S4). Notably, frontal theta 652 amplitudes in the time window immediately preceding the response did not appear to be 653 larger in trials with weaker evidence strength and by association slower RTs ( Figure  6C  654 right panel). To investigate this, we performed post--hoc mass--univariate comparisons of 655 response--locked theta amplitudes across strong and weak evidence strength conditions 656 (corresponding to trials with fast and slow RTs). Larger theta amplitudes were observed in 657 weak evidence strength conditions between --500 to --300 ms preceding the response 658 (uncorrected for multiple comparisons), associated with a less steep theta amplitude 659 build--up rate prior to the keypress response. However, these differences were not 660 statistically significant when applying a cluster--based correction for multiple comparisons. 661 Notably, Frontal theta amplitudes were not significantly larger in weak evidence conditions 662 from --300 ms to 50 ms relative to response onset, which is where the bulk of the effects in 663 S1 response--locked epochs were found. Instead, amplitudes in the weak evidence trials 664 tended to be smaller than in strong evidence trials. 665 In sum, the qualitative patterns of covariation between frontal theta, CPP and Mu/Beta 666 amplitudes differed across S1 and S2 task phases. Following S1 targets, there were larger 667 pre--response theta amplitudes in trials with slower RTs, due to a pattern whereby theta 668 amplitudes gradually rose at a steady rate until the time of the response (see also Tollner et 669 al., 2017). This also produced larger theta amplitudes for incongruent S1 stimuli, which co--670 occurred with smaller CPP amplitudes and highly similar Mu/Beta amplitudes prior to the 671 S1 response. In contrast, there were no clear theta--RT correlations for S2 targets, and theta--

694
We found evidence for two adjustments that occur following exposure to incongruent 695 stimuli: A temporary boost in the magnitude of sensory evidence represented in visual 696 cortex, indexed by effects on SSVEPs, and a slowing of evidence accumulation rates, 697 indexed by shallower CPP slopes. Here we note that, although CPP slopes significantly 698 differed by S1 congruency, the ERP waveforms in the group--averaged plots (in Figure  5B) 699 were highly similar, and we did not observe differences in RTs by S1 congruency. 700 One likely reason for why the observed effects on both the CPP and RTs were small 701 is that we only presented congruent stimuli at S2. Congruency sequence effects are 702 sometimes (but not always) smaller for congruent compared to incongruent Flanker al., 2013). Accordingly, we ran a behavioural follow--up experiment that also presented 705 incongruent stimuli at S2, which allowed us to further test the hypotheses that i.) evidence 706 accumulation rates are reduced following S1 incongruent stimuli, and ii.) the reduction in 707 evidence accumulation rates is associated specifically with the sensory information 708 provided by the distractor element of the modified flanker stimuli. 709 The latter hypothesis is associated with theorised shifts of attention away from the 710 distractor in conflict monitoring models (e.g., Shenhav et al., 2013). Assuming that 711 evidence accumulation rates reflect the sum of sensory evidence provided by the target 712 and distractor stimuli (e.g., as postulated in contemporary evidence accumulation models 713 designed for conflict task data, such as the Shrinking Spotlight Model; White et al., 2011, 714 2018), this would predict that accumulation rates would be relatively faster for S2 715 incongruent stimuli when S1 was also incongruent (due to reduced evidence weighting of 716 the unhelpful distractor), but slower for S2 congruent stimuli (whereby the distractor 717 would provide evidence for the correct response). Importantly, changes in the rate of 718 evidence accumulation would also predict larger congruency sequence effects on RTs in 719 trials with slower overall responses, indexed by shallower slopes of the delta functions for 720 trials in which S1 was incongruent. 721 722 5. Control Experiment: Effects of S1 Congruency 723 724 To further test for changes in evidence accumulation rates following response conflict, 725 we conducted a subsequent behavioural experiment (N = 29, 21 female, 8 male, aged 726 between 18--37, with 1 participant excluded from analyses due to poor task performance). 727 The trial structure was identical to the original experiment, except that incongruent 728 stimuli were also presented at S2, and that all S2 targets were of medium evidence strength 729 (for full methods and results see the Supplementary Material).

730
We observed the hypothesised pattern of effects. Differences in behavioural responses 731 to S1 targets by S1 congruency were consistent with the first experiment ( Figures  7A--C). For 732 RTs following S2 targets, there was an S1 congruency by S2 congruency interaction effect 733 (likelihood ratio test p < 0.001) reflecting the typical pattern of congruency sequence 734 effects (e.g., Gratton et al., 1992;Duthoo et al., 2013). Participants were faster when 735 responding to S2 incongruent stimuli when S1 stimuli were also incongruent (likelihood 736 ratio test p's < 0.001). For RTs to S2 congruent stimuli, there was little evidence of an effect 737 of S1 congruency (likelihood ratio test p = 0.620). Participants responded with similar 738 accuracy across S1 congruency conditions (see Figures  7D--E). 739 The S2 stimulus delta plots for S1 congruent and incongruent conditions revealed a 740 clearly visible reduction in the slope of the delta function when S1 was incongruent ( Figure  741 7F). For example, when the S1 target was congruent, the incongruent distractors in the S2 742 stimuli slowed RTs to a much larger degree in trials with overall slower (e.g., > 0.6 second) 743 RTs, compared to trials with faster RTs. This pattern is typical of Flanker tasks (Ulrich  et  744 al., 2015). However, this effect was much less pronounced (i.e. the delta function was 745 substantially shallower) when S1 was incongruent. Under the assumption that evidence 746 accumulation rates reflect a weighed sum of the sensory evidence provided by the target 1.0 S2 Congruent S2 Incongruent S1 C S1 I S1 C S1 I

S2 Congruent
S2 Incongruent S1 C S1 I S1 C S1 I of correct responses following S1 targets with congruent and incongruent distractors. B) 756

S2
Mean RTs for correct responses to S1 targets with congruent and incongruent distractors. 757 C) Delta plot displaying the [incongruent - congruent] differences in RTs for 10--90% 758 quantiles in steps of 5% on the Y axis and the average of each congruent and incongruent 759 condition RT quantile on the X axis. D) Proportions of correct responses to S2 targets split 760 by S1 and S2 congruency. E) Mean RTs for correct responses to S2 targets, split by S1 and S2 761 congruency. F) Delta plots displaying the S2 [incongruent - congruent] differences for each 762 RT quantile, plotted separately for S2 targets following S1 congruent and S1 incongruent 763 stimuli. Note that the slope of the delta function is shallower for trials whereby S1 was 764 incongruent. 765 766 6. General Discussion 767 768 To characterise the adjustments to decision--making processes that are implemented 769 following response conflict, we recorded EEG during a modified Flanker task and traced 770 the temporal dynamics of sensory evidence representation (measured using SSVEPs), following exposure to incongruent stimuli, including a boost in the magnitude of sensory 776 evidence represented in visual cortex and a small but statistically significant slowing of the 777 rate of decision evidence accumulation. Evidence consistent with reductions in evidence 778 accumulation rates following incongruent stimuli was also found in a foll0w--up 779 behavioural experiment, which also showed that this effect related to sensory evidence 780 provided by the incongruent (i.e. distracting) stimulus elements. Taken together, our 781 findings provide initial evidence that shifts of attention in conflict monitoring accounts are 782 implemented as adjustments to evidence accumulation rates, rather than other effects 783 which might be conceptually associated with attention.

784
In addition, we identified ramping motor activity at electrodes both contralateral and 785 ipsilateral to the response hand that started before S1 target onset ( Figure  4D  covariation with neural markers of evidence accumulation and motor activity, we discuss 795 some potential contributions to frontal theta amplitude measures that may be distinct 796 from effects that are specifically associated with response conflict. 797 798 6.1 Adjustments to Decision Processes Following Response Conflict 799 We found typical effects of stimulus congruency on behavioural and neural responses 800 to S1 targets in our modified Flanker task, consistent with existing work (e.g., Gratton et al., grating annulus that comprised the S1 distractor stimulus, which is conceptually similar to 822 shifts of attention away from distractor stimuli as described in conflict monitoring models 823 (Botvinick et al., 2004;Shenhav et al., 2013). Our results suggest that this so--called 824 attention shift might be implemented as a change in the accumulation rate of decision 825 evidence, and that this process is not necessarily a downstream effect of changes in the 826 magnitude of sensory evidence representations in visual cortex. However, we caution that 827 the observed size of these effects on the CPP were rather small, and did not co--occur with 828 expected differences in RTs across conditions. Before drawing any strong inferences, our 829 ERP results should be replicated in situations that produce large congruency sequence 830 effects on RTs, for example using the design in Jiang et al. (2018). 831 Despite these issues, we did find additional evidence consistent with post--conflict 832 changes in evidence accumulation rates in a subsequent behavioural experiment, whereby 833 we also presented incongruent S2 stimuli. Specifying changes in evidence accumulation 834 rates based on the ERP results allowed us to derive predictions regarding how congruency 835 sequence effects will be reflected in patterns of fast, medium and slow RTs. This is in 836 contrast to existing models, which could only generate predictions relating to mean RTs 837 (Botvinick et al., 2001;Shenhav et al., 2013). Based on the notion that post--conflict 838 adjustments specifically reflect a down--weighting of sensory evidence provided by the 839 distractor stimulus elements in the evidence accumulation process, we expected 840 congruency sequence effects on RTs for S2 targets to be progressively larger in trials with 841 slower overall responses. This would be indexed by less steep delta functions (indexing 842 effects of S2 congruency across fast, medium and slow RT quantiles) for trials where the S1 843 stimulus was incongruent, compared to when S1 was congruent. We observed precisely this 844 pattern of effects, lending further support to the notion that post--conflict adjustments are 845 associated with changes in how decision evidence is accumulated over time. This indicates 846 that cognitive control processes described in conflict monitoring theories (e.g., Botvinick the evidence accumulation stage of decision--making. It also suggests that the concept of 849 attention shifts in these models does not necessarily correspond to shifts in visual 850 attention as understood through, for example, normalisation models of attention (e.g., 851 Reynolds & Heeger, 2009). We propose that the concept of 'attention' in these accounts 852 should be revised to explicitly describe the decision--processes that are influenced 853 following conflict. 854 Interestingly, our findings of changes in evidence accumulation rates broadly agree 855 with mathematically--formalised evidence accumulation models that were developed 856 specifically to describe decision processes in conflict tasks ( to target and distractor elements separately, in order to better dissociate the dynamic 870 allocation of spatial attention from effects related to pupil dilation. 871 Another caveat is that we did not observe a slowing of responses to S2 stimuli following 872 incongruent S1 stimuli in the EEG experiment, which differs from reports of post--conflict 873 slowing effects in many (but not all) previous studies (e.g., Gratton  In our experiment there was a critical difference between S1 and S2 task phases that 909 determined how much this ramping motor activity influenced the timing of behavioural 910 responses. During the S1 task phase, when there was a strict (800 ms) response deadline, 911 the (negative--going) Mu/Beta amplitudes indexing motor activity had traversed almost 912 halfway to the motor execution threshold by the time of S1 target onset, meaning that even 913 small additional effects of ramping motor activity could trigger a keypress response 914 following S1 targets. However, this was not the case during the S2 task phase, whereby the 915 level of motor activity was further from the threshold at S2 target onset (see also of ramping motor activity (as collapsing decision bounds) may improve model fits in these 924 situations (e.g., Kelly et al., 2020), and should be considered when developing 925 computational models of response conflict effects (for a review of contemporary models 926 see White et al., 2018). 927 We also note that the extent of Mu/Beta at contralateral and ipsilateral electrodes to 928 the response hand did not appear to significantly differ by the presence/absence of 929 response conflict in the S1 stimulus. This may appear as evidence against conflict 930 monitoring accounts that presuppose differences in ipsilateral motor activity for 931 incongruent stimuli. However, it is unclear whether Mu/Beta activity is a direct correlate of 932 those motor--related signals that provide input to medial prefrontal cortex in conflict 933 monitoring accounts (e.g., Cohen, 2014). To better elucidate the role of Mu/Beta--related 934 motor action preparation in relation to conflict monitoring it may be informative to trace By tracing the temporal profiles of frontal theta amplitudes across S1 and S2 task 940 phases, we also provide preliminary evidence for two contributions to frontal theta 941 amplitude measures that occur in addition to effects of response conflict detection. We 942 note here that these are based on visual inspection of the data, and should ideally be 943 replicated and further characterised in future work. 944 The first source of frontal theta appeared to track motor activity associated with a task--945 related motor action, as quantified using Mu/Beta spectral amplitudes in our experiment. 946 This resembled the build--up of decision evidence, which is likely why frontal theta has 947 been previously proposed as a neural correlate of evidence accumulation (van Vugt et al., 948 2012; Werkle--Bergner et al., 2014). However, our results suggest that this source of frontal 949 theta correlates with motor activity rather than decision evidence accumulation (as in the 950 model of Brown & Braver, 2008). This is because both theta and motor activity began to 951 increase before the onset of the S1 target ( Figures  4C,  4D) before decision evidence would 952 begin accumulating in our task. This assumption could be further tested by assessing 953 whether accumulation--to--bound frontal theta dynamics are also found in perceptual 954 decision tasks that do not require motor responses (e.g., O'Connell et al., 2012). 955 The second source of frontal theta relates to the response deadline used in our task, 956 where response deadlines are also present in most existing EEG--based studies of conflict observed this following S1 targets. This dynamic produces larger theta amplitudes 963 immediately preceding behavioural responses in trials with longer RTs. However, following 964 S2 targets, and in a previous study that did not use a response deadline (van Vugt et al., Here we present two potential explanations for this pattern of theta amplitude profiles. 968 The first explanation is that, in our experiment, the rate of evidence accumulation was 969 slower in trials with incongruent S1 stimuli, as evidenced by the shape of the delta plot and 970 differences in CPP build--up rates. At the same time, motor activity gradually increased over 971 the S1 task phase at a similar rate for both S1 congruent and incongruent stimulus 972 conditions, which can be attributed to ramping effects of evidence--independent motor