Neural mechanisms underpinning metacognitive shifts driven by non-informative predictions

Humans constantly make predictions and such predictions allow us to prepare for future events. Yet, such benefits may come with drawbacks as premature predictions may potentially bias subsequent judgments. Here we examined how prediction influences our perceptual decisions and subsequent confidence judgments, on scenarios where the predictions were arbitrary and independent of the identity of the upcoming stimuli. We defined them as invalid and non-informative predictions. Behavioral results showed that, such non-informative predictions biased perceptual decisions in favor of the predicted choice, and such prediction-induced perceptual bias further increased the metacognitive efficiency. The functional MRI results showed that activities in the medial prefrontal cortex (mPFC) and subgenual anterior cingulate cortex (sgACC) encoded the response consistency between predictions and perceptual decisions. Activity in mPFC predicted the strength of this congru-ency bias across individuals. Moreover, the parametric encoding of confidence in putamen was modulated by prediction-choice consistency, such that activity in putamen was negatively correlated with confidence rating after inconsistent responses. These findings suggest that predictions, while made arbitrarily, orchestrate the neural representations of choice and confidence judgment.


Introduction
Humans consistently make judgments and retrospectively assess the quality of their judgments.Making accurate metacognitive evaluations of one's knowing enables individuals to effectively monitor and adjust their behaviors.However, cognition and cognition about cognition may be susceptible to contextual factors, such as prior expectations, leading to decision biases (Bang and Rahnev, 2017).For instance, providing content-specific information about the object before stimulus presentation enhances the effective signal of visual input matching the expected content, thereby facilitating stimulus detection (Stein and Peelen, 2015).Moreover, this expectation-driven response bias further shapes metacognition, as the congruency between expectations and decisions leads to improved metacognition of perceptual decisions (M.Sherman et al., 2015).Integrating top-down expectations in decisions is important for us to adjust ongoing actions and improve the quality of decision-making, especially when the sensory inputs are noisy and ambiguous.Thus, it is crucial to better understand the neural mechanisms underlying this prediction-modulated decision process and introspection.
In most studies, predictions are induced based on the provision of valid information, such as the probability of stimuli occurrence (M.Sherman et al., 2015), visual primes (Stein and Peelen, 2015), and semantic hints (Hertz, Blakemore, and Frith, 2020), preceding the stimuli.As these predictions are constructed based on valid evidence, they have been shown to trigger attentional templates, thus facilitating perceptual decisions and the subsequent metacognitive evaluations of decisions.Evidence from brain research has shown that informative expectations directly affect stimulus representations in early sensory cortexes once the stimulus is displayed and even induce pre-stimulus sensory templates before stimulus onset (Kok, Brouwer, van Gerven, and de Lange, 2013, 2012, 2017).Importantly, this integration of prior information and early sensory process in perceptual decision-making may be implemented in high-level decision-related brain regions, particularly the prefrontal cortex (PFC) (Summerfield and de Lange, 2014).For instance, events with predictive cues compared with no cues revealed greater activity in dorsolateral PFC, and its enhanced effective connectivity with the sensory region when participants had an expectation about the upcoming stimulus (Rahnev, Lau, and de Lange, 2011).
It remains unclear whether humans still rely on predictions when no signal is provided to construct an informative prediction, i.e., when individuals make arbitrary and blind predictions.Evidence suggests that other factors that are not helpful in evidence accumulation also exert an impact on recognizing the stimuli and evaluating the accuracy of decisions.For instance, in a sequential perceptual choice task, after making a decision on the motion direction of a dot cloud, subjects' perceptions of the stimulus motion direction are systematically biased away from the decision boundary (Jazayeri and Movshon, 2007).Moreover, confirmation bias, which describes the finding that committing to a categorical choice biases subsequent decision-making, is also well documented in supporting perceptual bias induced by preceding decisions (Talluri, Urai, Tsetsos, Usher, and Donner, 2018).Indeed, perceptual choices depend not only on the current sensory input but also on one's own choice history, and choice history signals shift endogenous attention toward the previously selected interpretation (Urai, de Gee, Tsetsos, and Donner, 2019).All these biases in perception may serve as strategies for individuals to maintain a sense of internal consistency (Festinger, 1962;Luu and Stocker, 2018).We hypothesized that predictions, even when made arbitrarily, would bias one's perceptual judgment in favor of the predicted option.Since no valid information could be derived from such predictions for perceptual judgment, we anticipated that the influence of the prediction would manifest as a top-down modulation of brain responses aimed at maintaining self-consistency, rather than as a bottom-up sensory shift.Medial PFC (mPFC) has been extensively accounted for top-down cognitive control and behavioral monitoring in decision-making.In a recent model, mPFC is proposed as an action-outcome predictor that learns associations of actions and outcomes.Neural response in this region is then seen as a result of guiding adaptive behaviors, evaluating the probable and actual outcomes of actions, and adjusting emotional responses (Alexander and Brown, 2011;Euston, Gruber, and McNaughton, 2012).Hence, the top-down modulation of expectations on perceptual decisions may be tracked by mPFC, representing an attempt to integrate priors to optimize the decision.Particularly, the tendency for decisions to be consistent with expectations might reflect an internal reward process, which is encoded in the ventral mPFC that ensures actions are as expected (Rogers et al., 2004;Rushworth, Noonan, Boorman, Walton, and Behrens, 2011).In addition, as the expectation-driven decision process embeds strong motivational and emotional components that maintain consistency (Sharot, 2019), it might involve the participation of subgenual anterior cingulate cortex (sgACC), a region implicated in reward processing (Azab and Hayden, 2018;Etkin, Egner, and Kalisch, 2011).This hypothesis is supported by studies showing that neural responses in sgACC are related to monitoring the internal state (Gillath, Bunge, Shaver, Wendelken, and Mikulincer, 2005), regulating emotional responses (Ramirez-Mahaluf, Perramon, Otal, Villoslada, and Compte, 2018), and encoding of expected outcomes (Kosson et al., 2006).
Perceptual decisions are usually accompanied by metacognition, through which a decision-maker monitors the quality of a decision.While confidence in a decision is largely dependent on objective evidence, this subjective evaluation procedure is sensitive to priors and contexts.Individuals may adjust their confidence in decisions and even revise the initial decision given top-down modulations (Stephen M Fleming, Van Der Putten, and Daw, 2018;Schwartenbeck, FitzGerald, and Dolan, 2016;M. Sherman et al., 2015).As evidence has shown, confidence increases when expectations guide the decision (M.T. Sherman, Seth, and Kanai, 2016).Individuals might feel rewarded in keeping congruency between expectations and decisions whereas discrepancy between them indicates a mismatch of judgments.Research has shown that the reward-related process, which is represented in dopamine-rich midbrain regions, drives learning to adjust subsequent behaviors and beliefs to minimize future errors in predictions (Bromberg-Martin, Matsumoto, and Hikosaka, 2010;Schwartenbeck et al., 2016).
Predicting the upcoming stimulus, even arbitrarily, is a voluntary and self-determined action.This voluntary prediction may give individuals an illusion of control that they can predict future stimuli with high accuracy.If the prediction is matched with perceptual judgment, this congruency in decisions has the potential to boost confidence in their perceptual judgments.We hypothesized that if the assessment of confidence incorporates prior information, the reward-related regions would be activated.The putamen has been traditionally implicated in prediction error monitoring and performance adjustment on the detection of prediction errors in reward-related tasks (Daw and Doya, 2006;Hare, O'Doherty, Camerer, Schultz, and Rangel, 2008;McClure, Berns, and Montague, 2003;O'Doherty, Dayan, Friston, Critchley, and Dolan, 2003).Recent work has extended its function to internal reward representation built by social information and behavioral adaptation in the absence of feedback (Han, Huettel, Raposo, Adcock, and Dobbins, 2010;King-Casas et al., 2005;O'Doherty, Hampton, and Kim, 2007).For instance, Daniel and Pollmann (2012) found that activation in the putamen and nucleus accumbens was correlated with the prediction error of confidence, i.e., the difference between the expected and the actual confidence, in the absence of external feedback.Such involvement of putamen in self-generated internal signals related to confidence resembles a reinforcement learning process in which the prediction error signaling the associations between stimuli and outcome gives rise to arousal and guide learning behaviors (Haruno and Kawato, 2006;Wise, 2005).We expected that the encoding of confidence would be modulated by the prediction-choice consistency.
Here we aimed to delineate how the brain utilizes non-informative priors to guide perceptual decisions and confidence in decisions.Based on previous work, we first reasoned that the tendency to make decisions that are consistent with predicted stimuli might reflect a motivational control of behaviors to be consistent in responses and would elicit activity in top-down self-control regions.Further, the construction of confidence may involve a comparison process between priors and decisions such that individuals flexibly form their confidence based on whether decisions contradict or align with previously predicted responses.We expected interaction between response consistency and confidence over and above the strength of the perceptual signal itself.Specifically, the activity in the putamen following prediction error signals-the discrepancy between expected responses and perceptual decisions-should carry information useful for confidence judgments.

Participants
Twenty-four healthy volunteers (11 males; mean age ± SD: 25.92 ± 5.92) took part in the fMRI experiment.Twenty-nine healthy participants (10 males; mean age ± SD: 28.56 ± 8.56) took part in the behavioral experiment.All participants were right-handed and had normal or corrected-to-normal visions.Participants reported no history of psychiatric or neurological disorders, and no current use of any psychoactive medications.They received monetary compensation for their participation in the experiment.Participants gave written informed consent.The study was approved by the Institutional Review Board of National University of Singapore and was carried out in accordance with the approved guidelines.

Procedures
Participants conducted a face/house judgment task accompanied by predictions.Each trial began with a fixation cross for a duration ranging from 0.5 to 2.5 s (Fig. 1).After which, the word 'predict' prompted participants to make an arbitrary guess or prediction about the stimulus image of that trial by selecting either 'F' (for face) or 'H' (for house).They were instructed, "When making a prediction, simply go with your intuition and gut feeling".Participants were required to respond within a 2 s time limit.Their selection was highlighted with a white box, lasting for 2 s minus their reaction time.After prediction, the stimulus image was presented for 200 ms.Participants were then prompted to make the face/house judgment within 2 s.No feedback was given as to whether the face/house judgment was correct or incorrect.After that, participants indicated their confidence in the accuracy of their judgment, on a 1-4 scale (1 = 'very low confidence', 2 = 'low confidence', 3 = 'high confidence', and 4 = 'very high confidence').
Face/house stimuli were acquired from Stephen M. Fleming (Stephen M. Fleming, Huijgen, and Dolan, 2012).We used a set of 10 neutral faces (five males, five females) and 10 houses.Fourier transforms of each image were computed, producing 20 magnitude and 20 phase matrices.The average magnitude matrix of each face or house image was stored.A set of 100 stimulus images containing increasing amounts of noise was created for each face and house.This was achieved by recombining a variable proportion of white noise, P(noise), with the average magnitude matrix, such that P(image) = 1 -P(noise).P(noise) was adjusted in step sizes of 0.01.Hence, this produced 99 images of varying amounts of noise for each face and house, from the least ambiguous and noisy image [P(image) = 0.99] to the most ambiguous and noisy image [P(image) = 0.01].
A one-up two-down staircase procedure was used to control for the task difficulty.After one incorrect response, the amount of noise in the subsequent stimulus image was decreased by one step [(P(noise) = 0.01], decreasing its ambiguity/difficulty by one step.Conversely, after two consecutive correct responses, the amount of noise in the subsequent stimulus image was increased by one step, increasing its ambiguity by one step.Two independent staircases were maintained for face images and house images, independently (i.e., correct/incorrect identification of face images did not affect the ambiguity of house images, and vice versa).
Participants performed 50 trials of practice to familiarize themselves with the task before entering the scanner.The staircase procedure began with P(image) = 0.50 for both face and house staircases in practice.At the end of the practice, the last reached step for both face and house staircases were preserved.The main experiment continued the staircase procedure from these last reached steps.Participants completed 200 trials in the scanner, separated into 2 sessions by a 30 s interval.In each session, trials were presented in a random sequence, with face stimuli appearing in 50 % of the trials, and house stimuli in the remaining 50 %.

fMRI data preprocessing
Images were preprocessed using DPARSF (http://rfmri.org/DPARSF) and SPM12 (https://www.fil.ion.ucl.ac.uk/spm/) implemented in Matlab.The first eight volumes were discarded to allow for T1 equilibration.Preprocessing consisted of spatial realignment, normalization using the same transformation as structural images, and spatial smoothing using a Gaussian kernel with a full width at half-maximum (FWHM) of 6 mm.Two sessions of the scan were preprocessed independently.

General linear model (GLM)
We specified a GLM with six task-related regressors + six motorrelated regressors (GLM 1).The first two regressors aligned to stimuli onset and modeled face and house trials respectively, each parametrically modulated by the noise level of the image.Four other regressors aligned to perceptual decision response onset, modeling four types of events: match/correct, match/incorrect, nonmatch/correct, and nonmatch/incorrect.These events were parametrically modulated by corresponding reported confidence ratings, which were mean-centered within participants.Regressors were convolved with a canonical hemodynamic response function (HRF).Six motion correction parameters were entered as covariates of no interest along with a constant term per session.Low-frequency drifts were excluded with a 1/128 Hz high-pass filter.To further dissociate confidence-related brain activations, we reanalysed the primary GLM by adding a single separate regressor for confidence, without mean-centering (GLM 2).Finally, we built a GLM with eight regressors aligned to perceptual judgment response onset: prediction (face/house) × stimulus (face/house) × response (face/ house) (GLM 3).This GLM aimed to explore the effect of stimulus type, the effect of accuracy, the effect of response congruency, and their interactions.
Single-subject contrast images were entered into a second-level random-effects analysis using one-sample t-tests against zero to assess group-level significance.To identify brain activity with considerable individual differences in the effect of prediction on perceptual metacognition, we reported results surviving after a voxel-level height threshold at p < .005and cluster-level family-wise error (FWE) correction at p < .05.We also reported results at an uncorrected voxel-wise threshold of p < .001with a cluster-wise threshold of p < .05after FWE correction.In addition, small volume correction (SVC) with peak FWE corrected p-values (p < .05)was used on a priori regions of interest: medial prefrontal cortex (MNI, x, y, z = ±6, 45, 21 mm) (Hampton, Adolphs, Tyszka, and O'Doherty, 2007), subgenual anterior cingulate cortex (x, y, z = 5, 20, − 8 mm) (Will, Rutledge, Moutoussis, and Dolan, 2017), and putamen.These regions were implicated in choosing defaults.A spherical search space with an 8 mm radius was defined surrounding the independently defined coordinates reported above.The putamen was defined using the anatomic mask from the Wake Forest University Pick-Atlas toolbox (http://fmri.wfubmc.edu/software/Fig. 1.Perceptual decision task.In each trial, participants were asked to predict whether the image would be a face or a house in 2 s.After the prediction, the image was displayed for 200 ms, and the participant then judged whether the image shown was a face or a house in 2 s.Participants then rated how confident they were in their judgment in 2.5 s.All responses were immediately highlighted. C. Liu and R. Yu PickAtlas).False discovery rate (FDR, p < .05level) was applied for multiple corrections among ROIs where appropriate.Finally, the rfxplot toolbox (https://rfxplot.sourceforge.net/index.html) was used to extract parameter estimates in the region of interest for plotting data from within second-level analyses in SPM.
Nonparametric permutation statistics using the SnPM13 toolbox (https://www.nisox.org/Software/SnPM13/)were also conducted to verify the brain activation findings in the main GLM and to control for false positives.The first level contrast maps obtained from our main GLM for each subject were entered into a group-level one-sample T-test where cluster-wise inference was conducted using both the p < .005and p < .001cluster-forming thresholds, and then 5 % level FWE corrected thresholds.Variance smoothing was set as [6 6 6], and number of permutations was 5000.Equivalent prior regions of interest were applied as explicit masks for analyses.

Prediction biases choices and enhances metacognition of choices
In our study, the actual generation of the face/house stimuli was independent of the prediction.First of all, we validated that the ground truth of the stimuli remained unaffected by participants' predictions.When predicting faces, the face stimuli were presented in 49 % ± 0.9 % (mean ± SE) of trials, the house stimuli were presented in 51 % ± 0.9 % (mean ± SE) of trials, and there was no significant difference in the frequency of the two categories of stimuli, t(23) = 1.21, p = .237.Similarly, there was no difference in stimuli frequency when predicting houses, face trials (50 % ± 0.6 %) versus house trials (50 % ± 0.6 %), t (23) = 0.34, p = .738.These findings confirmed the independence of the predictions and the stimuli presented.
At the behavioural level, participants revealed a prediction bias in which the prediction biased subjects to choose predicted options more often.The proportion of trials in which the prediction response was congruent with the judgment response was 55 % ± 2 % (mean ± SE, match trials), whereas the incongruent proportion was 45 % ± 2 % (nonmatch trials).A paired t-test showed that the match proportion was significantly greater than the nonmatch proportion, t(23) = 2.74, p = .012(Fig. 2a).Moreover, there was a significant difference in perceptual performance between conditions, with the nonmatch condition (mean ± SE: 64 % ± 1 %) showing a higher level of accuracy compared to the match condition (mean ± SE: 61 % ± 1 %), t(23) = 2.15, p = .043(Fig. 2b).
To further examine the specific effects of predicting faces and houses, we analyzed the response frequency and accuracy of trials under four conditions: prediction (face or house) × perceptual response (face or house).For response frequency, the repeated measures ANOVA showed a main effect of prediction type, as participants predicted faces in 57 % (SD = 3 %) of trials and houses in 43 % (SD = 3 %) of trials, F(1,23) = 4.32, p = .049,η p 2 = 0.158 (Figure S1a).The main effect of response type was not significant, F(1,23) = 3.21, p = .086,η p 2 = 0.122.There was a significant interaction between the prediction type and response type, F (1,23) = 7.51, p = .012,η p 2 = 0.246.Specifically, when predicting faces, there was no significant difference in judging the stimuli as faces or houses, p = .489.However, when predicting houses, there were more house judgments than face judgments, p = .003.This interaction finding indicated that the prediction bias was mainly driven by the fact that participants were more likely to judge the stimuli as houses when predicting houses.We further examined task accuracy under these four conditions (Figure S1b C. Liu and R. Yu .104,η p 2 = 0.111.However, there was a significant interaction between the prediction type and response type, F(1,23) = 8.43, p = .008,η p 2 = 0.268.Specifically, when predicting faces, there was no significant difference in accuracy between face judgments and house judgments, p = .279.In contrast, when predicting houses, the accuracy of face judgments was significantly higher than the house judgments, p = .004.This finding indicated that the decreased perceptual performance in match trials was mainly due to the decreased accuracy in house responses when predicting houses.
In our study, although we used the continuous staircase procedure aiming at controlling task performance across trials, the actual result revealed an accuracy difference of 3 % between match and nonmatch trials.It is important to note that this performance difference was not caused by the objective sensory intensity differences.The actual displayed stimuli' sensory intensities were not different between the match (mean ± SE: 45.84 ± 1.52) and nonmatch (mean ± SE: 45.35 ± 1.65) conditions, t(23) = 1.19, p = .245.Further examinations of the prediction with a 2 (prediction: face vs. house) × 2 (response: face vs. house) ANOVA on sensory intensities showed no significant main effect of predictions [F(1,23) = 0.52, p = .477],no significant main effect of responses [F(1,23) = 3.54, p = .073],and no significant interaction between them [F(1,23) = 0.06, p = .816].
We further checked the distribution of confidence in relationship to both task difficulty (i.e., stimulus ambiguity) and task performance (i.e., accuracy), to examine the reliability of confidence judgments as indicators of participants' metacognitive abilities.Mixed effects logistic regression models were conducted, with the sensory intensity of stimuli and confidence rating serving as trial level predictors and a random intercept by subject.Stimuli data for faces and houses were analyzed separately in two models, given that independent staircase procedures were used for face and house images.The sensory intensity of the stimulus was centered within subjects in the model.The analysis of face stimuli showed a main effect of stimulus intensity (z = 2.20, p = .03)and a main effect of confidence rating (z = 6.28, p < .001) on the predicted probability of perceptual accuracy.The interaction between the two predictors was also significant (z = 4.55, p < .001),with the slopes of the logistic fit becoming sharper for higher confidence ratings (Figure S2).
Similarly, the analysis of house stimuli exhibited a main effect of confidence rating (z = 9.39, p < .001)and a significant interaction between two predictors (z = 2.37, p = .02).The slopes of the logistic fit for house judgment were larger for higher confidence ratings (Figure S2).In line with previous studies, these findings demonstrated that confidence judgments were reliable indices for reflecting task difficulty and perceptual performance in the current study.
The prediction bias did not affect the subsequent confidence rating, as there was no significant difference in confidence between match trials (mean ± SE: 2.23 ± 0.12) and nonmatch trials (mean ± SE: 2.18 ± 0.12), t(23) = 1.47, p = .156(Fig. 2c).However, results revealed an enhanced metacognitive efficiency (meta-d'-d') when predictions were congruent with perceptual decisions.The index of meta-d'-d' is a relative measure of metacognitive efficiency which quantifies the degree to which confidence ratings discriminate between correct and incorrect trials while controlling for first-order performance (d').There was a greater meta-d'-d' difference in the match condition (mean ± SE: 0.12 ± 0.09) than in the nonmatch condition (mean ± SE: − 0.29 ± 0.13), t(23) = 2.81, p = .010(Fig. 2d).However, the m-ratio (meta-d'/d') revealed no significant difference between the two conditions, t(23) = 1.72, p = .099(Fig. 2e).Considering that the m-ratio can yield rather extreme values when the denominator (d') is small or when the numerator (metad') is large, we employed a hierarchical estimation framework to estimate metacognitive efficiency.Our analysis showed a greater m-ratio in the match condition (m-ratio = 1.23) compared to the nonmatch condition (m-ratio = 0.78), with a 95 % HDI = [0.03,1.02]) (Figure S3).
Research has shown that staircase procedures with varied contrasts can lead to inflated estimates of metacognitive efficiency and this metacognitive inflation correlates with the degree of stimulus variability experienced by each participant (Rahnev and Fleming, 2019).In our study, we found no significant difference in the variability of sensory intensity between the match (SD = 11.14) and nonmatch (SD = 10.93)conditions, t(23)= 0.85, p = .405.Moreover, there were no correlations between sensory intensity variability and metacognitive efficiency in either the match (meta-d'-d': r = − 0.28, p = .191;meta-d'/d': r = − 0.19, p = .368)or nonmatch conditions (meta-d'-d': r = − 0.12, p = .584;meta-d'/d': r = − 0.07, p = .754).Therefore, there were no significant difference in stimulus properties between conditions in our study.
Finally, we assessed the strength of the prediction bias and its relationship with task difficulty, task accuracy, confidence rating, and metacognitive efficiency at the individual level.The strength of the prediction bias was defined as the proportion difference between match and nonmatch trials.First, no correlations were found between the prediction bias and task difficulty (overall difficulty, match difficulty, and nonmatch difficulty, all ps > 0.3).Second, although no correlation between the prediction bias and overall task accuracy (r = − 0.33, p = 0.121), a significant negative correlation was observed between the prediction bias and accuracy in match trials, r = − 0.62, p < .001,indicating that people are more likely to adjust their perceptual judgments to predictions when tasks become more difficult.(Fig. 3).Third, there were no correlations between the prediction bias and confidence rating (all ps > 0.5).Lastly, we found a positive correlation between the prediction bias and meta-d'-d' difference between match and nonmatch conditions, r = 0.53, p = .008(Fig. 3).In summary, while this prediction bias was associated with impaired perceptual performance, it facilitated subsequent metacognitive judgments.

Replication of behavioral results in an independent sample
To test the robustness of our results, we conducted a behavioral study in an independent sample of 29 participants.The behavioral task was the same as the fMRI task, except that the experiment was conducted outside the scanner.The result showed a prediction bias such that the match trials (mean ± SE, 54 % ± 1 %) were more than the nonmatch trials (mean ± SE, 46 % ± 1 %), t(28) = 2.73, p = .011(Fig. 4a).There was no significant difference in stimuli intensity between the match condition (mean ± SE: 35.28 ± 1.42) and the nonmatch condition (mean ± SE: 35.48 ± 1.59), t(28) = 0.25, p = .808.Further examinations of the prediction with a 2 (prediction: face vs. house) × 2 (response: face vs. house) ANOVA on sensory intensities showed no significant main effect of predictions [F(1,28) = 0.002, p = .967],no significant main effect of responses [F(1,28) = 3.11, p = .089],and no significant interaction between the two [F(1,28) = 1.95, p = .174].Moreover, we found no significant difference in the variability of sensory intensity between the match (SD = 13.90) and nonmatch (SD = 13.84)conditions, t(28) = 0.26, p = .794.
In brief, this independent behavioral experiment replicated our main findings of the fMRI experiment, such that prediction biased perceptual judgments while enhancing metacognition of perceptual judgments.
It is worth noting that the behavioral experiment revealed an overall higher accuracy (mean ± SE, 68 % ± 1 %) than the fMRI experiment (mean ± SE, 62 % ± 1 %), t(51) = 4.12, p < .001.These two experiments adopted the same 1-up 2-down staircase procedure, which should converge to an accuracy of 71 % theoretically (Levitt, 1971).However, both experiments revealed accuracy levels below the targeted 71 % (one-sample t-tests compared to 71 %, ps < 0.05).A plausible explanation is the limited number of trials in our experiments: each participant underwent 50 practice trials and 200 formal trials.Since independent staircases were used for face and house images, each category had only 25 practice trials plus 100 formal trials to adjust task difficulty.Indeed, analyses on the time course of trials showed that the latter half of trials had higher accuracy than the first half (p < .05),implying that more trials are needed to converge to the desired accuracy.In addition, our task was more demanding than simple perceptual tasks, as participants were required to guess the stimulus and rate their confidence in judgment, which may have interfered with performance.Finally, the scanning environment further deteriorated perceptual judgment, showing lower accuracy in the fMRI experiment.Of note, although we did not achieve the target accuracy, this does not compromise our main finding of improved metacognition in the match condition.Accuracy was controlled when estimating metacognition using either meta-d'-d' or meta-d'/d'.

Activities in mPFC and sgACC encode the congruence of predictions and perceptual decisions (versus incongruence)
Our behavioral results showed that participants were more likely to judge the image as predicted.We next turned to our fMRI data to investigate the brain regions that encoded this prediction bias.We focused on prefrontal regions which have been extensively involved in top-down modulated perceptual decisions (de Lange, Heilbron, and Kok, 2018;Hampton et al., 2007;Nicolle, Fleming, Bach, Driver, and Dolan, 2011;Will et al., 2017).
Fig. 5 depicts activation patterns in contrasting match versus nonmatch trials.Table 1 lists all significant peak activations related to the contrasts.When contrasting match > nonmatch, we observed increased activity in mPFC and sgACC (Fig. 5a).when contrasting nonmatch > match, we observed greater activation in the postcentral gyrus (Fig. 5b).
We next tested the neural correlates of individual differences in prediction bias using regression analysis.The strength of the prediction bias was defined as the proportion difference between match and nonmatch trials.We reasoned that if the mPFC acts to reflect the consistency between predictions and perceptual responses, activity in this region might predict the degree of the prediction bias across individuals.To test the hypothesis, we entered two between-subjects covariates in the whole-brain analysis of match > nonmatch activity.The first one was the accuracy difference between match and nonmatch trials, and the second one was the strength of the prediction bias (i.e., the proportion difference between match and nonmatch trials).Our results showed that behavioral prediction bias significantly modulated the match > nonmatch response in mPFC (Fig. 6 and Table S1).Activity in mPFC was positively correlated with the degree of prediction bias across individuals.However, the task accuracy showed no significant activations.No brain regions showed a negative correlation with prediction bias.
The whole-brain analysis did not find effects in face versus house contrasts.Correct versus incorrect responses also revealed no regions that could survive for FWE-correction at p = .05.Lastly, precuneus was found to be negatively correlated with confidence (peak MNI [− 12, − 60, 15], t = 4.18, 419 voxels).There were no supra-threshold clusters that positively correlated with confidence.

Reduced sensitivity to confidence level in putamen when participants' perceptual decisions are inconsistent (versus consistent) with their predictions
Our behavioral results showed an alteration of metacognition when perceptual decisions and predictions were congruent relative to incongruent.There are four situations here.When a participant judged the stimulus as predicted and reported high confidence, participants may feel that they have made a good prediction.However, when the participant responded consistently with the prediction but felt less confident in the choice, we consider it as a weaker signal of good predictions.By the same token, when a participant judges the stimulus inconsistent with the prediction but feels high confidence in the judgment, we consider it as a strong signal that participants may feel that they have made a bad prediction.When judgment is not aligned with prediction and subjects reported low confidence, we consider it as a weaker signal for a bad prediction.We expect activation in the reward region (i.e., putamen) would track prediction quality.
To test for the outlined interaction between response congruency and confidence, we contrasted the parametric modulator tracking participants' confidence ratings on match and nonmatch trials and found a significant effect in putamen (number of voxels = 27, peak MNI coordinate = [27, − 9, 12]) (Fig. 8a).To unpack the interaction effect, we extracted the β estimates at this cluster peak for each condition separately.We found that the interaction was characterized by significantly weaker activity in the nonmatch condition (β = − 1.74) than in the match condition (β = 0.09), F(1,23) = 14.34, p = .001. (Fig. 8b).To further illustrate the interaction between response congruency and confidence, the rfxplot toolbox was used to extract parameter estimates in the putamen (defined by an 8 mm sphere around the peak voxel [27, − 9,12]) for four confidence levels, separately for match and nonmatch conditions (Fig. 8c).Linear regression models were fitted to these average estimates, and results showed a significant decrease in activation across the levels of confidence in nonmatch trials (β = − 0.235, p = .021),but not in match trials (β = 0.016, p = .879).A further paired ttest on the slope parameters showed a significant difference between the match and nonmatch conditions, t(23) = 2.78, p = .011.In addition, we found no main effects on accuracy [F(1,23) = 0.16, p = .697]and no interaction between congruency and accuracy [F(1,23) = 0.368, p = .550],indicating that the brain activity difference between match and nonmatch conditions are independent of influence by objective performance.These findings suggest that putamen tracks participants' confidence reports more closely when their perceptual decisions are inconsistent than when they are consistent with their predictions.C. Liu and R. Yu The activation results of the GLM with eight regressors (a combination of 2 predictions × 2 stimuli × 2 responses) are reported in Table S4.The finding of vmPFC and sgACC in the match versus nonmatch contrast remained consistent with our primary GLM.Contrasting correct versus incorrect responses did not reveal brain activations in the prefrontal regions.Moreover, no activations were observed regarding the interaction of response congruency × accuracy.In brief, these findings indicate that the activations of vmPFC and sgACC are not attributed to the differences in task accuracy.

Discussion
This study aimed to examine the effect of noninformative predictions on perceptual and metacognitive judgments.In our work, the selfgenerated predictions were arbitrary and non-informative.Our paradigm is rather different from previous studies, where expectations are manipulated by providing valid cues (Stephen M. Fleming et al., 2012;Hertz et al., 2020; M. T. Sherman et al., 2016).Our objective was to investigate the impacts of merely making a prediction, while also controlling for the effects of any perceptual information that may be linked with these predictions.Therefore, in our approach, we asked participants to make uninformed predictions, predictions that are not founded on any valuable information.
Our behavioral results showed that participants were biased to judge the stimuli as predicted and such a response congruency between predictions and perceptual decisions further led to enhanced metacognitive sensitivity.The fMRI results showed that prefrontal regions including vmPFC and sgACC encoded the congruency of responses between predictions and perceptual decisions, and activities in vmPFC/mPFC and anterior insula further predicted the strength of the congruency bias across individuals.Importantly, there was reduced activity in putamen related to confidence ratings in incongruent events when compared with congruent events, indicating that the discrepancy between expected responses and perceptual decisions alters the neural representation of confidence strength.Altogether, our results demonstrated that prior expectations are flexibly integrated with perceptual decision-making and confidence evaluation, and the putamen reveals a critical contribution to metacognition shifts that are modulated by predictions.
Of note, we found inconsistent effects of predictions on confidence in the fMRI task and the behavioural task.In our fMRI experiment, prior predictions affected the perceptual judgments but not confidence.In an independent behavioral experiment, prior predictions led to decreased accuracy but increased confidence in the match condition.Although confidence did not reach statistical significance in the fMRI experiment, the observed trend was consistent with the behavioral experiment.This elevated confidence might reflect a confirmation bias in favor of information aligned with prior decisions (Rollwage et al., 2020).Indeed, this finding is in line with recent studies demonstrating that confidence is biased by suboptimal and false priors about perceptual decisions (Marcke, Denmat, Verguts, and Desender, 2022;Olawole-Scott and Yon, 2023).

Response consistency between predictions and perceptual decisions encoded in mPFC and sgACC
The fMRI results showed that response consistency between predictions and perceptual decisions was encoded in mPFC and sgACC.This finding indicates that the perceptual consequences of self-generated predictions are orchestrated by activities in the mPFC that track prediction-response consistency.Previous work has demonstrated that mPFC is vital for higher-order sequential relationship representations and evaluations, serving as a cognitive map of the top-down control of actions and generating adaptive responses (Alexander and Brown, 2014;Euston et al., 2012;Konovalov and Krajbich, 2018).Similarly, single-unit recordings also provide evidence showing that vmPFC signals internally driven motivational processes and reward-seeking choices (Bouret and Richmond, 2010;Strait, Blanchard, and Hayden, 2014).This interpretation is further supported by the finding that activity in vmPFC predicted the strength of response consistency at the individual level.Of note, these effects are not a result of a change in decision performance, as the comparison of correct and incorrect responses did not activate the mPFC.In brief, the BOLD signal in mPFC reflects a top-down contextual control to seek consistency and internal reward as making decisions as expected are deemed "good" predictions.
It is worth noting that some studies have shown the functional role of dorsal mPFC in representing conflicts and detecting errors (Holroyd, Coles, and Nieuwenhuis, 2002;Nee, Kastner, and Brown, 2011;Sun et al., 2017).We did not find dorsal mPFC responding to inconsistency.Previous research in conflict/error monitoring has primarily focused on learning tasks in which participants are given external and objective feedback indicating whether their actions are correct/good or not.In this type of tasks, feedback and action-outcome associations are salient and informative, and hence, participants are strongly motivated to pay attention to errors for the purpose of learning to enhance their performance as represented in mPFC.In contrast, in our task, participants generated the prediction with no external feedback about their predictions or perceptual decisions.As such, participants might evaluate the outcomes of a decision based on whether it is congruent with the prediction, reflecting an internal learning process to optimize behaviors.Our task may predominately elicit rewarding experiences in the consistent condition and weak conflict in the inconsistent condition (Euston et al., 2012).Indeed, a recent "action-outcome predictor" framework has been proposed to reconcile various findings, viewing mPFC as a region concerned with learning and predicting the likely outcomes of actions (Alexander and Brown, 2011).
In our study, another region that tracked response consistency was the sgACC.Prior work has demonstrated that activity in this region could be simply related to emotional regulation and arousal (Ramirez-Mahaluf et al., 2018;Rudebeck et al., 2014).Psychiatric patients with mood disorders reveal structural abnormalities in sgACC (Drevets et al., 1997;Drevets, Savitz, and Trimble, 2008).The clinical efficacy of brain stimulation targets for depression is negatively correlated with sgACC, demonstrating its intrinsic role for anti-correlated networks in depression (Fox, Buckner, White, Greicius, and Pascual-Leone, 2012).In addition, sgACC has been found to be involved in reward-related processes, such as encoding impending large rewards (Azab and Hayden, 2018), detecting positive valence of visual scenes (Vrtička, Sander, and Vuilleumier, 2011), and evaluating subjective rewarding outcomes in social contexts (Lockwood and Wittmann, 2018).Overall, the involvement of sgACC in both emotional and reward-related tasks indicates an "evaluative" role of sgACC in adapting behavior to emotional context.In our study, the activation of sgACC encoding an adaptive response that is consistent might reflect an internal reward with positive emotional experience, in accordance with previous findings.
The whole-brain analysis revealed significant activation in the postcentral gyrus when perceptual decisions were mismatched with predictions.This activation most likely reflects a motor switch component of the task (Smith, Taylor, Brammer, and Rubia, 2004).
It is worth noting that we did not find the confidence signal in the vmPFC, which is somewhat in contradiction with numerous studies in the literature (Morales, Lau, and Fleming, 2018;Shapiro and Grafton, 2020;Vaccaro and Fleming, 2018).However, a recent study by Hoven et al. (2022) demonstrated that motivational signals disrupted metacognitive signals in the vmPFC.Specifically, confidence was found to be correlated with the vmPFC only in the gain context, but not in the loss context.In light of this finding, we speculate that the motivation for pursuing consistency might inhibit vmPFC activation in encoding confidence in our study.Exploring the relationship between motivational processes and confidence estimation deserves future research (Hoven et al., 2022).

Different sensitivity to confidence in putamen for consistent versus inconsistent responses
We found that prediction-response consistency modulated the coding of confidence in putamen, such that activity in putamen was negatively correlated with confidence rating after inconsistent responses.This finding echoes well with the reinforcement learning model that describes how prediction errors drive learning and modify behaviors.In reinforcement learning tasks, the putamen has extensively been shown to encode prediction errors and related stimulus-action-reward associations, to promote further learning and mediate actions to seek reward (Daniel and Pollmann, 2012;Haruno and Kawato, 2006;McClure et al., 2003).Consistent with this notion, in our study, activation of putamen was negatively correlated with confidence when there was a prediction error, i.e., inconsistent response.This "error" signal was internally generated based on the mismatch between predictions and decisions.Such a negative correlation might reflect a motivational control of putamen to compensate for prediction errors so as to minimize further errors in confidence evaluation.In the nonmatch condition, at the behavioral level, the self-reported overall confidence ratings tended to be lower than those in the match condition.The negative correlation between confidence rating and putamen activity may indicate that lower confidence in the nonmatch condition is more rewarding.Our results, incorporating previous findings, highlight a flexible role of putamen in integrating information on prediction error with confidence judgment.
In general, the observed prediction bias in our study aligns closely with the predictive brain theories (De Ridder, Vanneste, and Freeman, 2014;Friston, 2012).Our findings are in line with a recent study demonstrating that priors hinder perceptual judgments but promote metacognitive judgments (Constant, Pereira, Faivre, and Filevich, 2023).In a dual-decision task, Constant et al. (2023) found that stronger informative priors in the first decision (lead decision) led to lower perceptual accuracy but higher m-ratio compared to stronger priors in the second decision (target decision), suggesting that priors have distinct effects on performance and metacognition.Whether the impact of priors on judgment is a conscious or unconscious process remains unclear.The brain has been described as a 'prediction machine' that attempts to match incoming sensory inputs with top-down expectations (O'Callaghan, Kveraga, Shine, Adams Jr, and Bar, 2016).In general, prediction biases perceptual decisions by involving both top-down conscious control prior to stimulus presentation and bottom-up sensory processing during evidence accumulation (Dunovan et al., 2014).
Finally, it is worth mentioning that the observed effects of prediction on perceptual and introspective judgments in our studies might be partly interpreted by factors such as priming or motor preparation.Specifically, as nothing useful could be learned from predictions in our experiments, participants might just randomly choose a category for prediction.Such an exerted prediction action might simply prime participants' response tendency and elicit motor preparation, thus biasing subsequent perceptual decisions.A control experiment in which the location of the face/house responses is randomized so that participants cannot prepare for the motor responses at the choosing stage before the presentation of stimuli would help disentangle the effects of prediction, Of note, the positive activation of the posterior gyrus was not significant.Activations were displayed using an uncorrected voxel-wise threshold of p < .005(warm color) and p < .001(green color) to show the full extent of the activations.
Fig. 2. Behavioural results of the face/house judgment task.The function of predictions on trials congruency proportion, accuracy of the perceptual decision, confidence rating, and metacognitive efficiency (mete-d'-d'), respectively."Match" refers to trials in which the prediction response and the judgment response are consistent whereas "nonmatch" refers to inconsistent prediction-judgment trials.Error bars reflect the standard error mean.

Fig. 3 .
Fig. 3. Prediction bias and performance.The strength of the prediction bias was negatively correlated with the task accuracy in match trials, but was positively correlated with the meta-d'-d' difference between match and nonmatch conditions.

Fig. 4 .
Fig. 4. Replication of behavioral results in an independent sample of 29 participants."Match" refers to trials in which the prediction response and the judgment response are consistent whereas "nonmatch" refers to inconsistent prediction-judgment trials.Error bars reflect the standard error mean.

Fig. 6 .
Fig. 6.Brain activity in vmPFC/mPFC [peak (x, y, z) = 0, 51, 18]  for the between-subject regression analysis considering the prediction bias (difference of trials proportion between match and nonmatch trials) as a covariate for match > nonmatch contrast.The scatter plots at right are not used for statistical inference (which were carried out in the SPM framework); and are shown solely for illustrative purposes.The color bars represent statistical t-values.Results were displayed using an uncorrected voxel-wise threshold of p < .005(warm color) and p < .001(green color) to show the full extent of the activations.vmPFC = ventral medial prefrontal cortex.

Fig. 7 .
Fig. 7. Confidence-related brain activations observed in GLM 2. Decreasing confidence was correlated with brain activity in the precuneus [peak MNI: − 2, − 60, 39].Of note, the positive activation of the posterior gyrus was not significant.Activations were displayed using an uncorrected voxel-wise threshold of p < .005(warm color) and p < .001(green color) to show the full extent of the activations.

Table 1
Whole brain activations of the GLM 1 with voxel-level height threshold at p < .005, 10 voxels extended.
Note.Abbreviations: MNI = Montreal Neurological Institute; L = left; R = right.aindicatesresults survived at an uncorrected voxel-wise threshold of p < .005withsmallvolume correction of p < .05peakFWEcorrected.False discovery rate (FDR, p < .05level)wasapplied for multiple corrections among ROIs where appropriate.bindicatesresults survived at an uncorrected voxel-wise threshold of p < .005with a cluster-wise threshold of p < .05afterFWEcorrection.cindicatesresults survived at an uncorrected voxel-wise threshold of p < .001with a cluster-wise threshold of p < .05afterFWEcorrection.C.Liu and R. Yu