The temporal dynamics of metacognitive experiences track rational adaptations in task performance

The study investigates how multiple metacognitive experiences (e.g., boredom, effort, fatigue, frustration, conflict aversiveness) relate to cognitive control mechanisms and decision-making during sustained task performance. In cognitive tasks with conflict (e.g., Flanker), irrelevant stimuli can capture attention and impair performance. Traditional accounts attribute adaptations to conflict signals, whereas newer perspectives propose that negative metacognitive experiences drive adaptive control. The central research question is which subjective experiences track changes in latent decision-process parameters over time-on-task and how these experiences may support rational, strategic optimization of performance. Understanding these relationships can guide training programs and provide normative insight into how humans monitor and regulate cognitive processes over time.

Prior work on metacognition and cognitive control highlights subjective experiences (confidence, effort, fatigue, frustration, boredom) as signals for regulation. Conflict monitoring theory and affective-signaling accounts suggest aversive feelings associated with conflict can trigger adaptations. Reinforcement learning and neuroeconomic frameworks link performance costs to reward expectations and control allocation. Sequential sampling models (e.g., drift diffusion) decompose decision formation into drift rate, decision boundary (caution), and non-decision time, with conflict tasks requiring extensions that account for transient activation from irrelevant stimulus dimensions. Emerging views consider control as a resource-rational optimization problem balancing effort costs and reward-rate gains. Despite theoretical advances, empirical studies rarely track multiple subjective experiences alongside behavioral, model-based, and neural indices over extended time-on-task, motivating the present comprehensive investigation.

Design: Two complementary studies using a time-on-task (TOT) Flanker conflict paradigm across 18 blocks (~2 hours), with block-wise self-reports of metacognitive experiences. Participants: Behavioral experiment online via Prolific: N = 66 after exclusions (31 male, 33 female, 2 unknown; mean age 29.73, SD 5.26). EEG experiment at University of Granada: N = 45 after exclusions (16 male, 29 female; mean age 21.42, SD 2.75); 39 Spanish, 6 English. Ethical approvals obtained; informed consent collected. Task and procedure: Standard horizontal-arrow Flanker task. Response mapping: left-target arrow → S (online) / left control (EEG); right-target arrow → L (online) / right control (EEG). Trial timing: fixation 600–800 ms (mean 700); flankers 83 ms; target 33 ms; response window up to 1200 ms. Practice: 60 trials with trial-wise feedback; required ≥70% accuracy to proceed. Main task: 18 blocks × 260 trials (balanced congruent/incongruent; left/right). After each block: 30 s break (strict) and visual analog scales (VAS) assessing conflict aversiveness (preference for congruent vs incongruent; rescaled from 0–100 to −50 to 50, with higher indicating stronger dislike of incongruence), effort, frustration, boredom, fatigue (each 0–100). Behavioral preprocessing: Removed first trial per block, errors, RT < 200 ms, and trials outside 1st–99th RT quantiles. RT locked to Flanker onset. Modeling: Diffusion Model for Conflict (DMC) fitted per subject and block (260 trials) using DMCfun (Monte Carlo predictions with 20,000 trials per congruency; cost function on 19 RT percentiles per congruency and 5 Conditional Accuracy Function bins; Differential Evolution optimizer, 500 iterations). Parameter bounds: drift rate 0.2–1.2; decision boundary 30–110; non-decision time 200–600 ms; peak amplitude 10–110; peak latency fixed at 28; non-decision time variability fixed at 25; diffusion constant 4. Introduced PTB-ratio = peak amplitude / decision boundary to index irrelevant capture independent of boundary. Time-on-task analyses: Linear Mixed Effects models predicting DMC parameters, metacognitive experiences, and EEG indices by block (1–18), testing linear, logarithmic, quadratic, cubic trends and selecting via BIC. Random intercepts and slopes (up to first-degree) as appropriate; two-sided tests. Intra-individual relationships: Bayesian Multivariate Linear Mixed Effects models (brms; 8 chains, 4000 iterations, 1000 warmup; weakly informative priors N(0, 2.5)). Predict DDM parameter residual (detrended) time series by residual metacognitive time series, controlling for experiment (Behavioral vs EEG). Random slopes for all predictors; modelled correlations among dependent variables’ slopes and residuals. Report standardized beta, SE, 95% HDI, and probability of direction (pd). Detrending via removal of best linear/log trend per series. Optimality simulations: Pseudo-reward rate computed from DDM via simulations as (1 − ER) / (DT + NDT + ITI), separately for congruent and incongruent and summed. Landscapes over combinations of peak amplitude (20–120 in 0.1 steps) and boundary (20–120), and block-wise optimal boundary given other parameters. Defined optimal trajectories as maxima in landscapes. Explored resource-rational extension by adding cost proportional to boundary height. EEG acquisition and preprocessing: 64-channel actiCap (Brain Products), impedances <10 kΩ, online reference C1, 500 Hz sampling (downsampled to 250 Hz), band-pass 0.1–40 Hz, epochs −1000 to +1500 ms around Flanker onset. ICA (1 Hz high-pass for ICA stability); blink component identified via correlation with Fp1/Fp2, manually confirmed; automatic artifact rejection (autoreject) plus 150 μV peak-to-peak threshold (mean 3.70% trials removed, SD 4.21%); bad channels interpolated; average reference. EEG features: Lateralized Readiness Potential (LRP) via double difference ([C3left–C4left] − [C3right–C4right]), baseline −200 to 0 ms; Laplacian spatial filter. Extracted features as neural proxies: LRP slope (proxy for drift rate) measured between first incongruent negative peak and second incongruent peak; LRP amplitude (positive peak; proxy for boundary, averaged ±50 ms around peak); LRP latency (onset via segmented regression, single breakpoint 0–276 ms; proxy for non-decision time); LRP dip (absolute negative peak on incongruent trials; proxy for irrelevant capture). Representational similarity analysis: constructed null distribution from 50,000 random beta matrices; Spearman correlation between observed beta matrix and theoretical model matrices (independent-parameter and correlated-parameter).

• Time-on-task (TOT) effects on DDM parameters (N = 111):
  • Drift rate: slight logarithmic decrease, β = −0.058, SE = 0.023, 95% CI [−0.103, −0.014], t(110) = −2.57, P = 0.012.
  • Non-decision time: logarithmic decrease, β = −0.084, SE = 0.022, 95% CI [−0.128, −0.041], t(110) = −3.80, P < 0.001.
  • Decision boundary: strong logarithmic decrease, β = −0.325, SE = 0.023, 95% CI [−0.371, −0.279], t(110) = −13.88, P < 0.001.
  • Peak amplitude: strong linear decrease, β = −0.300, SE = 0.023, 95% CI [−0.345, −0.255], t(110) = −13.09, P < 0.001.
  • PTB-ratio: cubic pattern with early increase then decrease, plateau (β_linear = −0.305, SE = 0.037, P < 0.001; β_quadratic = −0.022, SE = 0.015, P = 0.058; β_cubic = −0.111, SE = 0.029, P < 0.001).
• EEG LRP features (N = 45) mirrored model parameters over TOT:
  • LRP slope: logarithmic decrease, β = −0.139, SE = 0.031, P < 0.001.
  • LRP amplitude: logarithmic decrease, β = −0.107, SE = 0.029, P < 0.001.
  • LRP latency: logarithmic decrease, β = −0.101, SE = 0.044, P = 0.029.
  • LRP dip: linear decrease, β = −0.227, SE = 0.035, P < 0.001.
• Intra-individual DDM–EEG relationships (beta matrix):
  • Drift rate ~ LRP slope: β = 0.145, 95% HDI [0.053, 0.231], pd > 99.99%.
  • Non-decision time ~ LRP latency: β = 0.085, 95% HDI [0.009, 0.159], pd = 98.77%.
  • PTB-ratio ~ LRP dip: β = 0.183, 95% HDI [0.092, 0.268], pd > 99.99%.
  • Boundary ~ LRP amplitude: not significant, β = 0.022, 95% HDI [−0.068, 0.115], pd = 67.79%.
  • Boundary related to LRP slope: β = 0.116, 95% HDI [0.017, 0.208], pd = 99.09%.
  • Representational similarity: correlated-parameter model significantly outperformed null (R = 0.637, 95% CI [0.206, 0.860], P = 0.008; permutation P = 0.005).
• Metacognitive experiences over TOT (N = 111):
  • Conflict aversiveness: mean above zero (b0 = 16.52, SE = 2.11, P < 0.001); no significant TOT trend (log β = −0.008, SE = 0.028, P = 0.839); subgroups showed increases or decreases.
  • Boredom: quadratic increase (β_linear = 0.300, SE = 0.025, P < 0.001; β_quadratic = −0.059, SE = 0.010, P < 0.001).
  • Effort: small linear increase (β = 0.128, SE = 0.029, P < 0.001).
  • Fatigue: cubic pattern (β_linear = 0.361, SE = 0.035, P < 0.001; β_quadratic = −0.074, SE = 0.009, P < 0.001; β_cubic = 0.118, SE = 0.023, P < 0.001).
  • Frustration: quadratic increase then plateau/reversal (β_linear = 0.184, SE = 0.027, P < 0.001; β_quadratic = −0.056, SE = 0.009, P < 0.001).
• Intra-individual DDM–metacognition relationships:
  • PTB-ratio positively related to conflict aversiveness (β = 0.050, 95% HDI [0.004, 0.092], pd = 98.54%) and frustration (β = 0.099, 95% HDI [0.048, 0.149], pd > 99.99%).
  • Fatigue negatively related to drift rate (β = −0.059, 95% HDI [−0.108, −0.011], pd = 99.23%) and boundary (β = −0.069, 95% HDI [−0.121, −0.018], pd = 99.58%).
• Rational optimization analyses:
  • Strong local coupling between peak amplitude and boundary (β = 0.829, 95% HDI [0.796, 0.862], pd = 100%).
  • Optimal boundary decreases with TOT given decreasing peak amplitude; observed boundary trajectory follows optimal direction.
  • Pseudo-reward rate increases with TOT: optimal trajectory (log β = 0.903, SE = 0.107, P < 0.001); observed trajectory (cubic: β_linear = 0.818, SE = 0.112, P < 0.001; β_quadratic = −0.222, SE = 0.112, P = 0.067; β_cubic = −0.326, SE = 0.112, P = 0.011).
  • Granger analysis: past peak amplitude predicts current boundary beyond past boundary (χ² = 3.53, P < 0.001); reverse non-significant (χ² = 1.08, P = 0.281).
  • Reward-rate determinants: decreases in peak amplitude primarily drive increases; boundary alone reduces reward rate if not co-adjusted; resource-rational model (adding boundary cost) accounts for observed suboptimality.
• Path analyses (final model): direct frustration → fatigue (β = 0.257, 95% HDI [0.189, 0.326], pd = 100%); boundary mediates frustration–fatigue (β_indirect = 0.003, 95% HDI [0.001, 0.008], pd = 99.70%).

Findings show that as time-on-task progresses, participants become less cautious (lower decision boundary) and less susceptible to irrelevant capture (lower peak amplitude), with modest decreases in drift rate and non-decision time. Preparatory motor signals (LRP) tracked these latent decision variables, supporting neural plausibility of the extended DDM. The interplay between decreasing peak amplitude and boundary adjustments aligns with a rational optimization strategy: adapting boundary downward in proportion to reduced irrelevant capture maximizes pseudo-reward rate. Metacognitive experiences mapped onto distinct aspects of this adaptive regulation: conflict aversiveness tracked the level of irrelevant capture, while frustration and effort related to adjustments of decision boundary, and fatigue indexed broader performance efficiency and may reflect the cost of sustained adaptive control. The representational similarity analysis indicated systematic overlap between EEG-derived indices and model parameters when accounting for inter-parameter correlations. Normatively, observed policies were somewhat suboptimal relative to reward-rate maxima, consistent with resource-rational behavior that incorporates costs of time/effort associated with conservative decision policies. The work emphasizes the importance of measuring interference relative to decision boundary (PTB-ratio) to avoid confounding and improve reliability across tasks.

Metacognitive experiences dynamically track and potentially guide rational adaptations in cognitive control during sustained conflict tasks. Conflict aversiveness provides insight into irrelevant capture, frustration and effort relate to strategic boundary adjustments, and fatigue reflects the burdens of ongoing optimization. Neural (LRP) indices corroborate model-based latent decision signals. The decision boundary adjustments appear tuned to changes in irrelevant capture to increase pseudo-reward rate, consistent with a resource-rational perspective. These results inform computational models of metacognitive experiences and the optimization of decision policies. Future directions include experimental manipulations to establish causal pathways (e.g., altering reward structures or conflict strength), refining neural correlates of decision boundary (beyond LRP amplitude), testing boundary optimality across diverse tasks and populations, and developing training interventions to enhance awareness and utilization of metacognitive signals for performance optimization.

The analyses are correlational, limiting causal inference, particularly in the path models; experimental manipulations are needed. The interpretation relies on specific normative and evidence-accumulation frameworks; alternative models may capture additional determinants. The decision boundary lacked a direct neural correlate in LRP amplitude, suggesting more distributed or complex neural representations. Suboptimality interpretations may also be influenced by individual differences (e.g., risk preferences) or criterion miscalibration.