Confidence Freeze: Early Success Induces a Metastable Decoupling of Metacognition and Behaviour
This paper investigates why humans persist with failing strategies despite negative feedback, proposing 'confidence freeze'—a metastable state where early success decouples metacognitive confidence from behavior. Using a multi-reversal bandit task (N=332 across 3 experiments), the authors show that brief exposure to 90% success rates (vs. 60%) induces lock-in behavior where participants endure ~6 consecutive losses while reporting plummeting confidence, suggesting a dynamic mechanism rather than stable individual traits.
The paper presents an intriguing mechanistic account of maladaptive persistence with solid effect sizes (Cohen's d ≈ 0.7 for interventions, d = 0.97 for policy stickiness). However, the statistical evidence for the central claim—early success suppressing loss sensitivity—rests on a barely significant interaction (p = 0.048) that falls within the 'uncanny valley' of p-values. The experimental design is elegant, but confidence ratings sampled only every 3 trials provide coarse temporal resolution for a phenomenon emphasized as dynamic. The reproducibility claim is weakened by data being promised 'upon publication' rather than available now, and key test statistics (e.g., manipulation check t-values) are reported as placeholders (X.XX) in the provided text.
The convergent evidence across behavioral, metacognitive, and computational measures strengthens the core claim. The finding that the same individuals transition in and out of lock-in states across reversals is methodologically important—it directly contradicts trait-based explanations. The computational modeling results showing elevated policy stickiness (φ: 0.76 vs. 0.16) with intact learning rates (α actually higher in high-success group) support the specific mechanism of belief-action decoupling rather than general learning deficits. The dual-pathway intervention finding—environmental (explicit trajectory) and cognitive (prompts) showing equivalent efficacy—is conceptually valuable for applications.
First, the statistical threshold for the key interaction (p = 0.048) is concerning given the recent emphasis on p < 0.005 for new discoveries and the well-documented inflation of false positives near p = 0.05. This marginal significance, combined with moderate sample sizes per experiment (~99-123), warrants replication before mechanistic claims are accepted. Second, the sparse sampling of confidence (every 3 trials) risks missing the actual dynamics of the freeze state—participants could have switched confidence levels between ratings, making the 'freeze index' potentially noisy. Third, the paper lacks formal model comparison metrics (AIC/BIC) for the computational models; claiming the stickiness parameter captures the effect requires showing it outperforms simpler alternatives. Fourth, multiple comparisons across experiments and thresholds (Δ = 1, 2, 3) are conducted without clear correction procedures.
The evidence supports the broad phenomenon—early success induces persistence—but the specific 'confidence-freeze' mechanism requires stronger validation. The claim that confidence and behavior 'decouple' relies on measuring confidence at sparse intervals and inferring dissociation from group-level patterns. The paper appropriately compares to related work on exploration-exploitation trade-offs and sunk-cost fallacies, positioning confidence freeze as a dynamic alternative to trait-based accounts. However, the comparison to normative Bayesian updating (mentioned in Discussion) lacks formal hierarchical model fitting to participants' actual switching data; it remains a verbal theory rather than a quantified benchmark. The interventions in Experiments 2 and 3 are well-designed, but neither includes a no-intervention control group from the same subject pool, limiting causal claims about intervention efficacy.
Reproducibility is seriously compromised. The Data Availability statement indicates materials 'will be made publicly available on OSF upon publication'—meaning the data, code, and analysis scripts are not currently accessible for independent verification. This violates current standards for computational reproducibility. Additionally, the manuscript contains placeholder statistics (e.g., 't(97)=X.XX') that suggest incomplete reporting. Critical hyperparameters for the reinforcement learning model (initial values, fitting procedure, bounds on parameters) are not specified in the extract provided. The between-subject design across three experiments means effect sizes could be sensitive to unmeasured sampling differences. Without open data and code, independent reproduction is currently impossible.
Humans must flexibly arbitrate between exploring alternatives and exploiting learned strategies, yet they frequently exhibit maladaptive persistence by continuing to execute failing strategies despite accumulating negative evidence. Here we propose a ``confidence-freeze'' account that reframes such persistence as a dynamic learning state rather than a stable dispositional trait. Using a multi-reversal two-armed bandit task across three experiments (total N = 332; 19,920 trials), we first show that human learners normally make use of the symmetric statistical structure inherent in outcome trajectories: runs of successes provide positive evidence for environmental stability and thus for strategy maintenance, whereas runs of failures provide negative evidence and should raise switching probability. Behaviour in the control group conformed to this normative pattern. However, individuals who experienced a high rate of early success (90\% vs.\ 60\%) displayed a robust and selective distortion after the first reversal: they persisted through long stretches of non-reward (mean = 6.2 consecutive losses) while their metacognitive confidence ratings simultaneously dropped from 5 to 2 on a 7-point scale.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.