Confidence Freeze: Early Success Induces a Metastable Decoupling of Metacognition and Behaviour

Zhipeng Zhang; Hongshun He

cs.LGMar 22, 2026

Confidence Freeze: Early Success Induces a Metastable Decoupling of Metacognition and Behaviour

Authors: Zhipeng Zhang, Hongshun He

Organizations: China Mobile Research Institute, Beijing 100053, China · China Mobile GBA (Greater Bay Area) Innovation Institute, Guangzhou 510656, China

Abstract

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.

Figures & tables

Figure 1: Conceptual overview of the confidence-freeze framework. This schematic illustrates the three core components of the proposed mechanism. Left: Ideal learning system. Learners combine sequential outcome patterns (success and failure trajectories) to infer environmental stability, increasing switching after extended failure streaks and maintaining strategies after success streaks. Middle: Early success distortion. Exposure to unusually high success inflates the internal prior on environmental stability and reduces the effective weight placed on negative evidence, leading to weaker behavioural updating following failures. Right: Cognitive lock-in. When a strategy begins to fail, confidence decreases slowly, behavioural updating is suppressed, and switching is delayed, producing reversible episodes of lock-in. The framework illustrates how early experience can alter the mapping from evidence to policy adjustments, rather than the evidence itself, thereby generating a metastable learning mode.

Figure 2: Trajectory-level statistical evidence guides switching behaviour. Panels summarize how sequential outcome patterns shape behavioural adjustments. (a) Conditional switching probability increases monotonically with loss-streak length in the normal-success group, indicating sensitivity to accumulating negative evidence. The high-success group shows a flatter function, consistent with reduced integration of failure information. (b) Hazard-rate analysis reveals the instantaneous probability of switching at each streak length, showing parallel normative behaviour in controls but attenuated loss sensitivity among high-success participants. Error bars represent bootstrapped 95% confidence intervals. These results validate the hypothesis that humans use trajectory-level evidence when adjusting strategies and demonstrate that early success selectively distorts this normative sensitivity.

Figure 3: Early success induces behavioural lock-in. (a) Distribution of persistence lengths shows that high-success participants persisted through substantially longer runs of losses following reversals. (b) Group comparison confirms significantly longer mean persistence lengths in the high-success condition (Mann–Whitney

p=.008

). (c) Survival analysis reveals a lower switching hazard among high-success participants, indicating a delayed transition out of failing strategies (log-rank

p=.013

). (d) Single-participant trajectories illustrate that lock-in is not a stable trait: the same individuals exhibited deep lock-in at some reversals and adaptive switching at others. Together these results demonstrate that early success creates a state-dependent distortion in evidence integration, yielding reversible episodes of cognitive lock-in.

Figure 4: Metacognitive–behavioural decoupling reveals confidence freeze. (a) Trial-level confidence–behaviour relationship shows that switching probability declines with higher confidence, replicating normative metacognitive control. However, many high-success participants exhibit periods where confidence drops sharply yet behaviour remains fixed. (b) Confidence-freeze index is substantially elevated in the high-success group, quantifying the proportion of loss-streak trials in which confidence falls by ≥2 points while choices remain unchanged. (c) Representative participant trajectory illustrates the signature pattern of freeze: confidence declines as evidence accumulates, but switching does not occur. These dissociations reveal that early success disrupts the translation of metacognitive belief updates into behavioural adjustments, producing a metastable freeze state.

Figure 5: Mixed-effects modelling reveals altered integration of negative evidence. (a) Fixed-effect coefficients from the logistic mixed model show strong loss-streak sensitivity in the control group but a significant reduction in the high-success group (loss-streak × group interaction

p=.048

). Confidence reliably predicts switching (

p<.001

), and trial number captures mild temporal adaptation. (b) Model-comparison analysis indicates that including confidence improves model fit substantially (ΔAIC = 72.6), and the full model with random slopes provides the best overall fit. (c) Random-slope distributions reveal considerable individual variability in loss-streak sensitivity, supporting the interpretation of lock-in as a reversible learning mode rather than a stable trait. (d) Odds-ratio estimation confirms the causal effect of early success on lock-in likelihood (OR = 2.18,

p=.022

). Together these results formalize the mechanism illustrated in Fig. 1 : early success reshapes the internal weighting of negative evidence, altering the mapping from belief updates to behavioural switching.

Figure 6: Policy stickiness (

\phi

) across experiments and groups. Boxplots show the distribution of stickiness parameters estimated from the reinforcement learning model. High-success participants in Experiment 1 (baseline) exhibited substantially elevated stickiness compared to normal-success controls (***

p<0.001

). This elevation was significantly reduced in both Experiment 2 (explicit outcome trajectories) and Experiment 3 (metacognitive prompts), with no difference between the two interventions. Group means are indicated by white diamonds; individual participants are shown as overlaid points.

Exploration-Exploitation Tradeoff

The exploration-exploitation tradeoff studies how decision-makers balance gathering information with using what they know to earn rewards.

See all 59 papers

Neural Succession: A Mesoscopic Theory of Invasion, Coexistence, and Stabilization in Continual Learning
Sep 28, 2026Shoaib Ahmed Dipu, Md Salman Shamil, Sayeed Shafayet ChowdhuryContinual Learning Catastrophic Forgetting
Indiana University Indianapolis · North South University
Surprise as a Signal for Plasticity and Metacognition
Jun 30, 2026Louis MouchonEpisodic Memory Metacognition in Language Models
Independent Research
Last Step Matters: Early Uncertainty Cannot Predict Failure in Long-Horizon Agents
Aug 30, 2026Zongyue Li, Chengyue Yu, Lei Zang +3Deep Research Agents Long-Horizon Agent Evaluation
Zhejiang University · Ant Group

Sep 28, 2026J/K move · Enter open · S save

Neural Succession: A Mesoscopic Theory of Invasion, Coexistence, and Stabilization in Continual Learning

Shoaib Ahmed Dipu, Md Salman Shamil, Sayeed Shafayet Chowdhury

Indiana University Indianapolis · North South University

Continual Learning Catastrophic Forgetting

Open paper AskPDF arXiv

Continual learning is usually studied through mechanisms that preserve old knowledge. We develop Successional Learning Theory (SLT), a mesoscopic account in which the current representation is a resident community, the incoming task is an invader, forgetting is resident displacement, joint retention is coexistence, replay is resident reinforcement, and training moves from establishment toward stabilization. Its empirical coordinate is directional pre-invasion compatibility, measured on the resident model before the incoming task is learned. Across eight experiments, compatibility orders later forgetting on the 20 directed Split-CIFAR-10 transitions (three-repeat r=-0.789, incoming-task cluster 95% CI [-0.90,-0.72], every repeat alone r<=-0.67), forecasts held-out forgetting with 24% lower error than a no-information baseline, and reproduces under controlled MNIST permutations and CIFAR-10 rotations (r=-0.804, -0.718). On an 84-transition suite, compatibility separates coexistence from exclusion at every retention threshold (AUC 0.93-0.97). Replay repairs every transition with at most 325 stored examples and is most efficient where displacement is largest. Compatibility reaches |r|=0.720, while activation, representation, Jacobian, and fixed-coefficient Lotka-Volterra specializations do not. Plasticity and feature turnover fall reliably from early to late training (15/15 and 14/15 runs). We formalize a minimum habitat-modification bound, a displacement floor, a sufficient coexistence condition, an identifiability law with a range-restriction corollary, successional stabilization, and local reinforcement. The identifiability law also predicts where the coordinate loses leverage, and the prediction matches three CIFAR-100 partitions and five optimizer regimes. SLT is a pre-adaptation diagnostic that complements replay, regularization, and projection methods.

Abstract

Exploration-Exploitation Tradeoff

Explore similar work

Neural Succession: A Mesoscopic Theory of Invasion, Coexistence, and Stabilization in Continual Learning

Surprise as a Signal for Plasticity and Metacognition

Last Step Matters: Early Uncertainty Cannot Predict Failure in Long-Horizon Agents