Fig. 1: Design and computational models.

a After learning word–object pairs, participants performed a memory suppression task in which they were asked to prevent the memory of the images associated with the cue words from entering awareness. They then rated the presence or absence of intrusive memories during suppression attempts. The estimation that an upcoming cue will trigger an intrusive memory (i.e., belief) can be inferred from previous encounters, providing an adaptive advantage in the form of the deployment of optimum memory control and proactive prevention of memory retrieval (i.e., predictive control). Reactive control is engaged when intrusive memories unexpectedly cross the proactive gate, resulting in a prediction error (PE) that triggers additional inhibition and updating of future expectations. It should be noted that recall cues (i.e., think items) are not displayed here (see the “Methods” section). The apple and the chair items are selected from the Bank Of Standardized Stimuli (BOSS) and published under CC BY SA license (https://creativecommons.org/licenses/by-sa/3.0/)⁵⁵. b Toy example. Standard contrast analyses of intrusive and nonintrusive cues cannot identify the contribution of these critical computational quantities on the disruption of the connectivity markers of inhibitory control. c Computational model space. Binary intrusion ratings across the suppression task were fed into computational models to track belief formation across the suppression task. In the two-level hierarchical Gaussian filter (HGF; pale blue panel), beliefs are hierarchical and dynamically weighted by uncertainty. The perceptual parameter ω regulates the speed of belief adjustment throughout the task. The Kalman filter (KF; pale orange) also includes dynamic belief updating, which is regulated by two free perceptual parameters, π and ω, encoding belief reliability and uncertainty, but it does not assume hierarchical beliefs. The Rescorla–Wagner model (RW; pale yellow) is a simpler non-hierarchical model with a fixed, participant-specific learning rate α. The response model describes the log-probability of the behavioral outcomes (i.e., intrusion or nonintrusion rating) given beliefs through a beta density function. These trial-wise log-probabilities are used to compute model accuracy.