<p>Exploratory hierarchical policy-arbitration analysis examining continuous alignment with the multi-feature (MF) policy versus integrated -value representations across individuals and conditions.</p>

The -based policy representation was operationalized using posterior predictions derived from the fitted Markov Decision Process (MDP) model based on integrated state–action value differences (). The MF policy representation was operationalized using posterior predictions derived from the fitted hierarchical multi-feature model integrating reward probability (), threat probability (), and the boundary-condition features binary energy state (BES) and wait when safe (WWS).

Arbitration weights were estimated hierarchically across participants and task conditions (approach vs. avoidance) while additionally controlling for block number and trial-wise policy similarity. Trial-wise policy similarity was quantified using the absolute difference between the MF and -based policy predictions (), thereby accounting for trial-wise variation in overlap between the competing policy accounts. Importantly, policy-similarity distributions were highly comparable across conditions, indicating that the observed condition-dependent arbitration effects were unlikely to be explained solely by differences in overlap between the candidate policy representations across approach and avoidance environments. <b>(A)</b> Distribution of baseline arbitration weights across participants.

Most participants showed greater overall alignment with the MF policy representation, consistent with the primary model-comparison analyses reported in the main text. However, arbitration weights formed a continuous distribution without evidence for sharply separable strategy classes, suggesting graded inter-individual variability in relative policy alignment. <b>(B)</b> Posterior probabilities of condition-dependent shifts toward increased -based policy alignment in avoidance contexts.

Most participants showed posterior probabilities substantially above 0.5, indicating a consistent directional tendency toward greater alignment with integrated-value representations under avoidance conditions despite substantial overlap between the MF and -based policy predictions.</p> <p>(PDF)</p>