Pavlovian threat acquisition and extinction are fundamental processes by which individuals learn about threat and safety in their environment. Research has shown that humans learn more rapidly and persistently to associate threatening and—somewhat counterintuitively—positive rewarding stimuli with aversive events, supporting predictions derived from appraisal theories of emotion (Stussi et al., 2018; Stussi, Pourtois, et al., 2021). Here, the present study aimed to provide a confirmatory analysis of these findings and further characterize their algorithmic bases. Data from the four original experiments (N = 247) using a differential Pavlovian threat conditioning paradigm were combined and reanalyzed. In this paradigm, threat-relevant (angry faces, snakes), positive-relevant (baby faces, happy faces, erotic images), and neutral (neutral faces, colored squares) stimuli were used as conditioned stimuli, and skin conductance response was measured as an index of learning. Computational modeling was applied to identify signatures of learning biases in Pavlovian threat acquisition and extinction. An expanded model comparison indicated that a reinforcement-learning model differentiating between excitatory (learning from reinforcement) and inhibitory (learning from the absence of reinforcement) learning best explained the observed data. Although no evidence for differences in excitatory learning rates was found between stimulus categories, both threat- and positive-relevant stimuli exhibited a lower inhibitory learning rate compared to neutral stimuli, contributing to the persistence of the conditioned response during extinction. These results confirm the robustness of the original findings and further validate the appraisal-based approach, thereby informing the affective and computational determinants of Pavlovian threat extinction biases and their translational relevance.