by Lioba C. S. Berndt, Daisy Crawley, Ruslana Tymchyk, Anna Hall, Jakub Onysk, Tore Erdmann, Elliott Wimmer, Isabel M. Berwian, Agnes Norbury, Quentin J. M. HuysBackground Behavioral Activation therapy is an effective treatment for major depressive disorder. Conceptually, the mechanisms through which it acts are thought to involve alterations to reinforcement learning, and several recent studies have supported this in laboratory settings. However, it remains unclear whether reinforcement learning mechanisms are involved in a realistic treatment setting. Methods In a randomized controlled observational study in the UK NHS talking Therapy service, 152 participants with low mood received reinforcement learning assessments using an affective Go/No-Go task. The assessment timepoints were randomized between subjects, with assessments occurring either before (control; n = 78) or during BA therapy (active; n = 74). Changes in Pavlovian biases were quantified using computational modeling. Results Anhedonia improved during treatment, but not before (p = 0.047; d = 0.53). The active treatment group showed significant changes in Pavlovian parameters compared to controls (pFDR = 0.012; d = 0.65). Pavlovian biases became more positive in the active group (M = 0.44, SD = 0.81) and more negative in the control group (M=−0.12, SD = 0.70). Changes in aversive Pavlovian bias statistically mediated treatment effects on symptoms of anhedonia (β=−1.77, p = 0.036, bootstrap; p = 0.0280). Conclusions Behavioral Activation modified Pavlovian biases early in treatment by simultaneously strengthening appetitive and diminishing aversive responses. Reduction in aversive bias statistically mediated improvements in anhedonia, suggesting a specific cognitive mechanism through which Behavioral Activation exerts its therapeutic effects. The study provides evidence that online measurements of cognitive function may have potential as mechanistic biomarkers.