Abstract Background Effective treatments for depression are available, yet many patients do not respond to treatment or experience relapse. Cognitive bias modification aims to ameliorate cognitive biases that contribute to the development and maintenance of the disorder. Objective This study examines the efficacy of a novel mobile approach-avoidance bias modification training with socioemotional cues for depression. Methods In this randomized clinical superiority trial, 75 inpatients with depression underwent 6 sessions of either active or sham approach-avoidance bias modification training with socioemotional cues over the course of 2 weeks alongside inpatient treatment as usual. The primary outcome was self-reported depressive symptoms, and the secondary outcomes included approach-avoidance bias based on reaction time and response force, anhedonia, and positivity. Outcomes were assessed before and after the training, and again at 2-week and 6-month follow-ups. The primary hypothesis was that active training would lead to a stronger decrease in symptoms of depression at the end of training. Results Both groups improved in depressive symptoms from baseline to the end-of-training assessment but did not differ in this regard ( B =−1.14, 95% CI −5.65-3.41; t 188.61 =−0.47; P =.64; d =−0.09, 95% CI −0.46-0.28). Changes in anhedonia, positivity, and approach-avoidance bias were also not different between training groups, neither at end of training ( P =.16) nor at the 2-week follow-up ( P =.69). However, the active training group showed a significantly greater reduction in depressive symptoms from the baseline assessment to the 6-month follow-up ( B =7.26, 95% CI 2.53-11.93; t 190.54 =2.95; P =.004; d =0.58, 95% CI 0.20-0.96). Permuted split-half reliability of the mobile assessment of approach-avoidance bias ranged from 0.77 to 0.94 for reaction times and from 0.81 to 0.93 for response force. Approach-avoidance bias was not altered by the training and did not mediate the training effects. Conclusions Mobile approach-avoidance bias modification training with socioemotional cues did not reduce depressive symptoms in the short term but did in the long term. Mobile training and assessment versions may be more feasible in the future, as they require no joystick setup and can be conducted on patients’ smartphones. Future work needs to further examine short- and long-term efficacy and the mechanisms driving long-term symptom change in larger multicenter trials.