Reducing negative impacts of stress, for example through mindfulness training, benefits physical and psychological well-being, and is becoming ever more crucial due to large-scale societal uncertainties (e.g., COVID-19). While extensive research has focused on mindfulness-related reductions in self-reported negativity, essentially no research has targeted task-based behavioral outcomes throughout long-term mindfulness trainings. Responses to emotionally ambiguous signals (e.g., surprised expressions), which might be appraised as either positive or negative, provide a nuanced assessment of one’s emotional bias across diverse contexts, offering unique leverage for assessing the effects of mindfulness. Here, we compared the effects of short- and long-term training via Mindfulness-Based Stress Reduction on ratings of faces with a relatively clear (angry, happy) and ambiguous (surprised) valence. Ratings became more positive for ambiguity from the start (Week 1) to end of training (Week 8; p <.001), but there were no short-term effects (from a single class session). This shift towards positivity continued through an additional eight-week follow-up (Week 16; p <.001). Notably, post-training valence bias (Week 8) was uniquely predicted by the non-reactivity facet of mindfulness (p =.01). Together, mindfulness promotes a relatively long-lasting shift toward positivity bias, which is uniquely supported by reduced emotional reactivity.