Learning-based therapies, such as cognitive-behavioral therapy, are used worldwide, and their efficacy is endorsed by health and research funding agencies. However, the mechanisms behind both their strengths and their weaknesses are inadequately understood. Here we describe how advances in computational modeling may help formalize and test hypotheses regarding how patients make inferences, which are core postulates of these therapies. Specifically, we highlight the relevance of computations with regard to the development, maintenance, and therapeutic change in psychiatric disorders. A Bayesian approach helps delineate which apparent inferential biases and aberrant beliefs are in fact near-normative, given patients’ current concerns, and which are not. As examples, we formalize three hypotheses. First, high-level dysfunctional beliefs should be treated as beliefs over models of the world. There is a need to test how, and whether, people apply these high-level beliefs to guide the formation of lower level beliefs important for real-life decision making, conditional on their experiences. Second, during the genesis of a disorder, maladaptive beliefs grow because more benign alternative schemas are discounted during belief updating. Third, we propose that when patients learn within therapy but fail to benefit in real life, this can be accounted for by a mechanism that we term overaccommodation, similar to that used to explain fear reinstatement. Beyond these specifics, an ambitious collaborative research program between computational psychiatry researchers, therapists, and experts-by-experience needs to form testable predictions out of factors claimed to be important for therapy.