How does Bayesian reverse-engineering work?

Bayesian models of cognition and behavior are particularly promising when they are used in reverse-engineering explanations: explanations that descend from the computational level of analysis to the algorithmic and implementation levels. Unfortunately, it remains unclear exactly how Bayesian models constrain and influence these lower levels of analysis. In this paper, we review and reject two widespread views of Bayesian reverse-engineering, and propose an alternative view according to which Bayesian models at the computational level impose pragmatic constraints that facilitate the generation of testable hypotheses at the algorithmic and implementation levels.