Evidence-accumulation models have enabled strong theoretical advances in our understanding of decision-making. Yet most of these models assume stationary performance; when concerned with performance improvements, dynamic versions of these models are critical to inform the mechanisms of learning. This is particularly true because most domains of learning, from low-level perception to more complex skills, involve changes in both choices (e.g., accuracy) and the speed at which those choices can be made (i.e., response time). Evidence-accumulation models provide frameworks for understanding these joint changes. Using data from participants completing a dynamic random dot-motion direction discrimination task over 4 daily sessions for a total of 2800 trials, we characterized alterations in both perceptual sensitivity [drift rate] and response caution [boundary separation] across multiple possible timescales. Nonlinear mixed-effect Bayesian Wiener diffusion models were applied to estimate individuals’ trajectories of performance change in joint RT and accuracy distributions, with different models allowing for various dynamics. For example, the models could be stable across the entirety of the training, changing as a continuous function of trial number, or changing continuously within sessions, but varying across sessions. Model comparisons then used estimates of Bayes Factors as well as approximations to leave-one-out cross-validation. The best-fitting model of perceptual improvements included sensitivity changing as a continuous, exponential function of trial number (i.e., 1 through 2800), but response caution changing continuously within each daily session and in an independent manner across sessions. Such results uncover two different processes producing the full pattern of behavior across the entire trajectory of experience, one involving a continuous tuning of perceptual sensitivity, and another more idiosyncratic process describing how participants convert perceptual evidence into behavioral decisions. Theories of perceptual learning involving multiple timescales of change are thus supported, with distinct processes and outcomes associated with each time scale.