How do we learn what features of our multidimensional environment are relevant in a given task? To study the computational process underlying this type of "representation learning," we propose a novel method of causal model comparison. Participants played a probabilistic learning task that required them to identify one relevant feature among several irrelevant ones. To compare between two models of this learning process, we ran each model alongside the participant during task performance, making predictions regarding the values underlying the participant's choices in real time. To test the validity of each model's predictions, we used the predicted values to try to perturb the participant's learning process: We crafted stimuli to either facilitate or hinder comparison between the most highly valued features. A model whose predictions coincide with the learned values in the participant's mind is expected to be effective in perturbing learning in this way, whereas a model whose predictions stray from the true learning process should not. Indeed, we show that in our task a reinforcement-learning model could help or hurt participants' learning, whereas a Bayesian ideal observer model could not. Beyond informing us about the notably suboptimal (but computationally more tractable) substrates of human representation learning, our manipulation suggests a sensitive method for model comparison, which allows us to change the course of people's learning in real time.