Evaluative conditioning (EC) is defined as a change in the evaluation of a neutral „conditioned“ stimulus (CS) after its pairing with a valenced „unconditioned“ stimulus (US). EC effects are known to depend on associative memory for the CS-US pairings and to generalize to stimuli that share features with the originally presented CSs (generalization stimuli, GSs). In the present research, we introduce a multinomial processing tree (MPT) approach that combines these two lines of research. We introduce an MPT model (and its underlying measurement procedure) separating CS/US detection based on specific features vs. non-specific features (which are shared by CSs and GSs), associative memory for the specific US vs. US valence, and general response tendencies (guessing old vs. new, guessing positive US valence vs. negative US valence). We also present an experiment testing the effect of CS variability (during learning) on the generalization of EC as well as on the parameters estimated by the MPT model. Theoretical implications and possible applications in future research are discussed.