This paper addresses the problem of mastering the complexity of interacting with a large set of applications on smartphones. In one hand, number of applications increases. In the other hand, screen size reduces. To tackle this paradoxical evolution, we investigate adaptive user interfaces. We assume that it is possible to predict the applications of interest for a user in a given situation. Based on this hypothesis, our challenge is to accelerate user interaction when prediction is correct, without penalizing it when prediction is wrong. The paper proposes the concept of Evanescent Adaptation. The principle is a two-layer based representation: the predicted items (first layer) are displayed above the full list of items (second layer). The first layer is said to be evanescent in the sense that it automatically disappears progressively. The paper claims for putting this disappearing process under the control of the end-user. Thereby the user can close the first-layer as soon as s/he perceives prediction as irrelevant.