This paper proposes a novel framework to support the generation of strategies for multi-criteria longterm improvement. It can be applied to general preference models but it is illustrated in this article on a Multi-Attribute Value Theory model. The novel contributions to the literature are twofold. Firstly, the framework addresses the issue of resistance to change that may arise during the implementation of a strategy. It constrains a step of improvement to be focused on a single criterion, and minimizes the intensity of operational changes. Secondly, it addresses the realism of the improvement scenarios by treating three types of structural dependencies differently: the positive synergies, the negative synergies and the bottlenecks. The scenarios are generated by finding a set of efficient solutions to a shortest path problem in a graph whose edges represent possible steps of improvement. Each edge is characterized by an increase of rank or level and two penalty functions relating to the difficulty of its execution, one representing a risk associated to bottleneck mechanisms and the other to operational change relative to a previous edge. A case-study using the Shanghai Academic Ranking of World Universities is presented in order to illustrate how this framework could be useful to generate a sequence of strategic actions for the Université libre de Bruxelles.health care (Georges et al., 2015), etc.Within operational research, another field called Data Envelopment Analysis (DEA) also focuses on the comparison of entities based on multidimensional evaluations (Banker et al., 1984;Charnes et al., 1978). In this context, indicators are considered to be either inputs (that the entity consumes) or outputs (that the entity produces). Here the focus is to assess the efficiency of entities (that are called Decision Making Units) in relation to each other. Typically, the aim is to give a recommendation for an inefficient entity to become efficient (in one or several steps). DEA is strongly related to benchmarking which is a systematic search for the best practices that will help the compared entities improve and reach the best performances (Camp, 1989). If several consecutive performance modifications are suggested, it is called stepwise benchmarking.This has been introduced in order to ease the process of improvement, or to make it more realistic, with regard to the similarity of each intermediate step (see Park (2012)), to the context-dependency (see Seiford & Zhu (2003)), or to the risk of failure (see Ghahraman & Prior (2016)), among others (Alirezaee & Afsharian, 2007; Estrada et al., 2009; Lim et al., 2011). Let us note that all these contributions propose that each intermediary step match exactly an observed entity's performance. Alternatively, other works allow fictitious (not observed) steps (Lozano & Villa, 2005), (Lozano & Calzada-Infante, 2017).Stepwise benchmarking has mainly been studied for problems that fit the DEA assumptions (i.e. the existence of indicators that can be explicitly viewed as inputs and ou...