A recurrent problem in the development of reasoning agents is how to assign degrees of beliefs to uncertain events in a complex environment. The standard knowledge representation framework imposes a sharp separation between learning and reasoning; the agent starts by acquiring a “model” of its environment, represented into an expressive language, and then uses this model to quantify the likelihood of various queries. Yet, even for simple queries, the problem of evaluating probabilities from a general purpose representation is computationally prohibitive. By contrast, this study embarks on the learning to reason (L2R) framework that aims at eliciting degrees of belief in an inductive manner. The agent is viewed as an anytime reasoner that iteratively improves its performance in light of the knowledge induced from its mistakes. By coupling exponentiated gradient strategies in online learning and weighted model counting techniques in reasoning, the L2R framework is shown to provide efficient solutions to relational probabilistic reasoning problems that are provably intractable in the classical framework.