This paper presents the participation of the IRIT laboratory (University of Toulouse) to the Real-Time Summarization track of TREC RTS 2017. This track aims at exploring prospective information needs over document streams containing novel and evolving information and it consists of two scenarios ( A: push notification and B: Email digest). In this year the live mobile assessment was made available in real-time which provides opportunities for techniques based on active learning and relevance feedback. Our team submitted three runs for both scenarios. For scenario A, we explore two different approaches. The first one is a naive strategy that returns only the first tweet that matches all terms of the title of an interest profile. The second method is based on a binary classifier that predicts the relevance of an incoming tweet with respect to an interest profile. We examine the impact of the use of the live relevance feedback to re-train the classier each time new relevance assessments are made available. For scenario B, the summary generation is modeled as an optimization problem using Integer Linear Programming.