The impacts of agricultural research for development are long-term and very diverse – positive, unexpected, sometimes negative. To assess and understand these impacts, ImpresS, a participatory evaluation method that incorporates the viewpoints of actors on the ground, was tested on 13 research case studies conducted by CIRAD and its partners in different countries. The central conclusion is that research institutions and their funders need to change their practices if they wish to achieve long-term impacts at scale. For research, this means accepting to play multiple roles, collaborating with innovation and policy actors, fostering learning, and developing explicit hypothetical but plausible ex ante impact pathways. For sponsors and funders, it implies considering a wider range of impacts, planning action in the long term, fostering articulation between projects with similar goals, and supporting adaptive learning and management.