Boosting Vision-Language Models with Transduction Benoît Gérin