National audienceIn this paper, we explore how to control text generation at decoding time to satisfy certain constraints (eg. being non-toxic, conveying certain emotions...) without fine-tuning the language model. Precisely, we formalize constrained generation as a tree exploration process guided by a discriminator that indicates how well the associated sequence respects the constraint. We propose several original methods to search this generation tree, notably the Monte Carlo Tree Search (MCTS) which provides theoretical guarantees on the search efficiency.Through 3 tasks and 2 languages, we show that discriminator-guided MCTS decoding achieves state-of-the-art results without having to tune the language model. We also demonstrate that ot...