MAIR Lab — Mila, Quebec AI Institute

The Multimodal Artificial Intelligence Research (MAIR) Lab at Mila advances the science of foundation models that can see, interact, and act in the physical world.

Our research explores how these models understand the visual world, and how they can be adapted through fine-tuning, parameter-efficient methods, reinforcement learning, and other approaches to unlock new capabilities. We apply these techniques across a range of multimodal tasks — from visual question answering and instruction-guided image editing to reasoning-intensive re-ranking and multimodal content generation.

Beyond developing methods, we create datasets and benchmarks that challenge models to reason deeply, generalize across modalities, and operate with cultural awareness in diverse global contexts.

Our goal is to move beyond surface-level recognition toward AI systems that truly understand, reason, and interact — bridging vision, language, and human values.

→ Explore our models and datasets to help shape the future of multimodal AI.