The models from the paper "Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning", trained to play Super Mario Land