Keze Wang

VLMs can now reason more faithfully in high-stakes domains like meteorology, thanks to a novel reinforcement learning approach that penalizes self-contradictory reasoning.