Reinforcement learning

Researchers propose VIGIL, a reinforcement-learning post-training framework that reduces hallucinations in multimodal large language models by maximizing mutual information between visual input and ge…