Meta AI Researchers Introduce Token-Level Detective Reward Model (TLDR) to Provide Fine-Grained Annotations for Large Vision Language Models