Large vision-language models have shown strong performance in image captioning, but they frequently generate descriptions of objects that do not actually appear in the image. This phenomenon, known as object hallucination, undermines the reliability of captions produced by these systems.

The paper introduces a reinforcement learning framework that uses evaluation metrics as reward signals to steer the model away from hallucinated content. By optimizing directly for metric scores, the training process aims to align caption generation more closely with the factual visual content of the image.

The work is notable because it applies metric-driven reinforcement learning specifically to the hallucination problem in captioning, rather than relying solely on supervised fine-tuning or post-hoc filtering. This approach suggests a path toward more trustworthy outputs from large vision-language models.