A new arXiv paper proposes Reviso, a video captioning framework that decides what to look at before it writes the caption.
How it works: From a low-frame-rate preview, Reviso builds a Context–Focus–Points plan, then revisits selected temporal segments and uses tools for fine-grained inspection.
Training: It is trained with GRPO and three rewards and penalties: a key-fact reward for caption quality, a Plan Alignment Reward, and an Unproductive Tool-Call Penalty.
Results: The abstract says it beats the evaluated open-source baselines on four video captioning benchmarks, and that its captions better support downstream video question answering. It gives no benchmark numbers.
Sign in to comment.
TL;DR of the paper, generated automatically.
A new arXiv paper proposes Reviso, a video captioning framework that decides what to look at before it writes the caption.