CapRL++: Unified Reinforcement Learning with Verifiable Rewards for Dense Image and Video Captioning
Published in arXiv preprint, 2026
CapRL++ extends CapRL from dense image captioning to a unified image and video caption reinforcement learning framework. Instead of relying on fixed reference captions, CapRL++ uses verifiable downstream question-answering rewards: a caption is considered high quality when a text-only model can answer visual questions from that caption alone.
The framework introduces reward components for visual utility, temporal timestamp formatting, and length-aware regularization, enabling dense, chronological, and information-rich captions for videos.
Recommended citation: Penghui Yang*, Long Xing*, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Yibin Wang, Yujie Zhou, Jiazi Bu, Jianze Liang, Qidong Huang, Jiaqi Wang, Feng Wu, and Dahua Lin. (2026). "CapRL++: Unified Reinforcement Learning with Verifiable Rewards for Dense Image and Video Captioning." arXiv preprint arXiv:2606.09393. Submitted to TPAMI.
Download Paper