CapRL was proposed in September 2025 by a research team from the University of Science and Technology of China, the Chinese University of Hong Kong, and the Shanghai Artificial Intelligence Laboratory, among other institutions. The related research findings were published in the paper "...".CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning".

CapRL is a novel training framework that redefines description quality through practicality: high-quality descriptions should enable non-visual language models to accurately answer questions about the corresponding image. It employs a decoupled two-stage process where a large visual language model (LVLM) generates the description, while the objective reward derives from the accuracy of a separate, non-visual large language model (LLM) in answering multiple-choice questions based on that description. Pre-trained on the CapRL-5M description dataset annotated with CapRL-3B, CapRL achieves significant improvements across 12 benchmarks. Furthermore, within the Prism framework for description quality evaluation, its performance is comparable to Qwen2.5-VL-72B, outperforming the baseline by an average of 8.41 TP3T.