The paper introduces multi-reference image-grounded video captioning, requiring factual video descriptions with phrase-level grounding to multiple reference images. RefCaptioner uses a two-stage post-training pipeline combining mixed-data supervised fine-tuning with Hierarchical Coverage-Discounted GRPO. The authors construct a corpus of 20,000 videos and 171,354 reference images, and release MRVBench for evaluating factuality and multi-reference grounding across real-world and AI-generated videos. According to the abstract, RefCaptioner performs best overall among open-source models while remaining competitive on standard video-captioning benchmarks. Human evaluations also report preferred captions and more source-faithful video reconstruction.
No heat snapshots are available in the last 24 hours.