Reproduce “Claim-Level Rubric Rewards for Video Caption Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingqi Gao, Hongyuan Dong, Yifei Chen, Zhisheng Zhong, Zheng Ruan, Wenjin Hou, Yu Chen, Han Hu and 1 more
CURATE: Claim-Level Rubric Rewards introduces a framework for reinforcement learning in dense video captioning that uses claim-level rubrics to provide fine-grained, interpretable feedback. This approach improved factual faithfulness and descriptive density, achieving a 10.6-point gain over its baseline on EventHallusion and consistently enhancing video-language model pretraining by an average of 1.75-2.18 points across multiple datasets.
Key insights
- Decomposing complex caption evaluation into simpler, category-aware atomic claim verifications provides more reliable and precise reward signals than holistic or strictly reference-aligned evaluation.
- Reference-anchored calibration effectively balances the incentive for alignment with key reference content against the allowance for additional, visually grounded details, preventing both factual inaccuracies and unwarranted verbosity.
- Structured, fine-grained reward feedback guides the captioning model towards improved factual accuracy and comprehensive description, leading to higher quality captions that are also better for downstream VLM training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.