ClaimDiff-RL: Fine-Grained Caption Reinforcement Learning through Visual Claim Comparison
The paper introduces ClaimDiff-RL, a novel framework for fine-grained caption reinforcement learning that addresses the reward granularity problem in long-form image captioning. By using reference-conditioned atomic claim differences as the reward unit, it allows for a more nuanced evaluation of captions, distinguishing between hallucinated claims and omitted facts. Experiments demonstrate that ClaimDiff-RL improves the balance between factuality and coverage, outperforming existing methods in several key areas.
- ▪ClaimDiff-RL uses reference-conditioned atomic claim differences as the reward unit for caption reinforcement learning.
- ▪The framework enables separate measurement of hallucinated claims and omitted salient facts.
- ▪Experiments show that ClaimDiff-RL improves the hallucination-missing-fact balance and surpasses Gemini-3-Pro-Preview in fine-grained capability dimensions.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20278 (cs) [Submitted on 19 May 2026] Title:ClaimDiff-RL: Fine-Grained Caption Reinforcement Learning through Visual Claim Comparison Authors:Tianle Li, Xuyang Shen, Yan Ma, Rongxin Guo, Shaoxiang Chen, Jiacheng Chen, Haochen Wang, Hongyang Tang, Yucong Zhou, Yu Cheng View a PDF of the paper titled ClaimDiff-RL: Fine-Grained Caption Reinforcement Learning through Visual Claim Comparison, by Tianle Li and 9 other authors View PDF HTML (experimental) Abstract:Long-form image captioning exposes a reward granularity problem in RL: captions are judged as whole sequences, while the important errors occur at the level of individual visual claims.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
