Reproduce “Group Relative Policy Optimization for Image Captioning” 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
Xu Liang
Xi'an Jiaotong University researchers demonstrate significant improvements in image captioning through Group Relative Policy Optimization (GRPO), achieving 2.4% better CIDEr scores while requiring 75% fewer training epochs compared to traditional self-critical sequence training methods.
Key insights
- Multiple candidate captions per image enable more robust optimization
- Group-based comparison reduces variance in advantage estimation
- KL divergence constraints help maintain model stability and caption diversity
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.