The study introduces RAPO, a reflection-anchor policy optimization method based on GRPO, to strengthen the ability of vision-language models to retain visual information during long-chain reasoning. The method selects high-entropy reflection anchors based on token entropy and the visual propagation potential of suffixes, and optimizes a chain-masked, limited-window KL surrogate objective. Experiments show that RAPO outperforms strong baselines across multiple LVLM backbones and reasoning-intensive and general benchmarks.
Source:Hugging Face 社区论文精选 · Read original
Leave a Reply