Reproduce “Why Reinforcement Fine-Tuning Enables MLLMs Preserve Prior Knowledge Better: A Data Perspective” 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
Zhihao Zhang, Qiaole Dong, Qi Zhang, Jun Zhao, Enyu Zhou, Zhiheng Xi, Senjie Jin, Xiaoran Fan and 7 more
The authors examine post-training for learning new tasks while preserving prior abilities. Using 3×3 jigsaw puzzles, they compare reinforcement fine-tuning (RFT) with supervised fine-tuning (SFT) on different reasoning examples. For Qwen2.5-VL-3B, both rollout-based SFT methods reach 70% jigsaw accuracy, while RefCOCO validation scores are 84.6 for RFT-generated examples and 74.2 for GPT-4o-generated examples, supporting a data contribution to retention without showing data alone explains forgetting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.