Why Reinforcement Fine-Tuning Enables MLLMs Preserve Prior Knowledge Better: A Data Perspective

@misc{zhang2025reinforcement,
  title = {Why Reinforcement Fine-Tuning Enables MLLMs Preserve Prior Knowledge Better: A Data Perspective},
  author = {Zhihao Zhang and Qiaole Dong and Qi Zhang and Jun Zhao and Enyu Zhou and Zhiheng Xi and Senjie Jin and Xiaoran Fan and Yuhao Zhou and Yanwei Fu and Tao Ji and Tao Gui and Xuanjing Huang},
  year = {2025},
  eprint = {2506.23508},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.23508},
}