Boosting Reasoning in Large Multimodal Models via Activation Replay

If our project helps you, please give us a star ⭐ on GitHub to support us. 🙏🙏

🎉 News

• [2025/11/27] Arxiv is released.

🕹️ Approach

• we propose Activation Replay, a simple yet effective training-free solution that boosts multimodal reasoning of post-trained LMMs, without requiring expensive policy optimization. By modulation of visual tokens at test time, Activation Replay enforces RLVR low-entropy activations to mimick the distributions from paired base LMMs at test time.
• We validate our approach across diverse LMMs post-trained by RLVR and diverse reasoning scenarios, including mathmatics, multi-turn agents that perform visual search in high-resolution images, and video reasoners that think across frames, where our approach showcases consistent performance gains across these setups.

🔥 Performance

✒️ Citation

@article{xing2025boosting,
  title={Boosting Reasoning in Large Multimodal Models via Activation Replay},
  author={Xing, Yun and Hu, Xiaobin and He, Qingdong and Zhang, Jiangning and Yan, Shuicheng and Lu, Shijian and Jiang, Yu-Gang},
  journal={arXiv preprint arXiv:2511.19972},
  year={2025}
}

❤️ Acknowledgement

Thanks for their wonderful work!

• VLMEvalKit
• DeepEyes
• Video-R1