MARC: Memory-Augmented RL Token Compression for Efficient Video Understanding
Published in International Conference on Learning Representations (ICLR) 2026, 2026
Recommended citation: Wu, P., Yu, Z., Liu, Y., Wu, C.H., Zhou, E., Shen, J. (2026). MARC: Memory-Augmented RL Token Compression for Efficient Video Understanding. In International Conference on Learning Representations. https://arxiv.org/abs/2510.07915
Peiran Wu, Zihong Yu, Yiling Liu, Ching-Hui Wu, Enmin Zhou, Junxiao Shen
This work presents MARC, a novel approach for efficient video understanding through memory-augmented reinforcement learning token compression, enabling better processing of long-form video content.