MARC: Memory-Augmented RL Token Compression for Efficient Video Understanding

The rapid progress of large language models (LLMs) has laid the foundation for multimodal models. However, visual language models (VLMs) still face heavy computational costs when extended from images to videos due to high frame rates and long durations. Token compression is a promising solution, yet most existing training-free methods cause information loss and performance degradation. To overcome this, we propose $\mathbf{\text{Memory-Augmented}}\mathbf{\text{Reinforcement}}\mathbf{\text{Learning-based}}\mathbf{\text{Token}}\mathbf{\text{Compression}}\mathbf{\text{(MARC)}}$, which integrates structured retrieval and RL-based distillation. MARC adopts a $\mathit{\text{retrieve-then-compress}}$ strategy using a $\mathbf{\text{Visual}}\mathbf{\text{Memory}}\mathbf{\text{Retriever}}\mathbf{\text{(VMR)}}$ to select key clips and a $\mathbf{\text{Compression}}\mathbf{\text{Group}}\mathbf{\text{Relative}}\mathbf{\text{Policy}}\mathbf{\text{Optimization}}\mathbf{\text{(C-GRPO)}}$ framework to distil reasoning ability from a teacher to a student model. Experiments on six video benchmarks show that MARC achieves near-baseline accuracy using only one frame’s tokens – reducing visual tokens by $\mathbf{\text{95\%}}$, GPU memory by $\mathbf{\text{72\%}}$, and latency by $\mathbf{\text{23.9\%}}$. This demonstrates its potential for efficient, real-time video understanding in resource-constrained settings such as video QA, surveillance, and autonomous driving.