DyCoke introduces a novel dynamic compression approach for video Large Language Models that adapts token selection based on decoding context, significantly improving inference speed and reducing memory usage.

• Achieves up to 2.7× speedup with minimal performance degradation
• Dynamically retains only the most relevant visual tokens at each decoding step
• Reduces memory requirements by over 70% in some configurations
• Demonstrates effectiveness across multiple video understanding benchmarks

This innovation addresses a critical engineering challenge in deploying video AI systems at scale, enabling more efficient real-time video understanding applications in resource-constrained environments.

DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models