HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding
This work addresses the problem of real-time, low-memory video understanding for applications like streaming services, though it is incremental as it builds on existing MLLM frameworks.
The paper tackles the challenge of enabling Multimodal Large Language Models (MLLMs) to efficiently understand streaming video inputs by proposing HERMES, a training-free architecture that uses KV cache as hierarchical memory, achieving 10× faster time-to-first-token (TTFT) and up to 11.4% accuracy gains on streaming datasets while reducing video tokens by up to 68%.
Recent advancements in Multimodal Large Language Models (MLLMs) have demonstrated significant improvement in offline video understanding. However, extending these capabilities to streaming video inputs, remains challenging, as existing models struggle to simultaneously maintain stable understanding performance, real-time responses, and low GPU memory overhead. To address this challenge, we propose HERMES, a novel training-free architecture for real-time and accurate understanding of video streams. Based on a mechanistic attention investigation, we conceptualize KV cache as a hierarchical memory framework that encapsulates video information across multiple granularities. During inference, HERMES reuses a compact KV cache, enabling efficient streaming understanding under resource constraints. Notably, HERMES requires no auxiliary computations upon the arrival of user queries, thereby guaranteeing real-time responses for continuous video stream interactions, which achieves 10$\times$ faster TTFT compared to prior SOTA. Even when reducing video tokens by up to 68% compared with uniform sampling, HERMES achieves superior or comparable accuracy across all benchmarks, with up to 11.4% gains on streaming datasets.