Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

HERMES: KV Cache as Hierarchical Memory for Efficient Streaming Video Understanding

About

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.

Haowei Zhang, Shudong Yang, Jinlan Fu, See-Kiong Ng, Xipeng Qiu• 2026

Related benchmarks

TaskDatasetResultRank
Video UnderstandingMVBench
Accuracy65.53
635
Visual Question AnsweringGQA
Accuracy57.6
524
Optical Character RecognitionOCRBench--
486
Visual Question AnsweringRealworldQA
Accuracy50.2
327
Streaming Video UnderstandingStreamingBench
Overall79.44
308
Long Video UnderstandingLVBench
Accuracy49.8
267
Text-based Visual Question AnsweringTextVQA
TextVQA Accuracy58
141
Real-Time Visual UnderstandingStreamingBench
Overall Score79.44
134
Long Video UnderstandingVideo-MME Long
Accuracy65
120
Visual Question AnsweringMMVP
Accuracy45.3
82
Showing 10 of 28 rows

Other info

GitHub

Follow for update