HERMES: Real-Time Performance Benchmarking for Online Video MLLMs
Автор: AI Paper Review
Загружено: 2026-01-25
Просмотров: 41
Introducing the HERMES framework that manages KV cache hierarchically to innovate the understanding of streaming video. Unlike conventional methods that caused latency due to the use of external memory, this model utilizes an internal memory system to achieve instantaneous response speeds. In particular, the video token is efficiently compressed by focusing on the fact that the shallow, middle, and deep layers of the model play the role of sensory storage, task memory, and long-term memory, respectively. HERMES has increased accuracy by significantly reducing data volumes without additional learning, and has achieved up to 10 times faster response speeds compared to previous technologies. As a result, the system enables you to perform sophisticated analysis in real time without missing the key information of a long video, even within a fixed memory capacity.
https://arxiv.org/pdf/2601.14724
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
