11 upvotes
New Benchmark Tests MLLM Memory in Long Video Chats
What is VideoMemBench: How Well Do MLLMs Remember in Video Conversations??
VideoMemBench measures how well multimodal large language models keep track of information across multi-turn conversations about hour-long videos. It probes whether models can retrieve and reuse facts introduced early in the dialogue.
Who it is for
Researchers and developers building video-aware assistants need a way to gauge long-term recall, so they can improve or compare their models.
How to try it
Visit the Hugging Face page https://huggingface.co/papers/2610.40131 to download the benchmark data.