跳到正文
HuggingFace Daily Papers(社區熱門論文)·· 2026-10-15AI 評分48

Memory as Weights: Internalizing Long-Term History for Streaming Videos

Memory as Weights: Internalizing Long-Term History for Streaming Videos

AI 導讀

MaLoW(Memory as LoRA Weights)將串流影片歷史逐步編碼至固定大小的 LoRA 權重,以權重承載長期記憶。其模型在 OVO-Bench 和 StreamingBench 分別取得 74.5% 和 84.8%,並在參數規模由 0.8B 至 11B 的基礎 VLM 上帶來一致的效能提升。

正文

💾 What if a streaming AI remembered the past in its weights—not its context window?

Excited to introduce MaLoW (Memory as LoRA Weights) in our new paper, Memory as Weights: Internalizing Long-Term History for Streaming Videos. MaLoW incrementally encodes video history into evolving, fixed-size LoRA weights. The pre-trained backbone stays frozen, while current observations remain in the input—separating long-term memory from real-time perception. 💾 History as weights. The current observations as input. 📥

Highlights from our experiments:
🏆 Strong streaming video performance: Our variants achieve 74.5% on OVO-Bench and 84.8% on StreamingBench, leading the methods compared in our paper.
🌈 Across base VLMs ranging from 0.8B to 11B parameters, MaLoW yields consistent performance gains.
🤖 Beyond video understanding: Incorporating memory improves robot-task success from 59.4% to 68.8% on Simpler WidowX.
📚 A new training resource: MaLoW-Instruct contains 381K question–answer pairs across 56K videos. 

Long-term memory is not just about what to remember—but where to store it.

🔗 Explore MaLoW:
📄 arXiv: https://arxiv.org/abs/2610.22070
💻 GitHub: https://github.com/dragonlzm/MaLoW
🌐 Project Page: https://dragonlzm.github.io/MaLoW

🤗 Hugging Face: https://huggingface.co/collections/zhuomingliu/malow
#StreamingVideo #VideoUnderstanding #MultimodalAI #LoRA #Robotics #interactionmodel

teaser_structure

來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co