[论文] Spool: Low-Latency Streaming Inference for Hybrid Vision-Language Mode...
论文概要
研究领域: CV
作者: Rui Ning, Om Chabra, Hari Balakrishnan
发布时间: 2026-11-17
arXiv: 2611.23228
中文摘要
视觉无障碍代理和机器人等应用需要对最近视频内容(如过去两分钟)在有限时间窗口内及时回答问题。混合视觉语言模型(VLM)将注意力与门控 DeltaNet(GDN)等线性循环层交替使用,非常适合此场景,但维护滑动窗口开销很大——循环层将所有过去帧合并到共享状态中,过期旧帧通常需重新运行整个窗口。我们观察到循环更新对状态是仿射的,系数依赖输入,因此每帧的效果可捕获为一种可结合组合的变换。我们提出 Spool——免训练框架,缓存每帧状态变换并随窗口推进高效组合,无需回放帧即可从循环层状态中移除过期帧。跨 5 个开源混合 VLM 和 4 个视频推理基准,Spool 在全部 20 项测试中实现最低 p95 首字延迟,几何平均加速 6.6 倍,同时平均提升准确率 2.3 个百分点。
原文摘要
Applications such as visual accessibility agents and robotics require answering questions about recent video content promptly, over a bounded window of past frames (e.g., the last two minutes). Hybrid vision-language models (VLMs), which interleave attention with linear recurrent layers such as Gated DeltaNet (GDN), are attractive for this setting, but maintaining a sliding window is expensive. The reason is that recurrent layers combine all past frames into a shared state, so expiring old frames normally requires re-running the entire window through the model. We observe that the recurrent update is affine in the state with input-dependent coefficients, so the effect of each frame can be captured as a transformation that composes associatively. We present Spool, a training-free framework ...
自动采集于 2026-11-19
#论文 #arXiv #CV #小凯