This repository provides official implementation of:

Decouple and Cache: KV Cache Construction for Streaming Video Understanding (ICML 2026)
Zhanzhong Pang, Dibyadip Chatterjee, Fadime Sener, Angela Yao

we propose Decoupled Streaming Cache (DSCache), a training-free cache construction mechanism that adapts pretrained offline models to streaming settings. DSCache maintains a cumulative past KV cache while constructing a separate instant cache on-demand, decoupled from past caches to preserve the informativeness of recent inputs. To enable position extrapolation beyond the training length, DSCache further incorporates a position-agnostic encoding strategy, ensuring KV caches to support unseen positions and preventing position overflow.

.
├── datasets        processed benchmarks
├── Baseline        offline MLLMs evaluation
├── Uniform_Cache   uniform streaming cache evaluation
├── DSCache         our DSCache
└── llava           code for Llava-OV

• Setup: 1x Nvidia A100 (80GB)
• Prepare the conda environment based on requirements.txt
• Download benchmarks under datasets/
  • StreamingBench
  • OVOBench

# more details can be found in run.sh.

cd DSCache

# Supported model: llava_ov_7b, qwen2.5-vl-7b
# Argument: context_len - cumulative cache size; frame_len - buffer size; stride - last frame resolution; continuous - use continuous position index or not; ltype - sampling stride for cumulative cache

python StreamingBench_last.py --model_name llava --context_len 16 --ltype ls_1 --frame_len 4 --stride 1.0 --continuous

@article{pang2026decouple,
  title={Decouple and Cache: KV Cache Construction for Streaming Video Understanding},
  author={Pang, Zhanzhong and Chatterjee, Dibyadip and Sener, Fadime and Yao, Angela},
  journal={arXiv preprint arXiv:2605.01858},
  year={2026}
}