---

🔥 News

• 2026.09.25 🎉🎉 Our V-CAST has been accepted to NeurIPS 2026!
• 2026.03.26 🤗🤗 We opened the V-CAST repository.

---

🎯 Highlights

• Curvature-aware spatio-temporal pruning: V-CAST allocates temporal budget according to video curvature and performs coordinate-preserving spatial pruning.
• Training-free and plug-and-play: V-CAST can be integrated into VideoLLMs without retraining.
• Coverage-oriented compression: V-CAST explicitly addresses discontinuous coverage and token-merging-induced position drift.
• Strong accuracy-efficiency trade-off: V-CAST preserves 98.6% of original performance, outperforms the second-best baseline by +1.1% on average, and reduces peak memory and total latency to 86.7% and 86.4% of vanilla Qwen3-VL-8B-Instruct.

---

✨ Overview

V-CAST is a training-free and plug-and-play curvature-aware spatio-temporal pruning framework for efficient long-context video inference. It revisits token compression from the perspective of spatio-temporal information coverage, and combines:

• Curvature-guided temporal allocation to route more budget to semantic turns and event boundaries.
• Coordinate-preserving spatial pruning to retain informative tokens without breaking the original (t, h, w) grid.
• Compatibility with VideoLLMs through a clean pruning-based design that avoids token merging drift.

---

🛠 Preparation

1. Clone the repository:

git clone https://github.com/xinyouu/V-CAST.git
cd V-CAST

2. Create the environment:

conda create -n vcast python=3.10 -y
conda activate vcast
pip install --upgrade pip
pip install -e ".[train]"

3. Install lmms-eval:

If you want to measure the latency and GPU memory, please use the custom installation.

cd lmms-eval
pip install -e .

Or you can also use the official installation.

pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git

---

🚀 Performance Evaluation

Our evaluation pipeline is built on top of lmms-eval, a unified toolkit for multimodal evaluation across text, image, video, and audio tasks. The current public release focuses on the Qwen3-VL evaluation path with V-CAST enabled by default, while additional model-family integrations will be released in follow-up updates. Some comparison baselines are also available in the open-source VidCom2 repository.

Quick Start

bash examples/v_cast/inference_qwen3vl_v_cast_64.sh

Core Public Paths

---

📌 Citation

If you find this repository useful, please cite:

@article{lin2026vcast,
  title={V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models},
  author={Lin, Xinying and Liu, Xuyang and Wang, Yiyu and Ma, Teng and Ren, Wenqi},
  journal={arXiv preprint arXiv:2603.27650},
  year={2026}
}

---

👍 Acknowledgment

We extend our gratitude to the open-source efforts of LLaVA-OneVision and Qwen3-VL.

---

📩 Contact

For any question about our paper or code, please email xinyinglin@slai.edu.cn.