Reproduce “PAVE: Patching and Adapting Video Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhuoming Liu, Yiquan Li, Khoi Duc Nguyen, Yiwu Zhong, Yin Li
PAVE introduces an efficient framework for adapting pre-trained Video Large Language Models to tasks requiring diverse side-channel signals such as audio, 3D, or high frame rate video. The method integrates these auxiliary inputs with minimal computational overhead, achieving state-of-the-art performance across various multimodal video understanding benchmarks.
Key insights
- Pre-trained Video LLMs possess a latent capacity for complex reasoning that can be unlocked by effectively leveraging relevant side-channel information, rather than requiring full retraining.
- Maintaining the original input token count for the LLM through a residual fusion mechanism is crucial for achieving high efficiency, preventing significant increases in the LLM's computational load.
- A parameter-efficient adaptation strategy can achieve state-of-the-art performance on a wide range of complex multimodal tasks, making powerful Video LLMs more versatile and practical to deploy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.