Reproduce “Towards Universal Video MLLMs with Attribute-Structured and Quality-Verified Instructions” 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
Yunheng Li, Hengrui Zhang, Meng-Hao Guo, Wenzhao Gao, Shaoyong Jia, Shaohui Jiao, Qibin Hou, Ming-Ming Cheng
Researchers from Nankai University, ByteDance Inc., and Tsinghua University developed ASID-1M, a million-scale, attribute-structured, and quality-verified dataset of audiovisual instructions, alongside a progressive supervised fine-tuning scheme for video Multimodal Large Language Models. Their ASID-Captioner model, trained with this data, achieved state-of-the-art performance among open-source models and rivaled proprietary models on various fine-grained video understanding benchmarks.
Key insights
- High-quality, attribute-structured, and rigorously verified data is crucial for advancing universal video understanding in MLLMs, surpassing limitations of data quantity alone.
- A multi-stage, automated pipeline with systematic verification steps can effectively generate large-scale, fine-grained multimodal instruction data.
- Progressive supervised fine-tuning, starting with attribute-wise learning and extending to longer contexts, significantly enhances MLLM capabilities for fine-grained perception and instruction following.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.