Reproduce “Mid-Training Language Models on Raw Video” 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
Jaedong Hwang, Xiaoqian Shen, Ernie Chang, Changsheng Zhao, Chong Zhou, Saksham Suri, Qi Qian, Zechun Liu and 4 more
Video mid-training lets a pretrained language model learn from ordered frame-derived visual tokens by predicting each next feature vector without captions or a text loss. In Qwen3-1.7B, after shared image-text instruction tuning, average accuracy rises 2.9 points across four video benchmarks and 5.1 across ten image benchmarks. These results concern this setup and do not establish temporal order caused gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.