Reproduce “Turns Matter: Advancing Streaming Video RL via Turn-Level Credit” 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
Shuai Dong, Kanghui Tian, Linghan Chen, Wenqi Shao, Hongrui Wu, Haowen Hou, Yongfu Zhu, Xuehui Wang and 15 more
TurnWise trains streaming video assistants to decide whether to respond or wait by assigning group-relative advantages to each decision from remaining returns after later actions are replaced with silence, and averaging loss across turns. On Qwen2.5-VL-3B, it scores 43.56 PAUC on ProactiveVQA, 21.59 Timing F1 on OVO-Timing, and 49.43 accuracy on StreamingBench, leading non-TurnWise results on these benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.