Reproduce “TwinPast: Benchmarking Historical Evidence Use in Streaming Video Understanding with Evidence Contrastive Video Pairs” 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
Haotian Yang, Yong Chen, Cheng Peng, Haobo Wang, Zhongle Xie, Xuanru Li, Gang Chen
TwinPast tests whether video answers follow earlier events by pairing videos with identical recent endings, questions, and options but different histories and correct answers. The authors found that recent frames sufficed for 128 of 400 sampled historical questions. For Qwen3-VL-8B, four recent frames produced 42.00% accuracy and 0.00% PairAcc; up to 64 sampled frames raised PairAcc to 34.00%, showing a remaining challenge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.