Reproduce “ANCHOR: Audio-Visually Grounded Chain-of-Thought Reasoning Benchmark” 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
Joel Julin, Souraja Kundu, Liza Dahiya, George Z. Wei, Liuyue Xie, Jingru Yi, Xu Zhang, Hao Yang and 1 more
ANCHOR evaluates whether video question-answer explanations are supported by temporal, spatial, visual, and audio evidence, alongside answer correctness. It includes 2,319 videos and 8,272 question–answer pairs, with human-refined annotations for 975 pairs. In cue-removal tests, removing video reduced accuracy by 38.4–41.8 percentage points across three models, showing cue dependence under the tested protocol, not that video alone determines performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.