Reproduce “VisualTrust: Video Reasoning under On-Screen Claim Framing” 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
Huachao Zhu, Junjielong Xu, Sicheng Li, Liang Yang, Mingtao Chen, Zhichao Hu, Li Jiang, Pinjia He
The authors introduce VisualTrust, a benchmark testing whether video models let on-screen claims override conflicting visual evidence, and DEAR, a training-free method for evidence arbitration. Across 11 models, mean accuracy is 90.1% with no claim, 98.5% with an accurate claim, and 51.9% with a misleading claim. DEAR improves misleading-claim accuracy for three proprietary models, within this benchmark of screened, visually grounded questions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.