The authors constructed VisualTrust by taking five video sources and generating three versions of each clip with identical typography and placement: no on-screen claim, an accurate claim, and a misleading claim that alters a single grounded fact.
Video AI models trust misleading on-screen claims over what the footage shows
A new counterfactual benchmark, VisualTrust, shows that 42.8% of initially correct model answers flip when a rendered claim is edited to assert a false fact.
Chinese Tech
Huachao Zhu · Junjielong Xu · Sicheng Li · Liang Yang · Mingtao Chen · Zhichao Hu · +2 more
The Chinese University of Hong Kong, Shenzhen · Tencent
Research Digest··2 min read
The authors introduce VisualTrust, a counterfactual benchmark of 5,063 video-question cases in which the same clip appears with no claim, an accurate claim, or a misleading claim.
Why this paper
From Tencent and The Chinese University of Hong Kong, Shenzhen
In one line
Misleading on-screen text causes most video MLLMs to flip to wrong answers, but a decoupled arbitration agent restores accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§