Most video benchmarks fail to actually test video understanding

A systematic audit of 115 benchmarks reveals that many can be solved without watching the video, and the authors release a curated meta-benchmark of the hardest items.

Independent
Enxin Song · Yinuo Xu · Shusheng Yang · Wenhao Chai · Jiatao Gu
Research Digest··3 min read
Song et al.

The authors propose the attack pyramid, which organizes five levels of shortcut attacks with increasing access to a benchmark item: option level (only option layout), text level (adds question text), pool level (adds other items in the benchmark), frame level (adds one frame or caption), and order level (adds shuffled or truncated video).

Why this paper

Independent

In one line

Video benchmarks are vulnerable to shortcut attacks; Video-Index keeps only the 840 items that no attacker solves.

What we could check

  • ·No code link found
  • ·No weights link found
  • ·No dataset link found
  • ·No compute details found
  • ·No stated limitations found
  • ✓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.

§

Research Digest

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.