MLLM eval pitfalls

[sam-motamed]

Very cool work! I am wondering how robust MLLM evals are given the best models even make lots of mistakes when videos are physically implausible?

as reference;

• TRAVL
• Impossible Videos
• VideoHallu

[123sio]

That’s a very fair question.
We do see clear limitations in current MLLMs when they act as judges for Physical Rationality. We’ve experimented with quite a few variants on this axis (different prompt designs, targeting critical questions related to physical rationality, adding priors like whether the scene is physically real or more semantic, as well as using different models as judges).
We then compared these judging strategies against human perception, and the current setup turned out to be the one that aligns best with human judgments among the options we tried.
We also include concrete case studies comparing different judges on Physical Rationality in Appendix A.3, which might be helpful to look at.