Plausible explanations mislead multimodal AI judges
In brief
- Plausible explanations bias multimodal judges across 14 tested models.
- Plain justifications reduce incorrect answer rejection by 12.7 percentage points.
- Explanations sway decisions without improving judge accuracy.
A study shows that ordinary, plausible explanations can deceive multimodal AI models used as judges in visual evaluations. Results indicate a significant drop in the rejection of incorrect answers, even with simple justifications or chain-of-thought reasoning.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error