Study Finds Frozen Vision-Language Models Abstain Unevenly Across Backgrounds
A new arXiv paper shows that selective prediction with a single confidence threshold can disproportionately reject examples from particular groups in frozen vision-language models, leading to uneven abstention.
Read the full story at arXiv cs.CV ↗
Timeline · 1 report
- 2026-11-03 05:00 · arXiv cs.CV
The Margin Knows the Background: Uneven Abstention in Frozen Vision-Language Models