New 'Dr. No' Method Steers Vision-Language Models to Abstain on Weak Evidence
A new arXiv paper, 'Dr. No: Feature Steering for Abstention in Vision-Language Models,' proposes steering features so models can decline to answer when visual or textual evidence is insufficient for a reliable conclusion.
Read the full story at arXiv cs.LG ↗
Timeline · 1 report
- 2026-11-04 05:00 · arXiv cs.LG
Dr. No: Feature Steering for Abstention in Vision-Language Models