Activation Matters: Test-time Activated Negative Labels for OOD Detection with Vision-Language Models

1. 1Identify the limitations of traditional negative labels in capturing OOD characteristics
2. 2Propose test-time activated negative labels to improve activation on OOD samples
3. 3Evaluate the effectiveness of the proposed approach on vision-language models

AI engineers and researchers working on vision-language models can benefit from this approach to improve out-of-distribution detection, which is crucial for reliable model performance

💡 Traditional negative labels may not effectively capture OOD characteristics, and test-time activated negative labels can improve detection performance

🔍 Improve OOD detection with vision-language models using test-time activated negative labels