CLIP-like Model as a Foundational Density Ratio Estimator 事件

CLIP-like Model as a Foundational Density Ratio Estimator arXiv:2506.22881v3 Announce Type: replace Abstract: Density ratio estimation is a core concept in statistical machine learning because it provides a unified mechanism for tasks such as importance weighting, divergence estimation, and likelihood-free inference, but its potential in vision and language models has not been fully explored. Modern vision-language encoders such as CLIP and SigLIP are trained with contrastive objectives that im