#CLIP
2 related articles
·2 min
Simplex-Constrained Representation: 97% Elimination of the Multimodal Embedding Modality Gap
A new paper proposes replacing the unit sphere with a probability simplex as the multimodal embedding domain, using a softmax adapter and total variation similarity to reduce the centroid distance of the modality gap by 97–99% on five retrieval benchmarks, with retrieval performance matching or exceeding baselines.
Read more →
·1 min
RADC: Risk-Aware Dual Caching for Test-Time Adaptation in Vision-Language Models
RADC introduces risk-aware dual caching for vision-language model test-time adaptation, using semantic foreground prototypes and Gaussian risk admission to achieve SOTA on cross-domain and OOD benchmarks.
Read more →