A new attribution framework combines sparse autoencoder components with gradient-based attribution to reveal how CLIP models rely on semantically unexpected concepts.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
From What to How: Attributing CLIP's Latent Components Reveals Unexpected Semantic Reliance
A new attribution framework combines sparse autoencoder components with gradient-based attribution to reveal how CLIP models rely on semantically unexpected concepts.