The paper proposes information scope as a new interpretability axis for SAE features in CLIP and introduces the Contextual Dependency Score to separate local from global scope features, showing they influence model predictions differently.
Imagenet: A large-scale hierarchical image database
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
fields
cs.CV 2years
2026 2verdicts
UNVERDICTED 2representative citing papers
TokenTrace watermarks diffusion generations by jointly perturbing prompt embeddings and latent noise, enabling query-driven recovery of multiple independent concepts from one image.
citing papers explorer
-
Beyond Semantics: Disentangling Information Scope in Sparse Autoencoders for CLIP
The paper proposes information scope as a new interpretability axis for SAE features in CLIP and introduces the Contextual Dependency Score to separate local from global scope features, showing they influence model predictions differently.
-
TokenTrace: Multi-Concept Attribution through Watermarked Token Recovery
TokenTrace watermarks diffusion generations by jointly perturbing prompt embeddings and latent noise, enabling query-driven recovery of multiple independent concepts from one image.