LINE iteratively refines open-vocabulary neuron concepts in vision models via LLM proposals and text-to-image testing, achieving up to 0.11 AUC gains on ImageNet while uncovering 27% new concepts missed by fixed vocabularies.
This yields two sets of scalar activations: •A t: Activations from the synthetic concept images inP
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
LINE: LLM-based Iterative Neuron Explanations for Vision Models
LINE iteratively refines open-vocabulary neuron concepts in vision models via LLM proposals and text-to-image testing, achieving up to 0.11 AUC gains on ImageNet while uncovering 27% new concepts missed by fixed vocabularies.