ImageNet classes in LLaVA-Next are linearly decodable and steerable in the residual stream, and multimodal SAEs reveal that visual and textual features become increasingly shared in deeper layers.
Refusal in lan- guage models is mediated by a single direction, 2024
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Line of Sight: On Linear Representations in VLLMs
ImageNet classes in LLaVA-Next are linearly decodable and steerable in the residual stream, and multimodal SAEs reveal that visual and textual features become increasingly shared in deeper layers.