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.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 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.